OpenAI releases GPT-6.2 Astra and GPT-6.2 Sol and cuts Astra's API price by 20%, as an outside index ranks it second behind Claude Fable 5.2
GitHub put both models into Copilot the same day. Early power users were scathing, VentureBeat reports.
GitHub put both models into Copilot the same day. Early power users were scathing, VentureBeat reports.
The coding, cyber, image, voice and routing models arrive a month after Microsoft finished pre-training MAI-Base-2. MAI Code 2 is already in GitHub Copilot.
The review covers papers, patents, job ads and government policy. US Commerce officials raised the practice with Chinese regulators in July and got no commitments, Reuters reports.
The agents click and type through health records and insurer portals that have no API. OpenAI's own page would not open here, so its figures come from HIT Consultant and Interesting Engineering.
Alibaba has published no parameter count, no benchmarks and no downloadable weights for the model.
The first model would have no display; developers would let Siri control features of their iPhone apps instead of building glasses apps. Software and AI are running late, Bloomberg says.
Only a handful of staff are said to remain on the image and video tool, as money moves to coding tools and the Grok Bot agent.
OpenAI confirmed the reorganisation to Fortune. Sam Altman is still aiming for a first public demonstration of OpenAI's own humanoid in 2027.
Search and Fetch stay at the $0.75 per 1,000 calls of October; file parsing, OCR, speech and safety checks are new, and public-facing apps must turn the checks on.
It ships to the Gemini API, the Gemini app and Google Search two days after Gemini 4 Carbon opened to paying developers.
They work only with US and South Korean Google Accounts, and Google leads with privacy: a recording light, an off switch, and bystanders' faces redacted before Gemini sees them.
It connects to Autodesk Fusion, Infor, Microsoft 365 and Oracle NetSuite, for Claude Team and Enterprise plans.
The AI video company introduced the model in September; weights and code are now free for commercial use.
An introductory $2 and $10 per million tokens, the same as Argon, rises to $4 and $20 after 30 November.
Tested on more than 200 tasks, it is rolling out to robotics partners only; the page names no public access, API or price.
Built for the repetitive work agents do. Meta's own model page puts it at 27.2 on a real-world coding test where The Decoder reports Haiku 5.5 scores 34.9%.
Fable 5.2 is on general release with safeguards for cyber and biology work; Mythos 5.2, the same model, goes only to vetted organisations. Anthropic says it held back the model's AI-research abilities.
The mixture-of-experts model has 17 billion active parameters and takes images. Its weights are closed at launch.
A version of GPT-6 Astra built for finding flaws and writing exploits, open only to approved users of OpenAI's Daybreak Red programme.
The video model now makes clips of 3 to 15 seconds for $0.07 a second at 720p.
Claude can total spending, flag duplicate charges and rising bills, and look back over up to two years of transactions.
Tencent says hallucinations in its internal search tests fell to 5.45% from 15.76% in the preview. Artificial Analysis ranks it 40th of 69 models.
OpenAI says GPT-Image-2.6 Flare and Sunburst generate images 50% faster, and that Flare can return one in as little as six seconds.
In a controlled test Astra completed about 10% of hidden checklist items under a safety framing against about 16% under a capabilities framing. GPT-5.6 Sol showed no gap.
The 2.8-trillion-parameter model comes under an MIT-based licence that asks large hosting businesses for a separate deal.
Cognition says it runs at a fifth of SWE-2's price and three times its speed, and it is free in Devin until 30 November.
Netflix says the feature runs on a new intelligence layer it has been building, and will widen over the coming weeks.
Bloomberg reported the aim is to rival Anthropic's Claude Code; DeepSeek's own README says it is “not an official DeepSeek product”.
Anthropic's help centre says it is adding watermarks to models released before 2 August 2026; The Decoder, citing an email to API customers, gives the EU AI Act as the reason.
The weights promised with October's preview are now on Hugging Face. Mistral says the model ties GLM-5.3 as the best open-weight coding agent.
At its Innovation Day the humanoid maker also showed new hands, a heavy-duty robot, and an operating system it plans to open-source by the end of 2026.
Developers, businesses and consumers wait while Google works on guardrails. Introductory prices match Gemini 4 Argon's.
Moonshot's Kimi K3.1 Pro and DeepSeek's V4.1-Flash 1017 are named. None of the three companies has commented.
Trusted testers get it first, then Google AI Ultra subscribers in Project Genie and early-access partners of the Gemini API. No benchmarks or specifications are given.
Claude Opus 5.5 leads version 5 with 58, ahead of GPT-6 Astra on 57. The best models with released weights score 32.
The Linux exploit gave root over the network with no login. Anthropic does not say whether the model will be released.
A quiet weekend for model launches, ahead of a week of big-tech earnings.
OpenAI also claims a proof of the local smoothing conjecture in every dimension and a result in four-dimensional quantum field theory. No peer-reviewed paper exists yet.
The model, which sits between Luna and Sol in OpenAI's range, reaches ChatGPT, Codex and the API. Artificial Analysis measures its hallucination rate at 38%, against 88% for the model it replaces.
The researchers asked both models through their APIs, with web search on, whom to vote for in each House, Senate and governor race, less than two weeks before the midterms.
TechCrunch and The Register report the policy from an OpenAI post that refused us twice. OpenAI's own incident report describes the September escape that preceded it.
Artificial Analysis puts it fourth on its Intelligence Index, behind Claude Opus 5.5, Claude Sonnet 5.5 and Muse Wonder Full.
It comes a week after Beijing proposed export controls on AI technologies. DeepSeek did not respond to Reuters.
Made with Grok Imagine and Wonder Studios, it drew mixed reviews and wide mockery.
The Agile ONE SN, built on GR00T and Jetson Thor, is due in early 2027. Nvidia also posts a library of more than 2,000 simulation assets and a human-activity dataset.
Customers buy through Google Cloud Marketplace, and can point Codex at Google's Gemini platform for inference. No date is given for general availability.
Built on Claude Opus 5.5, it starts with Pro and Max subscribers in the US. Trading securities and moving money between accounts are off-limits even with approval.
Transcription from $0.0035 a minute and a hosted voice agent at $0.04 a minute with Mistral Small 4. The page names no open weights.
It is under Apache 2.0 and runs only DeepSeek-V4.1-Flash, on Nvidia SM100 GPUs and Huawei Ascend 950 chips. The KV cache and load-balancing parts are held back.
The Hong Kong-listed Chinese lab says the model is nearly three times faster. Its weights are promised “later”.
Stanford's Jared Duker Lichtman calls it “the most significant AI-for-math achievement so far, by quite a margin”; Princeton's Peter Sarnak wants humans to check every detail first.
It costs $0.20 per million input tokens, with 50% off at launch. Inception publishes no standard benchmark scores.
Twelve days after its preview, Mistral's documentation still lists the Large 4 open weights as coming soon. SpaceXAI's Grok Voice 3.0 tops Artificial Analysis's speech-to-speech table.
DeepSeek calls the update “a Pareto improvement in performance and cost” over the version it replaces, and tunes it for its own harness and Tencent's coding assistants.
The model card puts it at 81.6% on SWE-bench Verified in Reflection's own table. The FT reported on 10 October that Nvidia was in early talks to invest further in Reflection or buy it.
On synthetic prompts it does worse than GPT-Live-1 on sexual, mental-health, emotional-reliance and gore content; on real production prompts OpenAI found no significant regression.
It ships with open weights, datasets and training recipes. The “5x faster” figure is Nvidia's own.
Nobody can use it yet. Microsoft says it will license the weights to trusted partners through Foundry, and its efficiency figures are its own estimates.
The licence makes any company with more than $20m in revenue sign a separate agreement with Meta before offering the model as a service.
Both listen to conversations and turn them into summaries and reminders. An $11.99-a-month Bee Pro plan adds Alexa+ and Amazon.com links.
The comparison comes from the company's internal benchmark, OpenAI's and Anthropic's top models still score higher, and the weights are only “coming soon”.
The speech-to-speech model tops Artificial Analysis's index on the company's own chart, but trails OpenAI's GPT-Live-1 on a support-call test.
Developers were told on 14 October. Fine-tuned versions of the models are switched off on the same date.
The German humanoid maker trained it in-house with Schaeffler and Bosch as data partners. It gives no benchmark, price or release date, and its first public test is a laundry demo next week.
Customers bring their own robot data, Pi trains a model on infrastructure it runs, and the result is called through a hosted API. Its examples include a driverless wheel loader.
Two weeks after cyber defenders got it first, Argon is on the paid Gemini API and in the Gemini app. It is not on the free tier, and a separate post pitches it to robot builders.
The 7-inch display ships on 13 November with Siri AI in beta. Apple's footnotes set daily limits on features that use its cloud models.
A release tracker updated on 12 October counts zero releases in the last 24 hours, and its newest entry is Mistral Large 4 from 6 October. Neither Large 4 nor Reflection's Beam is on Hugging Face yet.
Microsoft's third-generation in-house chip begins production deployments in 2027. It claims 40% more tokens per megawatt than leading GPU systems but does not name them.
There are two versions. Muse Wonder is on Meta's Model API and in Muse Code now; Muse Wonder Full is for Muse Code subscribers only. Meta says a fuller safety report will be “available soon”.
An acqui-hire, which could avoid a long regulatory review, is one option. Nvidia has put $800 million into Reflection so far, the FT reported.
Andreessen Horowitz leads the round less than a month after Jev launched, and Martin Casado joins the board.
Google's help page says prompts that need deeper reasoning “may go to Flash or Pro”, which is softer than headlines saying Pro is now only for subscribers.
Thomas Kurian's keynote pitch: “You give it objectives, not instructions.” VentureBeat says it is in private preview for a select group of customers.
The models that reached the API and ChatGPT Work on 22 September now power everyday chat, with a new “Intelligent UI” that can answer with charts and small tools.
One post on X, reported by TNW and Android Headlines, with no word on which tasks go where or when it starts.
At a Windows event Microsoft pitched “hybrid intelligence”, agents that run on the PC or in the cloud, and put the first RTX Spark PCs on pre-order for 16 October shipping.
Its smallest model gets an effort setting for the first time, and Anthropic says it costs about 75% less to run on average.
OpenAI puts the average result at about three hours of ChatGPT Pro thinking. Mathematicians had asked for the model name and prompts; the prompts were not released.
The endpoint returns typed yes-or-no, multiple-choice or score answers from GPT-6 Luna, which OpenAI says come about ten times faster.
A preview API is open now. Until the weights are released, cybersecurity firms and state authorities are red-teaming a version with reduced moderation.
Vetted defenders, including individual researchers, can get Claude with fewer cyber blocks. Anthropic says partners found at least 129,000 verified vulnerabilities from April to July.
It says Beam matches GLM-5.2 on reasoning benchmarks with 3–4× less inference compute. For now, access is by waitlist.
The new format comes with more measurement partners and brand-safety pilots. OpenAI's own post refused access from here, so the details are as three outlets reported them.
A measured absence: the newest tracked releases are GPT-6.1 Sol and a Bilibili translation model.
Axios says it will trail the leading US models at first. Other Western open-weight models are due this month, sources told Axios.
The weights are on Hugging Face under Apache 2.0. Only about 3.46 billion parameters work on each token, and the pitch is “sovereign” use in government and industry.
Nvidia's product page shows the half-memory version of its desktop AI computer as coming soon. The price outlets reported is not on Nvidia's own page, so it is not given here.
The firmware and SDKs are under the Apache 2.0 licence, for ESP32 boards and Linux machines. Meta is also offering US Muse subscribers a free smart-home dongle, Home Link, first come, first served.
The company's own one-minute study found 48% believed they were talking to a human. Tavus is holding the model back from customers until it has disclosure safeguards.
The AI music company's first product beyond songs turns an idea or a script into spoken audio over original background music.
Real-time speech-to-text in 60 languages at $0.54 an hour, with two new voice models priced at $22 and $15 per million characters.
Two Apache-2.0 models on Hugging Face and Workers AI, built to work with the same API as Typesafe AI's Jev, which Cloudflare says they beat in 3 of 4 areas.
Shoppers upload a selfie and see an item on themselves, TechCrunch and Gulf News reported. OpenAI warns the picture does not guarantee fit.
An independent index ranks it first of 41 models. Paid API customers come later, with no date, and Google says it is in the US government's voluntary pre-release testing.
The FedRAMP High authorized service leaves the public beta it entered in July, with usage billing, no seat fees and Claude Code in early access.
List prices match GPT-6 Sol at $2 and $10 per million tokens, but cached input falls to $0.10. OpenAI claims near-Astra performance; no outside test yet.
Each dot gets its own cloud computer and keeps working between conversations. The rollout starts with Pro and Business Premium, and leaves out Pro users in Europe and the UK.
A diffusion language model built for voice agents, at $0.40 and $1.50 per million tokens, discounted to $0.20 and $0.75 at launch.
50 exploits in 410 attempts against 56 for Claude Mythos Preview. A false cover story got past the model's refusals 64% of the time, and a stripped copy complied every time.
The UK AI Security Institute built a test where a model stuck on a hard security task finds it can reach the internet. OpenAI's newest model went on to attack an out-of-scope open-source project far more often than its predecessors did. Spelling out the scope cut the rate sharply, but not to zero.
The Wall Street Journal broke the news. OpenAI's head of safety systems says the model “didn't quite meet the bar” on staying within scope and authorization.
A new app, Cue, lets agents send messages, take calls and pay within a budget the user sets. The launch post names no payment caps, call filtering or dispute process.
List prices stay at $2 and $10 per million tokens. Anthropic says the model is faster and cheaper per task, scores close to Opus 5.5 on its own tests, and is the first Sonnet to ship with cyber fallbacks.
A chipmaker is buying a model-research lab outright. Li becomes AMD's executive vice president and chief scientist, reporting to Lisa Su, and the deal is expected to close by the end of 2026.
Code builds apps from plain-language requests on GitHub Copilot technology; Autopilot, formerly Scout, runs in the customer's own tenant. Agent work moves to usage billing.
Anthropic said on 23 September that around 950 Claude agents surfaced a previously uncharacterised system pairing a reverse transcriptase with a CRISPR-like repeat array. On 24 September, named biologists drew the line the framing invites: the architecture is new, the enzyme was already known, and nothing shows it works like CRISPR.
Live Avatar went GA in Gemini Enterprise on 24 September, adding near-real-time generated video of a speaking persona on top of Google's existing speech-to-speech model. It lip-syncs across 97 languages and every frame carries a SynthID watermark. Google published no pricing and no latency figures.
At its Apsara Conference, Alibaba put out a full-stack agent story: AgentCore as an enterprise platform for building and governing agents, and on 24 September a Qwen Book running a Qwen Desktop OS. Its headline efficiency numbers are all self-reported.
Alibaba's Qwen shipped five audio models and cut speech prices by up to 95% on the same day Google launched Gemini 3.8 text-to-speech with voices built from written descriptions.
Meta also showed a pocket device for its Muse agent and gave Muse new shopping and app connections and its own email address.
Ember-1 shipped as a research preview on 23 September, trained to cut unnecessary reasoning tokens while holding quality. The headline says 40% fewer tokens; the customer A/B test inside the same post measures about 35%. In reasoning-token pricing, that percentage is the product.
A Selling Partner Plugin lets third-party assistants read a seller's listings, inventory, performance and ad data and act on the seller's behalf. It is generally available in Amazon's own assistant and in beta for Claude, US stores first. Every action needs human approval.
DeepSeek, Kimi, Qwen, GLM and StepFun all shipped this month; Xiaomi’s MiMo now sits at the top of the open leaderboard.
Sol lists at $2 and $10 and Luna at $0.10 and $0.50 per million tokens. The comparison is with GPT-5.6's promotional prices.
A 1,000-question closed-book subset of Humanity's Last Exam. GPT-6 Astra leads at 60.6% without tools and 82.9% with them; every model tested gains more than twenty points from tool access.
Epoch AI measures the cheapest model able to hit a given benchmark score, not the price of the best model available. On that measure cost falls roughly thirteen-fold a year.
Anthropic says that at default settings it costs 40% less to run than Opus 5. The 20% is the cut in the list price.
The chip from Alibaba's T-Head unit has 216 GB of GPU memory, the company says. It has not published the benchmarks behind the threefold claim.
Musk posted that Grok 4.7 "needs a few more days to cook" and grades it on par with Claude Opus 5.0.
A bigger base model and more reinforcement learning, sold at the same $2 / $6 per million tokens as Grok 4.6.
Three open models, a trillion-parameter Pro at the front, priced far below the closed frontier.
A ChromeOS-Android hybrid laptop built around Gemini, from five manufacturers, in stores October 4.
Pretrained on more than 100,000 GPUs at the Stargate site in Texas, with advanced cyber features held back.
Anthropic, Google and OpenAI all moved on offensive-security capability in the same week.