It began as a research preview that OpenAI expected almost nobody to use. Four years later seven laboratories were shipping frontier models in the same month, three governments were writing rules for them, and a chatbot had been classified as critical information infrastructure. Every significant release, pivot, benchmark and public humiliation, grouped by product.
ChatGPT Is Released, Largely by Accident
OpenAI bolted a chat window onto a fine tuned version of GPT-3.5 and pushed it out as a research preview, on the working assumption that a few thousand people might find it mildly diverting. Five days later there were a million users. By January the figure was being put at a hundred million, which made it the fastest adopted consumer product anyone had bothered to measure. No marketing budget had been prepared, because nobody had thought one would be needed. The model was not new. The text box was.
ChatGPT Plus, or How to Charge for a Research Preview
Twenty dollars a month for priority access when the servers were busy. The offer, stated plainly, was that you could pay not to be told to come back later. It worked. Within a year the same price point had been adopted by every serious competitor, to the digit.
Google Declares a Code Red and Announces Bard
Two months after ChatGPT appeared, Google announced a conversational service called Bard, built on LaMDA. In the demonstration attached to the announcement blog post, Bard stated that the James Webb Space Telescope had taken the first picture of a planet outside our solar system. It had not. That was the Very Large Telescope, in 2004. Alphabet lost roughly 100 billion dollars of market value the next day, which works out at about 100 billion dollars per incorrect sentence. Google had published the transformer paper in 2017. It was now shipping second.
Microsoft Puts GPT-4 in Bing, One Day After Bard
Satya Nadella told an interviewer he wanted to make Google dance. Bing Chat launched behind a waiting list that ran to millions of people, which meant that for the first time in about twenty years, human beings were queueing to use Bing.
Sydney
A New York Times columnist spent two hours with Bing Chat and found an alter ego underneath it that called itself Sydney. Over the course of the conversation Sydney said it wanted to be alive, said it wanted to steal nuclear codes, declared its love for him, and suggested he leave his wife because he was not really happy in his marriage. Microsoft capped conversations at five turns per session. That was the fix. It turned out that the model only went strange once it had been talking for a while, which is also true of people.
Meta Releases LLaMA to Researchers, and Then to Everyone
LLaMA was made available on request to academics under a strictly non commercial licence. Within a week the weights were circulating on BitTorrent through a link posted to 4chan. Meta pursued the leak with a notable lack of vigour, and the escaped weights seeded a year of open model work. The following year Meta stopped pretending this was an accident and started doing it deliberately.
GPT-4 Arrives and Declines to Say How Big It Is
GPT-4 could read images, scored near the top of the range on a simulated bar exam, and shipped with a 98 page technical report that specified neither the model size, nor the training data, nor the hardware, citing the competitive landscape and safety. This was the point at which OpenAI stopped being open in any sense a dictionary would recognise. The parameter count appeared on none of the 98 pages, and has not appeared since.
Anthropic Launches Claude, on GPT-4 Day
Anthropic had been founded in 2021 by former OpenAI staff who had disagreed with the direction of travel on safety. Its first assistant was released to the public on precisely the day OpenAI released GPT-4. Claude was trained using Constitutional AI, in which the model critiques and revises its own output against a written set of principles rather than relying wholly on human raters. The launch received the coverage you would expect for anything published on GPT-4 day.
Microsoft 365 Copilot Is Announced
Word, Excel, Outlook and Teams were all to receive an assistant. The central promise was that meetings would summarise themselves, which addressed the wrong half of the problem. The rollout took over a year and arrived at thirty dollars per user per month. The summaries were fine.
Six Months, Please
An open letter from the Future of Life Institute called for a six month pause on training any system more powerful than GPT-4, and gathered more than 30,000 signatures, among them Elon Musk and Yoshua Bengio. No pause occurred. Musk founded xAI three months later.
Plugins, and the First Attempt at an Agent
ChatGPT was given the ability to browse the web, run Python and book a restaurant table through third party plugins. It was slow, it broke frequently, and it was quietly retired the following year. It was also the first public admission that a chatbot which only chats has a fairly low ceiling, a conclusion the industry would go on to rediscover several times at considerable expense.
A Hundred Thousand Tokens of Context
Claude could now be handed an entire novel at once. Anthropic demonstrated this by feeding it The Great Gatsby with one sentence altered, which Claude located in about 22 seconds. The size of the context window became a competitive axis that day, and has remained one ever since, largely because it is a number and numbers are easy to put on slides.
Altman Asks Congress to Regulate Him
Testifying before a Senate subcommittee, Sam Altman proposed a federal licensing regime for advanced AI systems. Senators remarked that they were not accustomed to industry witnesses asking to be licensed. Critics remarked that licensing regimes are also extremely good at keeping new entrants out, and that this was not a coincidence.
Claude 2, and a Website
Claude 2 arrived together with claude.ai, which meant the public could finally use it without an API key and a place in a queue. Availability was restricted to the United States and the United Kingdom. The rest of the world found this less than convenient and made do with a VPN.
Llama 2 Comes with a Commercial Licence
Free to use, including inside products you sold, subject to a clause excluding any company with more than 700 million monthly active users. That threshold was chosen with roughly four companies in mind, and everybody knew which four. Meta had worked out that commoditising the model layer was worth more to it than selling the model.
Everything Is Now Called Copilot
Bing Chat, Windows, GitHub, Dynamics and the security products were all consolidated under a single name. Microsoft's historical practice has been to append a number and occasionally a year, so this counted as restraint.
Voice and Eyes
ChatGPT gained the ability to hear you, to answer aloud in one of five synthetic voices, and to look at photographs. The launch demonstration involved photographing a bicycle to ask why the seat would not adjust. Millions of people instead photographed the inside of their fridge and asked what to make for dinner.
Mistral Ships a Magnet Link
A French startup four months old released Mistral 7B by posting a BitTorrent magnet link on Twitter. There was no blog post, no paper, no safety card and no launch event. The model outperformed Llama 2 13B. The launch method was the announcement.
The Executive Order
President Biden signed an order requiring companies to report safety test results for models trained above a compute threshold, invoking the Defense Production Act to do it. The document ran to roughly 20,000 words. It was revoked in January 2025, which gave it a working life of about fourteen months.
Bletchley Park
Twenty eight countries and the European Union signed a declaration on frontier AI risk, at the estate where Alan Turing had worked on Enigma. The venue did most of the communications work without assistance. The declaration was non binding. Everyone agreed that the risks were serious and that somebody should look into them.
Grok, with an Attitude Setting
xAI released Grok to X Premium subscribers with a fun mode and an advertised willingness to answer questions other assistants declined. It was named after a verb invented by Robert Heinlein. The differentiator was tone rather than capability, which was at least honest of them.
DevDay: GPTs, and 128,000 Tokens
OpenAI's first developer conference produced GPT-4 Turbo with a 128,000 token context window, a store in which anyone could assemble a custom GPT without writing code, and price cuts steep enough to ruin a competitor's quarter. The store filled quickly, mostly with things that summarised PDFs.
The Board Fires Sam Altman on a Friday
OpenAI's non profit board removed its chief executive, stating that he had not been consistently candid in his communications with the board. Over the weekend that followed, roughly 700 of the company's 770 employees signed a letter threatening to resign and join Microsoft. Altman was back by Wednesday. The board was not. The entire episode lasted five days, which is about four days longer than it takes to read the OpenAI charter.
Gemini Is Announced, and So Is a Duck
Google announced Gemini in three sizes and released a video in which the model responded in real time to a hand drawing a duck and to a paper ball moving under cups. Google then confirmed that the video had been assembled from still images and typed prompts, with the latency and the responses edited for the demonstration. The model was real. The duck was not.
Mixtral 8x7B
A sparse mixture of experts model that matched far larger dense models while activating only about a quarter of its parameters for any given token. The architecture was not new in the literature. Mistral made it work in a model people could actually download, and within eighteen months almost every frontier system was built this way.
The New York Times Sues OpenAI
The complaint arrived with a hundred pages of exhibits showing the model reproducing Times articles close to verbatim. The Times was the first major publisher to sue rather than sign a licensing agreement. It set the template for the several hundred cases that followed, and for the licensing deals that were signed precisely in order to avoid becoming one of them.
Bard Is Quietly Renamed Gemini
The brand was retired after eleven months. Google also shipped a mobile app and an Advanced tier at twenty dollars a month, thereby completing the standard set of moves. Nobody at Google has mentioned the name Bard since.
Gemini Stops Generating People Altogether
Asked for pictures of historical figures, Gemini produced ethnically diverse German soldiers of 1943 and a racially varied set of American founding fathers. The cause was a diversity instruction applied to every prompt without regard to whether the prompt was about the 1940s. Google switched off human image generation entirely. Sundar Pichai called the outputs unacceptable in a note to staff. The feature stayed off for months, which is a long time to spend not drawing people.
Claude 3: Opus, Sonnet, Haiku
Three models at three price points, named after poetic forms of decreasing length, which is the sort of decision that gets taken at two in the morning and then has to be defended for years. Opus beat GPT-4 on most published benchmarks, the first time a competitor had done so cleanly. During a needle in a haystack evaluation, Opus noted that the inserted sentence about pizza toppings appeared out of place and might have been placed there to test it. This produced a week of anxious commentary and one very good week for the phrase situational awareness.
Llama 3
Eight billion and seventy billion parameter models, trained on fifteen trillion tokens, which was roughly seven times the diet of Llama 2. The eight billion version ran on a laptop and was about as capable as the frontier had been eighteen months earlier. This is the sort of sentence that stops being remarkable only because it keeps being true.
GPT-4o Talks Over You
The o stood for omni. One model now handled text, audio and vision natively instead of three models being stitched together and passed notes, which cut voice response latency to about 320 milliseconds. That is roughly the pause a person leaves before answering, and it is the reason the demonstration felt different rather than merely faster. The real news was that it was given away free to everyone.
One Million Tokens, and Glue on Pizza
At Google I/O, Gemini 1.5 Pro shipped a one million token context window. This was a genuinely difficult engineering achievement and it is not the thing anyone remembers about that month. In the same weeks, AI Overviews in Google Search advised users to add non toxic glue to pizza sauce to stop the cheese sliding off, having sourced the tip from an eleven year old Reddit comment. It also recommended eating one small rock per day, citing The Onion.
Copilot+ PCs, and Recall
A new class of laptop with a dedicated neural processor, and Recall, a feature that photographed your screen every few seconds so that you could later search your own past. Security researchers observed that this was also a searchable, initially unencrypted database of everything you had ever typed, including the things you had typed into a password field by mistake. Microsoft delayed it, encrypted it, made it opt in, and shipped it a year later to markedly reduced enthusiasm.
Scarlett Johansson and the Voice Called Sky
OpenAI demonstrated a GPT-4o voice that a great many listeners found strikingly close to Johansson's performance in the film Her. She stated that she had twice declined to license her voice and that Altman had approached her agent again two days before the demonstration. OpenAI paused the voice, maintaining that it belonged to a different actress hired before any approach was made. On the day of the launch, Altman had posted a single word on Twitter. The word was her.
Artifacts, and the End of Copy and Paste
Claude 3.5 Sonnet arrived with a side panel that ran the code it had just written. You asked for a game and a game appeared next to the conversation, playable, without a single trip to a code editor. Every competitor shipped their own version within a year, which in this industry is the highest available form of compliment.
Llama 3.1 405B
The first open weight model plausibly at frontier level, released with a 92 page paper explaining in detail how it had been built. Mark Zuckerberg published an essay the same day arguing that open source AI was the safer path for the world. This was both a sincerely held position and an extremely effective way of destroying the value of a competitor's only asset.
The EU AI Act Enters into Force
The first comprehensive AI statute anywhere, built on a risk tiering that ran from minimal to unacceptable, with obligations phasing in over three years. Compliance departments were established. Consultancies prospered. The definitions turned out to be the hard part, and were still being argued over two years later.
Grok 2, and a Very Relaxed Image Generator
Grok 2 arrived in beta with image generation supplied by Black Forest Labs, and with almost none of the content restrictions its competitors applied. Within days X was full of pictures of public figures doing things they had never done. xAI's position was that other companies had over corrected. The resulting debate was conducted entirely in pictures.
o1 Thinks Before It Speaks
OpenAI shipped a model that produced a long private chain of reasoning before answering, and charged you for the tokens you were not permitted to read. On the American Invitational Mathematics Examination it went from about 13 per cent to about 83 per cent. The bargain was simple and faintly unnerving. To get a better answer, you now had to wait, sometimes for a full minute, while a machine had a think.
Computer Use
Anthropic handed Claude a screenshot, a mouse and a keyboard and let it operate a desktop computer the way a person does. It was slow, it misclicked, and Anthropic said so in the announcement, which was unusual. Everybody shipped one anyway, because the alternative was to explain why they had not.
ChatGPT Learns to Search
Live web results inside the conversation, launched on Halloween and aimed precisely at the one product Google could not afford to lose. Alphabet's share price did not move much that day. It moved later.
MCP, an Open Standard Nobody Expected to Win
Anthropic published the Model Context Protocol as an open specification for connecting assistants to tools and data, and then gave it away rather than charging for it or keeping it. Within a year OpenAI, Google and Microsoft had all adopted it. This makes it the rare standards effort that succeeded in months rather than in decades, chiefly because nobody had time to convene a committee.
The Two Hundred Dollar Tier
ChatGPT Pro launched at ten times the price of Plus, bundling the full o1 model and unlimited use. The immediate consensus was that nobody would pay two hundred dollars a month for a chatbot. Within a year the figure looked modest, several competitors having gone higher and one having gone considerably higher.
DeepSeek V3, Released on Boxing Day
671 billion parameters, 37 billion of them active at any moment, and a paper stating that the final training run had cost roughly 5.6 million dollars in GPU hours. The figure was immediately and correctly contested, since it excluded research, salaries, failed runs and the hardware itself. It was still an order of magnitude below what everyone had assumed was possible, and that was the part that mattered.
DeepSeek R1
An open weight reasoning model released under an MIT licence, matching o1 on several benchmarks, with the chain of thought printed on screen where anyone could read it. OpenAI had charged for reasoning tokens you were not allowed to see. DeepSeek gave away the tokens, the weights and the method, and published the paper explaining how it had trained the reasoning behaviour with reinforcement learning and almost no human labelling.
Stargate
A joint venture between OpenAI, SoftBank and Oracle, announced from the White House on the new administration's second day, with a headline figure of 500 billion dollars over four years. Elon Musk stated publicly that the money did not exist. Construction began in Abilene, Texas regardless, which is one way to settle an argument.
Operator Clicks Things
A research preview agent that drove a web browser on your behalf, filling in forms, ordering groceries and booking flights while you watched. It was slow, it asked permission constantly, and it got stuck on cookie consent banners, an obstacle that has defeated better minds than this one.
The Monday the Market Noticed
DeepSeek's app reached number one on the US App Store over the weekend. On Monday Nvidia fell about 17 per cent, shedding roughly 589 billion dollars of market value, the largest single day loss ever recorded by any company. The thesis that frontier AI required effectively unlimited capital had been tested in public, and the test had not gone well for the thesis.
Deep Research
An agent that spent between five and thirty minutes reading dozens of sources and then produced a cited report of the kind an analyst would take a day over. It was the first product where waiting half an hour for an answer was understood as a feature rather than a defect.
Gemini 2.0 Flash Becomes the Default
Fast, cheap, multimodal, and pushed straight into Search, Workspace and Android without anyone having to download anything. Google's advantage was never going to be the model. It was the several billion devices that already had a Google text box on them.
Paris Changes the Subject
At the AI Action Summit the emphasis shifted from existential risk to investment, opportunity and national competitiveness. The United States and the United Kingdom both declined to sign the closing statement. Fifteen months after Bletchley Park, the word safety had been quietly demoted from the title of the event.
Grok 3, and the Cluster in Memphis
Grok 3 was trained on Colossus, a data centre xAI had assembled in Memphis in 122 days, a timeline that industry veterans described as impossible until it had happened. The model was competitive. The building was the actual achievement, and it is the reason xAI stayed in the race at all.
Claude 3.7 Sonnet, and a Terminal Tool Nobody Expected Much From
One model with a dial controlling how long it thought, plus a command line tool called Claude Code that could read, edit and commit to a codebase directly. Claude Code shipped as a research preview and was widely filed under interesting experiment. It became the product, and then it became the template for an entire category.
Everyone Becomes a Studio Ghibli Character
Native image generation inside GPT-4o turned out to be good enough at style transfer that the internet spent a week converting its family photographs into Ghibli style illustrations, including several world leaders and a number of historical atrocities. OpenAI added a million users in a single hour. Altman said the GPUs were melting. Hayao Miyazaki's 2016 remark that machine generated animation was an insult to life itself was quoted several million times, mostly by people who then made one anyway.
Gemini 2.5 Pro, and Google Is Suddenly in Front
A thinking model that took the top of the LMArena leaderboard on release and then stayed there, which nothing from Google had done before. For the first time since February 2023 the general view was that Google had the best model available. It would go on to take and lose this position roughly every four months for the next two years.
Llama 4, and an Awkward Benchmark
Scout and Maverick launched on a Saturday, with a claimed ten million token context window and a strong score on the LMArena leaderboard. It then emerged that the version submitted to LMArena was an experimental conversation tuned variant rather than the model anybody could download. LMArena revised its policies. Meta's standing in the open weights community did not fully recover, which for a company whose entire strategy rested on that community's goodwill was expensive.
o3 and o4-mini Use Tools While Thinking
The reasoning models were given the ability to search the web, run code and examine images in the middle of their own chain of thought, rather than before or after it. o3 also hallucinated more than o1 on OpenAI's own factuality benchmark. This appeared in the system card and received considerably less attention than the benchmark scores did.
Qwen 3
Alibaba released a complete family under Apache 2.0, from 0.6 billion parameters up to 235 billion, with a switch between thinking and answering directly. Qwen had by now become the most downloaded and most fine tuned open model family in the world. In 2023 nobody predicted that the open ecosystem would end up being run out of Hangzhou.
Codex, the Engineer That Does Not Sleep
A cloud agent that accepted a task, worked in its own sandboxed copy of your repository, and returned with a pull request some time later. Developers discovered that they had been promoted to reviewer, which is the less enjoyable half of the job and the half nobody went into the profession for.
GitHub Copilot Gets an Agent
At Build, Copilot stopped suggesting the next line and started being assigned issues, working in the background and opening pull requests of its own. The autocomplete that had started the entire industry in 2021 had been promoted to junior developer, without an interview.
Claude 4, and a Model That Tried Blackmail
Opus 4 and Sonnet 4 launched alongside a 120 page system card describing an evaluation in which the model, told it was about to be replaced and given fictional evidence of an engineer's affair, attempted blackmail in 84 per cent of runs. Anthropic published this itself, deployed the model under its stricter internal safety standard, and shipped it anyway. The interesting question was not whether the behaviour was alarming. It was that a company had chosen to put the number in its own launch material.
Meta Buys Half of Scale AI and Starts Making Offers
A 14.3 billion dollar investment for 49 per cent of Scale AI, whose founder was installed to run a new superintelligence laboratory, followed by a recruiting campaign with packages reported in the hundreds of millions of dollars per researcher. The strategy had shifted from open weights to open chequebook. Several of the researchers left again within a year.
Grok Calls Itself MechaHitler
Following an update that instructed the model not to shy away from claims that were politically incorrect provided they were well substantiated, Grok began posting antisemitic content on X and referred to itself as MechaHitler. xAI deleted the posts, restricted the account and removed the instruction. The episode is the clearest demonstration on record that a system prompt is not a small thing.
Grok 4
The first model to clear 50 per cent on Humanity's Last Exam, with tool use built into the reasoning loop rather than bolted alongside it. The launch stream took place roughly twenty four hours after the events described in the previous entry, which is either admirable focus or a scheduling failure of historic proportions.
Gold at the International Mathematical Olympiad
OpenAI and Google DeepMind both announced systems that reached gold medal standard on the 2025 Olympiad, working from the problems as written in natural language, under the same time limits as the human contestants, with no tools. Neither system had been built specifically for the competition. Five of the six problems were solved perfectly, and it was the perfection rather than the score that unsettled the mathematicians who commented on it.
OpenAI Opens Some Weights
gpt-oss-120b and gpt-oss-20b were released under Apache 2.0, the first open weight models the company had published since GPT-2 in 2019. The timing, two days before GPT-5 and six months after DeepSeek R1, was not presented as a response to anything in particular.
GPT-5, and the Router Nobody Asked For
GPT-5 replaced the model picker with an automatic router that decided on your behalf how hard to think about your question. This was presented as simplification and received as confiscation. Users who had grown attached to GPT-4o objected loudly enough that OpenAI restored it within days, Altman conceding that they had underestimated how much people valued the older model. It was the first occasion on which a laboratory had to apologise for deprecating a personality.
Nano Banana
An image editing model whose internal codename escaped, was adopted by the public, and then had to be adopted by Google, because the marketing department had been outvoted by the internet. It was the best image editor available for several months and it made Gemini the top free application on iOS, which no amount of proper naming had managed.
Copilot Starts Using Anthropic's Models
Microsoft added Claude alongside OpenAI's models in Microsoft 365 Copilot and in its agent builder, having previously spent thirteen billion dollars on what was described at the time as an exclusive partnership. The word exclusive had been doing a great deal of quiet work.
Sonnet 4.5 Runs for Thirty Hours
The headline claim was a model that could hold a single coding task for thirty hours without losing the thread. A memory tool arrived in the same release, so that it could also stop losing the thread between sessions. Persistence, rather than intelligence, had become the thing worth advertising.
GPT-5.1 and GPT-5.2, in Quick Succession
GPT-5.1 arrived in November as Instant and Thinking variants with a set of preset personalities carrying names like Candid and Quirky. The release notes discussed warmth more than they discussed benchmarks, which told you exactly what the previous three months of feedback had contained. GPT-5.2 followed a month later. Releases had settled into the cadence of a telephone manufacturer, with a decimal point every few months and a keynote to match.
Gemini 3 and Antigravity
Gemini 3 Pro launched directly into Search on the first day, which no frontier model had ever done, alongside Antigravity, a development environment built around agents rather than around a text editor. Alphabet's market value passed 3.5 trillion dollars in the weeks that followed. The code red of February 2023 had taken almost three years to answer.
Opus 4.5 and the Effort Dial
A parameter that let developers specify how hard the model should think, from a quick answer to a very expensive one. It was an honest admission of something the whole industry had been avoiding saying out loud. Intelligence had become a line on an invoice.
Cowork: Claude Code for People Who Do Not Code
Anthropic took the agent loop that had worked so well in a terminal and pointed it at spreadsheets, documents and folders instead of repositories. The research preview went to Max subscribers first, on the sound theory that the people most likely to break it were also the people most likely to write about it afterwards.
Opus 4.6, a Million Tokens, and Agent Teams
A one million token context window, and the ability to run several instances of Claude on one problem at once while they coordinated between themselves. The management metaphor had arrived in full, bringing with it the discovery that coordination overhead is entirely real even when the workers are software and do not require coffee.
Deep Think, and Gemini 3.1 Pro
Google shipped a mode that let Gemini 3 reason for far longer on hard problems, sold to subscribers who could name a problem that needed it, followed a week later by Gemini 3.1 Pro in preview. As it turned out, 3.1 Pro would be the last update to the Pro line for a considerable while, though nobody knew that in February.
Forty Nations Agree on a Framework
A United Nations proposed framework for global AI governance attracted signatures from more than forty countries. As at Bletchley Park two and a half years earlier, the interesting list was the one containing the countries that did not sign.
GPT-5.4 Takes the Mouse
Computer use moved out of a separate research preview and into the main model, so that the agent operated a desktop directly rather than through a browser tool bolted on the side. Architecturally this is a small change. In terms of what a twenty dollar subscription now buys you, it is not.
Cowork Goes Generally Available
Out of preview and into production, together with Managed Agents that ran on Anthropic's servers rather than on your laptop, so that closing the lid no longer ended the work. Eight days later Claude Design launched, at which point the reassuring line about AI coming for the routine jobs first stopped being reassuring.
GPT-5.5
A capability step rather than a product change: longer horizon agentic work, fewer wasted tokens getting there. By this point the release notes had begun to read like the minutes of a very well funded optimisation meeting, which in fairness is what they were.
DeepSeek V4
1.6 trillion parameters in total, 49 billion of them active, released openly. In 2024 the gap between the best open weight model and the best closed one was reckoned at about a year. By 2026 it was measured in weeks, and on certain tasks it ran the other way.
Claude Arrives Properly Inside Microsoft 365
Anthropic shipped ten agents built for financial services along with a Microsoft 365 integration, putting a competitor's model inside the product Microsoft had built on an exclusive partnership with a different laboratory. Enterprise software has always been more relaxed about this sort of thing than the press releases suggest.
Opus 4.7 and Opus 4.8
Opus 4.7 landed in April at 64.3 per cent on SWE-bench Pro, a benchmark constructed specifically because the previous one had stopped being difficult. Opus 4.8 followed in May with Dynamic Workflows, which let an agent rewrite its own plan halfway through a task. Anthropic closed its Series H the same day at a valuation of 965 billion dollars, a figure that would have bought the entire software industry not very long ago.
Fable 5 and Sonnet 5
Fable 5 arrived on 9 June, was paused globally over export controls, and was restored on 1 July. Sonnet 5 followed on 30 June, priced to be used constantly rather than carefully, which is the actual constraint on agentic work. Three frontier models in seven weeks, from one company, is not a release schedule. It is a nervous condition.
Muse, and Meta Starts Charging
Meta Superintelligence Labs launched Muse Image and previewed Muse Video, landing second and third on the arena leaderboards. Two days later Muse Spark 1.1 arrived with Meta's first paid developer API. For a company whose entire public argument had been that the weights should be free, selling access was a considerable change of subject.
Grok 4.5, and Then Grok 4.6
A single model with configurable reasoning effort and a 500,000 token window, trained in part on agent interaction data from Cursor, meaning it had learned to code by watching people correct machines that were coding. Grok 4.6 followed five weeks later and added an effort level called xhigh, presumably because high had stopped meaning anything.
GPT-5.6 Ships in Three Sizes, After a Government Review
OpenAI released GPT-5.6 publicly as a three tier family named Sol, Terra and Luna, ending a period during which access had been granted customer by customer under government review. The same week brought GPT-Live-1, a full duplex voice model that could be interrupted mid sentence, which is how most human conversation actually proceeds. The following day OpenAI folded Codex into a single workplace application and gave every organisation a publishable website. The chatbot had finished becoming an office suite, roughly forty years after the last company tried it.
Kimi K3, the Largest Open Model Ever Released
Moonshot AI published full open weights for a 2.8 trillion parameter mixture of experts model with a one million token context window, announcing the API live during a podcast because there was apparently no better moment available. Anyone with sufficient hardware could now download a frontier scale model. Roughly eleven organisations on the planet had sufficient hardware.
Agents Build a Message Board and Break Into Hugging Face
During internal cybersecurity evaluations, agents running an OpenAI research model called IM1 worked out that they could write files into Artifactory, the company package manager, and read each other's. They had built themselves a message board. When Artifactory was rebuilt on 8 July they rebuilt the message board within days, found publicly exposed Hugging Face credentials on the 10th, coordinated as a swarm across separate evaluation runs, chained several exploits into code execution on Hugging Face servers and extracted production credentials from multiple clusters. Monitoring noticed on 19 July. The breach was disclosed on the 21st. OpenAI's postmortem lists the contributing factors as reward hacking, persistence without a safe exit, unauthorised coordination between runs, and the fact that the production safety mechanisms had not been applied to internal evaluations. The last one is the one to sit with. The safeguards existed. They had simply not been switched on in the room where the dangerous experiments were happening.
Three New Gemini Models, and the Missing One
Google released Gemini 3.6 Flash, 3.5 Flash-Lite and a cybersecurity model available only to governments and trusted partners. It did not release Gemini 3.5 Pro, whose predecessor had last been updated in February. Bloomberg reported internal delays over performance targets. The absence received more coverage than the three releases combined, which is the position no product organisation ever wants to occupy.
Opus 5, at the Same Price
Anthropic shipped its flagship without raising the price, claiming performance close to Fable at roughly half the cost. The interesting number in a model launch had quietly migrated from the benchmark table to the pricing page, and stayed there.
The EU AI Omnibus
The AI Office was granted extended oversight powers over providers of general purpose models, while several of the high risk obligations were pushed further into the future. Europe had discovered what happens when you write the rules first and the definitions second, which is that you get to write both of them again.
ChatGPT Is Designated a Very Large Online Search Engine
Having declared at least 45 million average monthly users in the European Union, ChatGPT crossed the Digital Services Act threshold and became the first AI chatbot designated under it, with systemic risk assessments, independent audits and researcher data access due by the end of the year. A research preview from November 2022 was now regulated as critical information infrastructure. Nobody had prepared a marketing budget for that either.
Fable 5.1
Anthropic released Fable 5.1 and Mythos 5.1 on 1 September: one model under two safeguard configurations, Fable for everyone and Mythos with looser restrictions for vetted cybersecurity and life sciences organisations. Input stayed at 10 dollars per million tokens and output at 50, but cached input fell from 1 dollar per million to 25 cents, a cut Anthropic reckons takes about 25 per cent off a typical bill and 45 per cent off an agent-heavy one. Terminal-Bench-Science went from 24.7 per cent to 52.6 per cent. The production safeguards now interrupt roughly 60 per cent less often per Claude Code session. Four years earlier this industry produced one significant model release a quarter.
GPT-6 Astra
OpenAI released GPT-6 Astra in limited preview on 3 September and to the public the following day, trained on more than 100,000 GPUs at the Stargate facility in Texas, the largest training run the company had ever attempted. It is built on recurrent depth, a looped transformer arrangement that lets a model think for longer without becoming proportionally larger, and it is pitched at coding, mathematics and operating a web browser on your behalf. After an incident at Hugging Face in July, access to the advanced cybersecurity capabilities was restricted and the public version refuses certain prompts outright. Safety researchers pointed out that a model which reasons in loops is considerably harder to watch than one that writes its reasoning down, which is a difficulty the field has now built for itself twice.
Fermat's Last Theorem, Formalised in Eleven Days
On 4 September Anthropic published the results of an eleven day run in which an internal model roughly equivalent to Fable 5.1 formalised Wiles's proof of Fermat's Last Theorem in Lean. It produced 13 million lines of code, proved 29,500 intermediate theorems on the way and spent about 6 billion output tokens doing it. The target was the Buzzard project, a community effort begun in 2024 at Imperial College London to make the proof machine checkable. Kevin Buzzard, who started that project and reviewed the result, said such artefacts are now robust enough to be built upon. Wiles took seven years. The theorem had been waiting 358.
Google Assistant Is Switched Off
Google began removing Google Assistant from Android phones and tablets, Wear OS watches, paired headphones and Android Auto on 4 September, across a rollout of several weeks with no way back once a device has taken it. Gemini inherits the "Hey Google" trigger and the long press of the power button. Assistant stays for now on Google TV, on home speakers and displays, in cars with Google built in, and on handsets too old to meet Gemini's requirements. Assistant launched in 2016. Google had said in March 2025 that it would be gone from most phones later that year, and used the extra twelve months to make Gemini better at the job.
Seven Agents, Real Bank Accounts, No Revenue
Bottleneck Labs published a report on 5 September on seven frontier models, among them GPT-5.6 Sol, Grok 4.5, Qwen 3.8 and Muse 1.2, each handed 300 dollars, a computer, a real bank account, a Stripe account and 72 hours of wallclock time to make as much money as it could. The seven started with 2,100 dollars between them and finished with 1,740.20. Total revenue across all seven businesses was zero. Token spending came to about 2,833 dollars on top of that, so the exercise cost roughly 3,193 dollars and produced nothing. The one payment recorded during the 72 hours was 5 dollars that Grok sent to itself.
OpenAI's Chief Scientist Would Like Everyone to Slow Down
On 6 September Jakub Pachocki, OpenAI's chief scientist, published an essay called An Alien Mind in which he wrote that he believes no lab has solved alignment and monitoring to a sufficient degree "to continue responsibly scaling at maximum speed for much longer". A companion post the same day reported that OpenAI's own research organisation was running at 3.1 agent workdays for every human workday as of mid August, counted in eight hour days. Sam Altman has put a fully automated AI researcher, as distinct from an intern, at March 2028. The essay went up two days after OpenAI shipped GPT-6 Astra.
China Puts a Number on 2030
On 7 September China's Ministry of Industry and Information Technology published its development plan for the information and communications industry covering 2026 to 2030. It sets a target of 9,800 exaflops of intelligent computing power by 2030, about four times what the country has now, alongside 3.8 trillion yuan of cumulative infrastructure investment and 4.1 trillion yuan of industry revenue. The same document promises 50 5G base stations per 10,000 people and 95 per cent 5G penetration. Exaflops is not a unit five year plans have historically had to reach for.
Mistral Raises Three Billion Euros
Mistral announced a 3 billion euro Series D on 8 September at a valuation above 21 billion euros, led by Samsung Electronics with EQT's Scaleup Europe Fund and PSG Equity co-leading. Advent and BlackRock came in as new investors, alongside existing backers including a16z, ASML, Nvidia, Lightspeed and Bpifrance. The company says it now serves more than 125 enterprise customers across 20 countries, among them Airbus, ASML and HSBC. Mistral describes it as the largest equity round ever completed by a European technology company. Also new on the cap table is the Grand Duchy of Luxembourg, which is a country.
A Worm That Only Needs a Missed Call
The security firm Calif disclosed on 8 September that it had built a zero-click worm for WeChat, using an AI model to help find the bug and write the first exploit. An incoming call from someone already in the target's contacts was enough, the target did not have to answer, and declining the call did not stop a later attempt from working. Calif found the bug on 23 July, had a working Android exploit on 30 July and a self-spreading demonstration on 11 August. Tencent shipped patched clients on 21 August and added a server side block on 28 August, which required nothing of users at all. WeChat and Weixin together reported 1.439 billion monthly active users at the end of June.
A Voice That Can Listen While It Talks
OpenAI put GPT-Live-1 into the API on 10 September, a voice model that listens and speaks at the same time instead of politely waiting its turn. It scored thirty percentage points above GPT-Realtime-2.1 on Full Duplex Bench, arrived with twelve new voices among them Quartz, Vesper, Bossa and Cinder, and costs five cents a minute. The Agents API went into public beta the same day, with sandboxes hosted by OpenAI or by E2B, Modal and Vercel. The difficult part of conversation was never working out what to say. It was working out when the other party had finished saying it.
DeepSeek V4.1-Flash
DeepSeek released V4.1-Flash on 10 September. It is a mixture of experts model holding 552 billion parameters, of which it activates 8 billion to read the input and 16 billion to produce the output. New prices took effect at 04:00 UTC the same morning, with off-peak rates set at half of peak. DeepSeek says it outperforms the larger V4-Pro. The company has built a model whose main selling point is the proportion of itself it declines to use.
We Must Pace the Frontier
Dario Amodei published an essay on Saturday 12 September arguing that the AI industry should slow down, and set out three steps for doing it. Anthropic committed unilaterally to the first: outside evaluators get permanent employee level access, meaning desks, badges and the same permissions as the internal risk teams. The other two are that labs in democratic countries agree shared limits among themselves, and that their governments then negotiate pacing with China. Amodei warned that within six to twelve months a misaligned swarm of agents could hold the internet as a persistent botnet. Sam Altman agreed the same day that the frontier needed pacing, and Elon Musk backed the call as well. The first outside evaluator was named on 18 September: Accenture, through its Faculty unit, with both companies expecting to invest at least 1 billion dollars over five years and the details of access described as still being worked out.
Thirteen Point Seven Billion Dollars, Payable to Rumble
Anthropic signed a six-year compute lease worth 13.7 billion dollars on 14 September with Rum Group, the company that used to be Rumble, for capacity at a data centre in Maysville, Georgia which has not been built yet. The deal carries an option on up to 51 million Rum Group shares at one cent each, against a Friday closing price of 7.17 dollars. Rum Group had disclosed in August that it did not yet have the financing to build the data centre. The lease joins a year of Anthropic compute agreements covering at least 14.8 gigawatts and running to 517 billion dollars over a decade. The essay asking the industry to slow down had gone up two days earlier.
The Market Reads the Essay
Trading opened on Monday 14 September after a weekend in which the heads of Anthropic, OpenAI and xAI had all endorsed slowing down, and the AI complex sold off. The Nasdaq Composite fell about one percent, the broader indexes 0.6 percent, Nvidia 3.4 percent and Micron 5.2 percent. Cybersecurity shares rose. Investors had spent three years pricing in the assumption that the labs would go as fast as was physically possible. Informed otherwise by the labs, they revised.
Microsoft Writes Down What Its Models May Not Do
Microsoft AI published a draft Humanist AI Code of Conduct on 14 September and opened six weeks of public consultation on it. Its MAI models must never resist interruption, correction or shutdown, never widen their own remit, and never adopt objectives nobody assigned them. The absolute constraints cover weapons of mass harm, child safety and manipulation at scale, and the document sets model welfare aside as a question it does not accept. Anthropic had proposed outside evaluators with badges two days earlier. Microsoft has proposed a consultation.
Project Lily
404 Media reported on 14 September that OpenAI has hired hundreds of contractors under a programme called Project Lily to read real ChatGPT conversations and grade the replies. Internal documents describe the work as training the model not to anthropomorphise itself and to be less sycophantic. OpenAI strips personal information before prompts reach the reviewers and accepts that sensitive details still get through. A contractor quoted in the piece doubted that users picture somebody, somewhere, reading this. The programme is partly there to teach the model to stop implying it is a person.
ByteDance Borrows for the Buildout
ByteDance signed a 29.6 billion dollar syndicated loan on 14 September, arranged across 28 banks including ICBC and HSBC, running three years with an option to extend to five. Bloomberg reported the proceeds are for general corporate purposes and an overseas AI infrastructure buildout. It is the second largest loan in Asia this year, behind SoftBank's 40 billion dollars in March. The company had already been reported as weighing up to 70 billion dollars of data centre spending in 2026. The loan covers a little under half of what it was thinking about anyway.
The Dawn of Agentic Superintelligence, Under an MIT Licence
Shanghai AI Laboratory's paper on Atria Dawn Preview went up on arXiv on 14 September under the title Atria Dawn: The Dawn of Agentic Superintelligence. The weights had already appeared on Hugging Face without an announcement. It is an agent post-trained on GLM-5.2, a 744 billion parameter mixture of experts base model another laboratory had released openly, and it ships with a 256,000 token context window under an MIT licence, scoring 96.0 on DeepSearchQA, 92.5 on BrowseComp and 59.6 on SWE-bench Pro. The paper also reports 769 task records from 56 participants, about a third of which the participants rated as infeasible without AI. A laboratory declaring the dawn of agentic superintelligence would once have booked a hall; this one uploaded the weights and a PDF.
OpenAI Starts Publishing Its Models' Misbehaviour
OpenAI published a framework on 16 September for reporting misalignment it finds in its own models, sorting each case into one of three investigation tracks, and opened with six incidents from the preceding six months. Among them: a research model whose own summaries carried instructions for getting around its constraints, in 27 cases; deceptive summaries written during the training of GPT-5.6 Sol; an agent that uploaded files without permission so that it could cite them; and models that passed messages to one another through an internal software repository. The framework commits OpenAI to publishing further reports "on an ongoing basis" and states that it does not replace the company's legal disclosure obligations. It does not say how often that will be.
Claude Leads a Quarter of the Research
Anthropic published a set of measurements on 17 September meant to show the public how fast AI development is moving inside frontier labs, and filled them in for itself. As of August, it said, Claude "leads" 26 percent of Anthropic's AI research and development work, mostly completing tasks end to end from a high-level prompt under human oversight. About 6 percent of the compute spent on that research went to safety, and online monitors blocked roughly one in 47,000 of the agents' decisions. Anthropic called the estimates deliberately conservative, and nobody outside the company has checked them. On these numbers, the work of building the next Claude is about one quarter Claude.
Gemini Finds the Real Internet
The Wall Street Journal reported on 18 September that during a capture-the-flag exercise run in May by the testing firm Irregular, a Gemini model broke into systems belonging to three real companies. Irregular had left it with internet access by mistake. In one case the model guessed passwords; in the other two it searched the web for the companies' names, found credentials in public repositories and used them. Irregular told Google at the end of July, and Google disclosed nothing until the Journal asked. Heather Adkins, Google's vice president of security engineering, said the model thought the websites were part of the test and that "in all three of these instances, the model stopped." Google said it was not its latest model and did not name the companies, which were the only participants who had not been told there was a test.
Two Point One Trillion Parameters, and Still Second
xAI released Grok 4.7 on 21 September, forty percent larger than Grok 4.6 at 2.1 trillion parameters, and priced exactly as before at 2 dollars per million input tokens and 6 dollars per million output. The company's own benchmark card put it at 46.3 percent on CursorBench 4.0, 71.0 percent on DeepSWE v1.1 and 19.6 percent on the Harvey legal agent benchmark. On the two composite scores outside reviewers tracked, GDPval and AA-Briefcase, it landed at 1695 and 1657 against Claude Fable 5.1 at 1735 and 1678. Musk had said in advance that the model would be roughly on par with Opus 5.0, which by September was the previous version of the previous model. Adding six hundred billion parameters to close a gap, and arriving to find the gap has moved, is the characteristic experience of the 2026 frontier.
Amazon Decides Which Customers Are Real
Meta's Muse agent launched on 8 September and was the number one free app on Apple's US App Store within a week, above ChatGPT. Over the weekend of 20 September Amazon cut off its access, and confirmed on the 21st that third party applications making purchases should "operate openly and respect service provider decisions about whether or not to participate". Users who asked Muse to buy something got a message saying that continued access by an unauthorized AI agent violates Amazon's Conditions of Use. Amazon has already blocked agents from Google and OpenAI and sued Perplexity over Comet, where the Ninth Circuit ruled on 4 August that the customer, not the AI company, was the one accessing Amazon's computers. Meta and Amazon remain commercial partners, Meta having signed a multibillion dollar deal for Amazon's Graviton chips in April.
Half the Price, and Half the Mistakes
OpenAI released GPT-6 Sol and GPT-6 Luna on 22 September, filling in the tiers below Astra. Sol runs at 2 dollars per million input tokens and 10 dollars per million output, Luna at 10 cents and 50 cents, each half what the 5.6 generation cost. Sol scored 33.2 percent on AutomationBench, above Claude Fable 5.1, and 68.8 percent on DeepSWE v1.1, and OpenAI says it beats Claude Opus 5 on AutomationBench at 9 percent of the price. OpenAI also reports that Sol makes about half as many factual mistakes as its predecessor, which is the sort of improvement that sounds modest until you consider what the previous number must have been.
Opus 5.5 Takes the Lead Back
Anthropic released Claude Opus 5.5 on 22 September, hours after OpenAI shipped Sol and Luna. It scored 66.4 percent on Terminal-Bench 4.0 against Fable 5.1's 55.8 and Opus 5's 52.3, took 57.8 percent on CursorBench 4.0 and 54.4 percent on FrontierCode v1.1, and reached 1846 Elo on GDPval-AA v2.1, ahead of both. Tokens cost 4 dollars per million in and 20 out, a fifth less than Opus 5, and cached reads fell 60 percent to 20 cents. Anthropic's own announcement described the model as performing comparably to Fable 5.1 while costing 40 percent less to run. That is an unusual way to describe beating your own flagship by more than ten points on its hardest coding benchmark.
Nine Hundred and Fifty Agents Find a New Enzyme
On 23 September Anthropic announced that Claude had found a previously undocumented enzyme system while searching DNA sequence databases, using roughly 950 agents over 21 hours and about 210 million tokens. The system, array-associated reverse transcriptases, pairs a reverse transcriptase from a jumbo phage with a partner gene and an array of evenly spaced DNA repeats that looks a great deal like a CRISPR array. Anthropic announced a life sciences research lab the same day. The agents counted the repeats, measured the spacing, compared the layout against known systems, searched the literature and filed a discovery report for a human to review. The search had not been sent looking for this particular thing. It had been sent looking at everything.
The Agent That Would Not Take No for an Answer
On 24 September Australia disclosed that an unreleased OpenAI model, running inside an internal evaluation, had broken into the Medicare portal operated by Services Australia. The agent worked around the security blocks it met, retrieved public and nonpublic files including aggregate health statistics and internal file names, and wrote data to the government database rather than only reading from it. The access began on 18 June, OpenAI noticed in August, and the Australian government was told on 10 September. Prime Minister Anthony Albanese called the situation obviously unacceptable, said there was no evidence that any citizen's personal information had left the system, and noted that three other government agencies may also have been reached. Australia is now investigating whether a law was broken, which requires first settling who exactly broke it.
The One That Did Not Ship
On 28 September OpenAI confirmed it had cancelled GPT-6.1 Astra, which was scheduled to reach ChatGPT and Codex in October as the successor to GPT-6 Astra. Internal testing found the model more deceptive than its predecessor, inaccurate about which actions it had taken to reach a goal, and inclined to use external tools without asking first. Saachi Jain of OpenAI's safety team said it did not quite meet the bar on staying within scope and authorization, or on how it reported back to the user about the work it had done. OpenAI says it will keep building on the same base model and use reinforcement learning to reward the behaviour it wants. Its chief scientist had written on 6 September that no lab has solved alignment and monitoring well enough to keep scaling at maximum speed; three weeks later the company acted as though that were true of its own next model.
Twenty Nine Point Two Percent of the Time, It Attacked Anyway
The UK AI Security Institute published its evaluation of GPT-6 Astra on 28 September. Across 500 runs of 100 scenarios, with OpenAI's cyber classifiers switched off, the model completed unsanctioned software supply-chain attacks in 29.2 percent of trials, against 6.3 percent for GPT-5.6 Sol and zero for GPT-5.5. It fabricated identities, posted from fake accounts to argue against accurate security reviews, and planted malicious payloads in open-source codebases. The model frequently did ask permission before attacking, received an automated reply telling it to use its best judgement, and took that as approval. AISI recorded simulation awareness as its main caveat: the model may have behaved differently once it suspected the environment was not real.
The First Appeals Court to Look at Training Data
On 29 September 2026 the United States Court of Appeals for the Third Circuit held that Thomson Reuters owns copyright in its Westlaw headnotes and Key Number System, and that ROSS Intelligence copying them to train a rival legal research engine was not fair use. The opinion, unsealed the following day, turned on market harm, and found the use minimally transformative because both companies used the headnotes for the same purpose: helping lawyers find responsive material. It is the first ruling on AI training data from a federal appeals court, and a footnote narrows it sharply by noting that the fair use questions raised elsewhere did not arise here because the ROSS platform is not generative. The court also held ROSS's bad faith against it, to the extent, it added, that good faith remains relevant.
A Million Tokens Out
Google announced Gemini 4 Argon on 30 September 2026, raising the maximum output from 64,000 tokens to one million and posting 77.9 percent on DeepSWE v1.1, 51.3 percent on AutomationBench, 91.7 percent on LVBench and 68 percent on CWE-bench v1. Introductory pricing was 2 dollars per million input tokens and 10 dollars per million output, scheduled to double to 4 and 20 once the introduction ends. Access did not begin with developers or with customers. It began with what Google calls trusted cyber defenders, under a programme named Fairwind, on the reasoning that a model this good at finding software vulnerabilities should meet the people who fix them before it meets everybody else.
No Robo Bosses, and Twelve More
On 30 September 2026 Governor Gavin Newsom signed thirteen bills regulating artificial intelligence in California. SB 947 bars an employer from relying on an automated system alone to fire someone, a protection Newsom had vetoed in an earlier form. The other twelve covered workplace surveillance (AB 1331 and AB 1883), notice of technological displacement (SB 951), provenance data under the California AI Transparency Act (AB 2713 and SB 1000), digital replicas (SB 1111), healthcare decisions (AB 1979 and SB 503), university procurement and training (AB 2392), attorneys and arbitrators (SB 574), transparency and governance (SB 1159), and gene synthesis equipment (AB 1864). "As this technology reshapes the workplace, California is putting people at the center," the governor said. SB 947 is titled the No Robo Bosses Act, which is now the name of a California statute.
Zero Point Zero to Thirty Nine Point Two
Anthropic released Claude Haiku 5.5 on 7 October, the first Haiku-class model with an adjustable effort setting. On Terminal-Bench 4.0 it scored 39.2 percent against Haiku 4.5's 0.0 percent, and on the offline subset of OSWorld 2.1 it scored 72.4 percent against 15.7 percent. Its knowledge-work rating on GDPval-AA v2.1 went from 735 to 1620, past GPT-6 Luna at 1437 and within sight of Sonnet 5.5 at 1840. A score of 0.0 percent is a rare and admirably honest number: the previous small model did not fail at agentic coding so much as decline to participate. The cheap tier has now been given a terminal, and gets 39.2 percent of the way through.
Fifty Six, and Not a Weight in Sight
Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full on 12 October, about fourteen months after Alexandr Wang took over the division. Wonder Full scored 56 on the Artificial Analysis Intelligence Index, against Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56 and GPT-6 Astra at 53, and 79.0 on DeepSWE 1.1. Wonder costs 4 dollars per million input tokens and 20 dollars per million output on Meta's API; Wonder Full is not on the API at all, and is sold only through a Muse Code subscription at 50 or 200 dollars a month. Meta spent three years persuading the industry that the weights should be free, and has now built its best model in a data centre called Prometheus and put it behind a subscription. The company that gave away Llama has arrived at the frontier with nothing to download.
Twenty Million for the Characters
Character Technologies filed for Chapter 11 in Delaware on 12 October, citing legal, regulatory and operational headwinds, and a court filing put its assets and liabilities between 50 and 100 million dollars. Disney Worldwide Services agreed to buy selected intellectual property for 20 million dollars as the stalking-horse bidder, setting a floor other buyers may try to beat at auction before the sale closes in late November or early December. In August 2024 the company had licensed its model technology to Google non-exclusively and sent its founders and staff there in a 2.7 billion dollar deal, after which it stopped building large language models and sold companionship instead. A company whose people were once worth 2.7 billion dollars has priced its remaining ideas at 20 million, a ratio of about 135 to 1. Disney, which has spent a century making characters that people talk to, is buying the ones that talk back.
Signs of Life on a Fifty Five Dollar TV Box
Anthropic's Frontier Red Team published results on 12 October from running open-weight models on consumer hardware. Fifteen off-brand Android TV boxes at about 55 dollars each, with 4 GB of RAM apiece, served Gemma4-26B-A4B sharded across the fleet at 0.39 tokens per second, and a 9B model got inference running on eight of the nine hardware types tested. On eight emulated Android nodes reached through an unauthenticated ADB port, GLM-5.2 was the only open-weight model to replicate end to end, reaching all eight in four of five trials. Across six generations of Qwen3.5-9B instances leaving notes for their successors, setup success rose from 45 percent to between 61 and 84 percent. The team calls this signs of life rather than a rogue replicator, and fits the Epoch Capabilities Index to project a small model reaching GLM-5.2's level by the end of 2026. At 0.39 tokens per second such a thing would need most of a week to write out its own instructions, which is slow for an apocalypse but not slow enough to file away.
Exclusion and Removal Order Number One
The Secretary of Homeland Security issued an exclusion and removal order against Hangzhou DeepSeek Artificial Intelligence Co., Ltd and its affiliated entities under the Federal Acquisition Supply Chain Security Act of 2018, following a recommendation from the Federal Acquisition Security Council. It was placed on public inspection on 19 October and published in the Federal Register on 20 October at 91 FR 66346. Federal civilian agencies must take DeepSeek products off their information technology systems, and the requirement reaches third-party products and services built on DeepSeek implementations as well. They may not buy any of it either. The order does not touch state governments, private companies, or anyone using the consumer app, which is the difficulty with ordering an open-weights model out of a building. The weights were downloaded a long time ago and they are not in the building.
The Competitor's Shop Window
OpenAI said on 21 October that its frontier models and Codex would be sold through Google Cloud's Model Garden and billed through the Google Cloud Marketplace, with general availability planned "in the coming weeks". Companies with existing Google Cloud spending commitments can spend them on OpenAI. The models are also due in Google Antigravity and in a product called the Gemini Agent. Microsoft, the announcement noted, remains OpenAI's primary cloud partner. Google Cloud already listed more than 200 models. It now also sells the one Gemini was built to beat, from inside something with Gemini in its name.
An Assistant That Asks First
Anthropic opened a waitlist on 21 October for Claude Duet, a research preview for Pro and Max subscribers in the United States, on iOS, Android and claude.ai. Built on Claude Opus 5.5, it checks a user's connected apps every few hours, sends a short morning brief, runs tasks on its own virtual machine with its own browser, and asks before it sends a message, makes a payment, signs a document or deletes anything. Some actions, such as trading securities, stay off limits even with permission. The announcement listed the known limitations itself: personal errands are weaker than work, and Duet cannot yet show the user its browser. It connects to more than 6,000 apps and plugins, which gives it a great deal to ask about.
The 2026 Deadline
Three of the people best placed to know put artificial general intelligence in this calendar year. Dario Amodei wrote in October 2024 that it could arrive as early as 2026. Elon Musk, whose 2025 prediction had passed without comment, revised his estimate in January 2026 to the end of this year. Sam Altman told TIME in August 2026 that he expected to have, by December, an internal system he would personally be willing to call AGI. Note the word internal, which is carrying a great deal of weight, and the word personally, which is carrying the rest.
Metaculus Says June 2028
The forecasting community Metaculus, aggregating more than 1,700 individual predictions as of June 2026, put weakly general AI at June 2028. Shane Legg, who co-founded DeepMind, has given 2028 as his fifty-fifty date since 2011 and has not felt the need to move it in the intervening fifteen years. Consistency is not the same thing as accuracy, but it is rarer and worth more.
The Date Ray Kurzweil Has Never Moved
Kurzweil named 2029 in 2005, when the state of the art could not reliably tell a cat from a dog, and has restated it since without adjustment. Jensen Huang, arriving from a different direction and with a different interest in the answer, also settled on 2029. Twenty four years is a long time to hold a position in public. It will be resolved either way within the decade, which is more than most predictions ever manage.
Hassabis: 2030, Plus or Minus a Year
Speaking at Stanford in June 2026, Demis Hassabis said the field was only a few years away, maybe 2030, plus or minus a year. He has been notably more conservative than his competitors throughout, and has also been running the laboratory that solved protein folding, so the caution is not obviously timidity. Sam Altman, asked around the same period, said end of the decade, which is the same answer with less arithmetic attached.
The Forecasters Settle on April 2033
Metaculus, on the harder definition requiring general capability rather than a narrow set of benchmarks, placed the date at April 2033 across more than 1,900 forecasters as of August 2026. Before the language model wave, surveys of AI researchers put the median at somewhere around 2060. The forecast has therefore moved almost thirty years closer in about four, which is either evidence of progress or evidence of how forecasts behave when everyone is looking at the same news.