Morning edition — Friday 23 October 2026

MiniMax has published the weights of M3.1 Flash on Hugging Face, four days after launching the model through its API on 19 October, when it said weights would follow. The model card describes a natively multimodal mixture-of-experts model with about 428 billion total parameters and 23 billion active per token, text, image and video inputs, and a context window of up to 1 million tokens, aimed at coding, agentic work, computer use and cybersecurity. The weights come under MiniMax's own community licence rather than a standard open-source licence.

MiniMax reports scores such as 73.7 percent on DeepSWE v1.1 and 61.7 on OSWorld V2, and says that compared with its previous model, online inference is nearly three times faster and agentic usage costs 70 percent less. Its API price is $0.15 per million input tokens and $0.60 per million output tokens. The model card recommends serving it with SGLang, vLLM or TokenSpeed.

Separately, Yicai, citing the state-media-linked account Yuyuan Tantian, reported late on 22 October that DeepSeek and MiniMax have completed registration under temporary pre-export procedures that China introduced while it revises its catalogue of restricted technology exports, a draft that lists 23 AI technologies including large models and agents, and that both have completed open-source and overseas licensed use.

OpenAI made the Decisions API generally available on 22 October, after a beta that began on 6 October. Built on GPT-6 Luna, it answers bounded questions from text, images or JSON: the probability that a condition is true, a choice from a fixed list, or a score against a rubric. OpenAI says it is up to 10 times faster than calling GPT-6 Luna through the Responses API.

Since the beta, OpenAI has added a native integration with its GPT-Live voice model, prompt caching, US and EU inference residency, JSON and image inputs, and up to 400 questions per request, twice the previous limit. It costs $0.10 per million input tokens and $0.01 per million cached input tokens, with no charge for output.

OpenAI quotes customers including Box, Datadog, Ramp, Sierra and Speak. Datadog, for example, says it uses the API to score the risk of agent-generated shell commands before running them, with 40 times faster p95 latency than other frontier models at comparable accuracy.

The Agentic AI Foundation, hosted by the Linux Foundation, announced four new projects on 22 October at its AGNTCon + MCPCon North America conference in San Jose: Agent Skills, contributed by Anthropic; Open Responses, contributed by OpenAI; OpenEnv, contributed by Hugging Face; and OpenSandbox, contributed by Alibaba. The foundation now hosts 10 projects, up from three at its launch in December 2025, including the Model Context Protocol, AGENTS.md and goose.

Agent Skills is the format for packaging instructions, scripts and resources that agents load on demand. Anthropic says more than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot and VS Code, support it, and that about 1 million skills have been published and nearly 280 million installed. It is donating the name, logo, GitHub organization and agentskills.io domain; engineers from Anthropic, Google, OpenAI and other companies will maintain the standard as individuals.

Open Responses is an open specification, based on OpenAI's Responses API, for calling models from different providers through one shared schema. OpenEnv is a framework and hub for reinforcement-learning environments, and OpenSandbox is a sandbox platform with Docker and Kubernetes runtimes for coding agents, GUI agents and AI code execution.

Anthropic published research on 22 October, done with the formal-verification company Theorem, in which Claude wrote proofs about the compiled RISC-V binaries of open-source libraries tested by Google's OSS-Fuzz. Over a few weeks, Anthropic says, Claude patched the bugs it found and then proved that 239 fuzzing entry points across 123 libraries never crash on any input up to 64 MB. The Rocq proof checker, not the model, decides whether each proof is valid.

The finished proofs total about 180 million lines of Rocq, with a typical proof of about 600,000 lines for a binary with a median length of 18,000 RISC-V instructions; the appendix says Claude wrote around a billion lines along the way. Anthropic says Claude also found bugs that had escaped years of fuzzing. It is publishing a subset of the proofs with the RISC-V and Linux semantics, and Theorem is offering to run the pipeline for other open-source projects and publishing a Claude skill for developers.

On a set of 36 small verification tasks, Anthropic says Opus 5.5 proves half of them with about 40 percent fewer tokens than Opus 5, and that agents sharing a proof workspace needed 17 percent fewer tokens than agents working alone.

A peer-reviewed paper by Microsoft AI researchers, published in Science on 22 October, describes MAI-DxO, a system that works through a case like a panel of virtual doctors: it starts from a short presentation, asks questions, orders tests and weighs each test's cost before committing to a diagnosis. The researchers built a benchmark, SDBench, from 304 New England Journal of Medicine clinicopathological cases and 325 emergency-department admissions from the MIMIC-IV-ED dataset.

Running on GPT-5, the system reached 83.2 percent accuracy on the NEJM cases at an average estimated test cost of $5,431, against 36.1 percent for generalist physicians with search and reference tools and 22.3 percent without them. On the emergency cases it reached 80.3 percent, against 46.8 percent for physicians. On the NEJM cases, budget-limited settings kept accuracy near 80 percent at less than half that cost, and Microsoft says the gains held with newer models including GPT-5.5.

Microsoft is open-sourcing SDBench for selected NEJM cases and says a prospective study with Beth Israel Deaconess Medical Center in Boston will have patients interact with the system directly, with results measured against physicians' independent assessments.