The Week in AI
The week in brief
OpenAI published 722 mathematics manuscripts on 6 October, grouped into 372 families of results, all produced by an internal model it has not released.
OpenAI made three announcements on 6 and 7 October. It published a batch of machine-produced math results with many proofs formalized in Lean, rolled out GPT-6 in ChatGPT with answers that can include charts and small tools, and opened a cheap Decisions API for classification work. Anthropic released Claude Haiku 5.5 on 7 October at $0.10 per million input tokens for shorter requests, and on 8 October Google put a planning agent into Gemini Enterprise in private preview.
Two large open-weight models from Western labs were announced on 5 and 6 October, and neither can be downloaded yet. Mistral AI previewed Mistral Large 4 and Reflection AI unveiled Beam, and both say weights follow later in October. On 9 October Microsoft added a cheap model for fixed-choice decisions, Microsoft-Decision-1, at $0.042 per million input tokens. On 11 October Ant Group started a limited beta of LingGuang 2.0, a personal agent with long-term memory.
OpenAI published 722 math manuscripts from an unreleased model
OpenAI put 722 mathematics manuscripts, grouped into 372 families of related results, in a public GitHub repository on 6 October, all produced by an internal frontier model it has not released.
What happened
Each family holds a main result plus supporting arguments or alternative proofs. Many of the proofs come with Lean formalizations, which a proof checker can verify mechanically. OpenAI says more formal proofs will be added.
OpenAI also released 10 detailed summaries of how the model reasoned its way to particular results, along with compute estimates and statistics on the problems it attempted. It estimates the average result used about three hours of ChatGPT Pro thinking compute. Latent Space put the number of problems attempted at about 4,000, and other outlets reported the same figure, which the Atlas did not find on OpenAI's own page.
The release follows OpenAI's earlier Navier-Stokes announcement. According to Ynetnews, this batch covers algebra, number theory, computer science, mathematical logic and topology.
How OpenAI framed the verification
OpenAI consulted the IAS Advisory Group on Mathematics and AI on how to release the batch. It says the results are at different stages of verification and may contain errors.
Reported counts differ slightly. Ynetnews describes more than 700 papers solving 377 previously unsolved problems, and OpenAI's own grouping is 372 result families. The records do not explain the gap of five. Ynetnews also reports that some results relate to three of the Millennium Prize Problems, and no record says any of the three has been solved.
Why it matters
Mathematicians can read and check the work directly, and the Lean parts can be checked by software. That makes this batch easier to test than a benchmark score.
The model itself stays internal. Outside groups can verify the proofs, but they cannot rerun the model on new problems or see how often it failed beyond OpenAI's own statistics. The three hours of Pro thinking per result is the figure to use when someone asks what this costs, and it is OpenAI's estimate.
What we don't know
How many of the 372 families hold up under outside review is unknown. The records do not say what share of proofs are fully formalized in Lean, or when the additional formal proofs arrive.
Mistral previewed Large 4 and Reflection AI unveiled Beam
Mistral AI opened a preview API for the 1 trillion parameter Mistral Large 4 on 6 October, a day after Reflection AI unveiled the 501 billion parameter Beam, and both say open weights will ship later in October.
What happened
Both models are mixture-of-experts designs. The network is split into many specialist blocks and each token is routed through only a few of them, so the active parameter count is the part doing work on any single step.
Mistral Large 4 is Mistral's largest model so far, up from the 675 billion parameters of Mistral Large 3. It is multimodal and runs as a preview API on Mistral Studio. Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, and open weights are due by the end of October.
Mistral's blog gives 1 trillion total and 49 billion active parameters for Large 4. Its documentation lists 1.05 trillion total and 52 billion active.
Beam is Reflection AI's first model release. It is text-only, aimed at coding and agent tasks, and has 501 billion total and 23 billion active parameters with a 1 million token context window. Reflection says it pretrained Beam on 23.8 trillion tokens and then trained it heavily with reinforcement learning (RL). Beam is a research preview, and the weights and a technical report are due later in October.
What the two labs claim
Mistral says Large 4 significantly outperforms any open-weight model from the US or Europe and is competitive with the strongest open models globally. The second phrase points at the Chinese open-weight models, which Mistral does not claim to beat. Mistral did not publish scores the Atlas could check on 6 October.
Reflection compared Beam with Z.ai's GLM-5.2, which has about 744 billion total and 40 billion active parameters. Reflection says Beam matches GLM-5.2 on reasoning benchmarks while using 3 to 4 times less inference compute, and nobody outside the company has checked that. Beam has a little over half of GLM-5.2's active parameters, which does not by itself account for a 3 to 4 times saving, so the technical report will have to show where the rest comes from.
Why it matters
Both labs measured themselves against open models from China. For a practitioner picking an open model at this size, Large 4 and Beam are the European and US entries for October. StepFun added a Chinese model at this size on 8 October, Step 5 Preview, with 600 billion total and 27 billion active parameters, and it is available only through an API.
Until its weights ship, Mistral is red-teaming a version of Large 4 with reduced moderation together with cybersecurity partners. Anthropic, OpenAI and Google keep the weights of their top models.
What we don't know
Which of Mistral's two parameter counts is correct is unknown, and so is whether the released weights will be the moderated version or the one being red-teamed. The benchmark scores behind both companies' comparisons are not in the records. It is also unknown how Beam compares with GLM-5.3, the newer model in the same family.
OpenAI put GPT-6 into ChatGPT with Intelligent UI
OpenAI rolled out GPT-6 in ChatGPT on 7 October with a feature called Intelligent UI, with paid tiers running on Sol and free users on Luna.
What happened
With Intelligent UI, a ChatGPT answer can mix text with charts, diagrams, forms, tappable buttons and small interactive tools such as a calculator. The model decides which format fits the question. OpenAI built a library of streamable interface components and a compiler that draws the interface while the model is still generating.
The model can also start answering while it keeps thinking and send partial answers. OpenAI says GPT-6 resists attempts to bypass its safety training better than GPT-5.6, and the Atlas has no independent test of that. OpenAI says the model is built for ChatGPT's more than 1.2 billion weekly users.
The Decisions API and Ultrafast
On 6 October, a day earlier, OpenAI opened the Decisions API in public beta. It takes text, images or both and returns one of three typed answers, which are a probability that a condition is true, a choice from a fixed set, or a score against a rubric. That suits moderation, routing and grading jobs that currently go through a general chat endpoint.
The only supported model is gpt-6-luna, at $0.10 per million input tokens with no output charges or separate cache charges. OpenAI says it runs about 10 times faster than the Responses API and gives no benchmark for that. On 8 October OpenAI also added an Ultrafast service tier for GPT-6.1 Sol in the Responses API, which shortens the time between output tokens and is open to all API users within rate limits.
Microsoft released a model for the same job on 9 October. Microsoft-Decision-1 is a post-trained Qwen3.5-9B in Foundry that returns a calibrated probability for each of a fixed set of options, at $0.042 per million input tokens with output free. Microsoft reports a median latency about 35 times lower than GPT-6 Sol.
Why it matters
Free ChatGPT users got GPT-6 Luna two days before Google limited free Gemini users to Flash Lite on 9 October. Anyone comparing the free versions of the two products on stage is now comparing different tiers than a week earlier.
Intelligent UI hands interface choices to the model. A reply can now be something the user taps or fills in, so a demo or screenshot of a GPT-6 answer depends on what the model chose to render for that prompt.
What we don't know
The records name three GPT-6 variants, Sol, Luna and Astra, and do not give their parameter counts. OpenAI published no benchmark scores for GPT-6 in ChatGPT in the records the Atlas holds. Usage limits per tier and the Ultrafast price were not in the sources.
Anthropic released Claude Haiku 5.5 with adaptive thinking
Anthropic released Claude Haiku 5.5 on 7 October, its smallest model, with a 1 million token context window at $0.10 per million input tokens and $0.50 per million output tokens for requests up to 100,000 tokens.
What happened
Haiku 5.5 succeeds Haiku 4.5. It has a 1 million token context window and can write up to 128,000 tokens in one response. Anthropic says it is its fastest model, aimed at summaries, classification, browser use and subagent coding work alongside Opus 5.5 and Sonnet 5.5.
It is the first Haiku with adaptive thinking, where the model decides how much to reason within an effort setting the developer picks. Requests longer than 100,000 tokens cost $0.50 per million input tokens and $2.50 per million output tokens. Anthropic says the new model is about 75% cheaper than Haiku 4.5 on average.
What Anthropic reports
All of these scores are Anthropic's. Haiku 5.5 scores 39.2% on Terminal-Bench 4.0, where Haiku 4.5 scored 0.0%. On the offline subset of OSWorld 2.1, it scores 72.4% against 15.7% for Haiku 4.5.
Anthropic reports 45.9% on Humanity's Last Exam without tools, up from 10.2%. On GDPval-AA v2.1 it reports a score of 1620, against 735 for Haiku 4.5.
Anthropic also halved cache read prices for Sonnet 5.5, which launched on 28 September. It added a monthly API credit for Max and Team subscribers.
Why it matters
Terminal-Bench measures work in a command line and OSWorld measures operating a desktop, and both are the kind of work a subagent gets handed. Haiku 4.5 scored near zero on the first and 15.7% on the second.
The input price matches OpenAI's Decisions API, which also charges $0.10 per million input tokens and launched the day before. OpenAI charges nothing for output and Anthropic charges $0.50 per million output tokens. Microsoft-Decision-1 came in below both on input price two days later.
What we don't know
The records do not say how Anthropic calculated the 75% average saving or which usage mix it assumed.
Google previewed a Gemini Enterprise agent that delegates to subagents
Google announced a universal agent in the Gemini Enterprise app on 8 October, in private preview for enterprise customers, that takes an objective, plans the work and hands parts of it to subagents.
What happened
Until now, Gemini Enterprise was a place for companies to build and govern their own agents. The new agent takes a stated goal, plans the steps itself, and delegates pieces to job-specific subagents that use custom skills and tools. It picks a model automatically, and users can choose another one, including Anthropic's Claude.
The agent has its own Workspace account with an email address, so its actions appear in the audit trail under the agent's name. It connects to Workspace, Microsoft 365, Slack, Jira, Snowflake, BigQuery and any server that speaks the Model Context Protocol (MCP), and users follow its progress in a tasks inbox.
Google Cloud's post from its Gemini at Work 2026 event adds that the agent can run on schedules or event triggers and keeps working in the cloud when a laptop is closed. It lists role-based permissions, sandboxing, an Agent Gateway and spend caps per project among the controls.
Sundar Pichai, Google's chief executive, said at the event that Gemini has over 1 billion monthly active users and that nearly 90% of Fortune 100 companies use Gemini Enterprise. Both figures are Google's own.
Why it matters
Companies deploying agents keep asking who did what. Giving the agent its own account means its actions are logged under its identity, so an administrator can separate them from the employee who set the goal.
Google also lets customers pick Claude inside its own product. On 11 October, Ant Group's LingGuang 2.0 beta showed a consumer version of the same idea, according to Pandaily, which cites people close to the project. It is built to remember users across sessions, move tasks forward on its own and call Skills and MCP tools. Ant has not announced it, and the sources say a full rollout is planned for October.
What we don't know
Google gave no price, no general availability date and no count of preview customers in the records. The records do not describe which model the agent picks by default or how it decides.
Also worth knowing
Security and provenance
- Anthropic reportedly expanded its Cyber Verification Program on 6 October and folded Project Glasswing into it, with three access tiers for vetted security professionals, six days after Google gave Gemini 4 Argon to cyber defenders first.
- Project Glasswing found more than 5,500 verified vulnerabilities at Anthropic from April to October, and partners found more than 129,000 from April to July, according to company figures reported by Reuters and summarized by Techmeme.
- Critical Infrastructure Defense Program is Anthropic's program launched on 8 October with 11 founding partners including CrowdStrike, Dragos, Hitachi and Rockwell Automation, giving defenders of grids, water utilities, factories and transport networks frontier Claude models, threat research and on-site engineers.
- OSS Scanner is reportedly a free opt-in Anthropic service launched on 8 October that scans open-source projects with models including Claude Mythos and sends reports with no human review. Anthropic reports that expert reviewers found 85 of 97 critical and high-severity findings met the bar for coordinated vulnerability disclosure.
- Goodfire launched monitors for Baseten customers on 8 October that read a model's internal activations with small classifiers called probes. It reports catching 94% of malicious hacking sessions, and monitoring about 1,500 Kimi K3 sessions cost about $51, against about $10,000 for a top-tier model checking every step.
- SynthID Detector is a public Google site opened on 7 October that checks images, video and audio for SynthID watermarks from Google, OpenAI, Nvidia and Kakao, with about 10 checks per user per day and Apple support promised later.
- textGrain is OpenAI's statistical text watermark, announced on 5 October for the EU AI Act. OpenAI reports its detector catches about 95% of 400-token passages at a 1% false positive rate but only 17% once a quarter of the words are replaced.
- Satya Nadella, Microsoft's chief executive, said companies should assume all AI models are compromised, The Verge reports.
Open and on-device models
- Qwen-Image-2.1-Turbo is an open-weight checkpoint of Alibaba Qwen's 7 billion parameter image model, released on 9 October under a restricted license. Qwen says it generates and edits images in 8 denoising steps.
- Mellum2.1 is an open 12 billion parameter coding model from JetBrains with 2.5 billion active parameters, released on 8 October and trained mainly with RL. JetBrains reports almost twice the throughput of Qwen3.5-9B under heavy load.
- EmbeddingGemma 2 is an open 740 million parameter embedding model from Google DeepMind, released on 6 October and built on Gemma 4, that maps text, code, images, video and audio into one 768-dimensional space. Google reports its text-only part runs in about 191MB of active RAM on a Pixel 11 Pro.
- Google AI Edge Foresight is a free experimental Mac app released on 8 October that transcribes meetings and writes notes offline with on-device models including EmbeddingGemma 2.
- pplx-embed-v2-late is reportedly a pair of open-weight multimodal retrieval models from Perplexity at 0.6 billion and 9 billion parameters under the MIT license, and Perplexity reports 92.4% on MADQA for the larger one.
- DeepSeek V4 Flash will run locally on RTX Spark PCs quantized to 1.6 bits, Microsoft said at its 7 October Windows keynote, alongside a Nvidia Nemotron of more than 70 billion parameters and llama.cpp support in Windows ML.
- RTX Spark is an Arm-based Windows platform from NVIDIA and Microsoft with up to 128GB of unified memory, announced on 7 October. The Surface Laptop Ultra on it starts at $2,599 and ships 16 October.
- Falcon-Emirati-7B was reportedly announced by the Technology Innovation Institute, a model built on Falcon-H1-Arabic and specialised in the Emirati Arabic dialect and culture.
Products and pricing
- Gemini's free tier dropped to Flash Lite only on 9 October. Standard Flash moved to the $4.99 per month Google AI Plus plan, and Gemini Pro and Deep Think now need Google AI Pro at $19.99 or Ultra at $99.99 per month.
- TRAE is ByteDance's merged app, launched on 9 October, which combines the TraeWork office agent and TraeCode coding apps into one client with an Agent mode and an IDE mode. TraeWork stops service on 23 October.
- Doubao Work got an infinite creation canvas on 9 October, powered by ByteDance's new Seedream 5.0 Flash image model, alongside Doubao 2.1 Lite for office tasks.
- Vidu Q4 Preview is ShengShu's video model released on 8 October, with output up to 4K, clips up to 16 seconds and up to 15 reference images, at a launch price from RMB 0.09 per second.
- ChatGPT image ads were announced by OpenAI on 5 October, with labeled ads next to image generation results for Free and Go users and US testing later in October.
- Nano Banana 2.1 is Google DeepMind's new Flash-tier image model, listed on OpenRouter at $1.50 per million input tokens, $7.50 per million output tokens and $30 per million image output tokens.
- gpt-rosalind-discovery appeared on OpenAI's API pricing page on 10 October under life sciences models, at $5 per million input tokens and $25 per million output tokens. Access is limited to approved organizations, and OpenAI has not said how it differs from gpt-rosalind-research.
- Liquid Inference from Architect Financial Technologies auctions each request among providers serving the named model and charges the lowest offer that meets the buyer's limits on price, latency and region.
- Playground is a Google Labs service launched in the US on 7 October that builds playable browser games from text prompts.
Agents
- Claude Managed Agents reportedly gained dynamic workflows in beta on 9 October, so an agent can write a program that runs many agents in phases while the server runs it in the background.
- openJiuwen, the Huawei-backed agent platform, released AgentOS for Enterprise on 9 October as Apache 2.0 source code with no model weights, bundling an agent runtime, work and coding agents, a sandbox and an access gateway.
- Personal Agent Protocol is an open standard published on 6 October by Meta with Walmart, Stripe, Sierra and others to help websites tell legitimate agents acting for a user from malicious bots.
Research and evaluations
- OpenAI reports that GPT-6 Astra scored 55.0% on 11 contracting tasks built with Ironclad, against 41.6% for GPT-5.6 Sol, on a page that carries no publication date.
- Vals AI reports that a team of Claude Opus 5.5 agents found two candidate room-temperature antiferromagnetic semiconductors for computer memory, predicted by calculation and not yet tested in a lab.
Money
- Qualcomm is reportedly in advanced talks to buy the AI startup Modular for $4 billion, according to Stocktwits.
People this week
- Rohit Prasad, who led Amazon's AGI organization and the Nova models, will become CEO of Boston Dynamics, the Hyundai Motor Group robotics company, announced on 7 October.
- Luo Fuli, who leads Xiaomi's MiMo model team, was reportedly promoted to vice president level, according to 36Kr.
What did not change
The top models from Anthropic, OpenAI and Google remain closed. The two large Western open-weight models of the week, Mistral Large 4 and Beam, are still previews without downloadable weights. OpenAI's math model is still internal, and the Gemini Enterprise agent has no price or general availability date.
What to watch next
- Open weights for Mistral Large 4 by the end of October, and whether they match the moderated or the red-teamed version.
- The weights and technical report for Reflection AI's Beam, which should show where the claimed 3 to 4 times inference saving comes from.
- Outside review of the 372 OpenAI math result families, and the additional Lean proofs OpenAI has promised.
- 15 October, when the Nvidia Nemotron of more than 70 billion parameters is expected on RTX Spark PCs, and 16 October, when the Surface Laptop Ultra ships.
- 23 October, when TraeWork stops service, and the full rollout of Ant Group's LingGuang 2.0 planned for October.