AI News

AI news brief: Tuesday, November 17, 2026

35 stories, most important first.

OfficialModels · · Updated

OpenAI releases GPT-6.2 Astra and GPT-6.2 Sol in its API

OpenAI's API changelog adds GPT-6.2 Astra for demanding reasoning and coding and GPT-6.2 Sol for everyday coding and agent work.

Why it matters: Teams that build products on OpenAI models now have two new options, and any integration that relies on tool calling has to run on the Responses API.

6 sources
ReportedResearch · · Updated

Google says Gemini 4 Carbon powered agents that verified Signal's cryptographic code

DeepMind says Gemini 4 Carbon powered agents that verified Signal's cryptographic code, a result it presents with wet-lab and protein findings.

Why it matters: Models are being applied to formal verification and scientific checking. That suggests verification may be a practical use for teams beyond code generation.

2 sources
ReportedProducts · · Updated

OpenAI's Astra for Law ranks first in a 24-area legal evaluation at launch

According to OpenAI, Astra for Law ranked first in a launch evaluation spanning 24 areas of law, and every ChatGPT Enterprise customer can now request access.

Why it matters: Legal teams at companies on ChatGPT Enterprise can now request access to a product OpenAI positions for legal work. The ranking is OpenAI's own result, so buyers should test the tool on their own matters before relying on it.

2 sources
ReportedSafety & Security ·

Google says hackers built an AI tool that turns security fixes into attacks

Google says attackers built an AI tool that turns newly released security fixes into attacks before companies can install them, with no human involved.

Why it matters: If Google's account holds, the gap between a fix being published and a patch being installed becomes a direct attack surface. Slow patching now carries more risk than before.

1 source
ReportedResearch ·

Physicist Suvrat Raju says GPT-6 Astra solved nearly 90% of his physics benchmark

Raju says GPT-6 Astra answered nearly 90% of the unpublished theoretical physics problems in TheoPhysEval, which he now considers saturated.

Why it matters: Expert-written research benchmarks can saturate within months of release. Buyers who rely on public tests to judge research-grade AI may be measuring less than they assume.

1 source
ReportedResearch ·

OpenAI model's 1,147-page proof on pure Yang-Mills checked in Lean, panel says

OpenAI says one of its models proved the four-dimensional gluon mathematics behind pure Yang-Mills, and a panel found the Lean version of the proof correct.

Why it matters: A machine-checked Lean proof of a long-standing result shows that AI-generated mathematics can be verified formally. Organisations need that kind of check before trusting AI output in technical work.

1 source
ReportedProducts ·

Apple launches health AI research app with 200,000 participants

Apple's health AI research app, which has 200,000 participants, uses OpenEvidence to answer medical questions from clinical sources.

Why it matters: Health organisations and their vendors may look at how consumer medical AI handles data retention and training use when evaluating similar tools.

1 source
ReportedModels ·

Alibaba's Qwen4-Max ships, scoring 43 on the Intelligence Index at lower cost

Alibaba's Qwen4-Max is now shipping, scoring 43 on the Intelligence Index at its highest reasoning setting, up from 34 at preview.

Why it matters: A lower cost per task at a higher benchmark score makes Qwen4-Max a price-competitive option for agent workloads. The legal benchmark shows that checklist accuracy can overstate task completion.

1 source
OfficialResearch ·

Anthropic says Claude Fable 5.2 solved two open problems in black hole physics

Anthropic says Claude Fable 5.2 solved two open theoretical physics problems on black holes and quantum gravity, one a long-unfinished 2003 calculation.

Why it matters: The claim is that long-running, largely autonomous agents produced checked research results in a hard field. Teams planning agent-driven R&D may find that a useful signal. The results are company-reported, and the company itself says they do not settle the underlying question.

1 source
OfficialSafety & Security ·

Prime Radiant releases three external sandbox servers for coding agent code

Prime Radiant has released three open sandbox servers that run agent-generated code outside the agent's own environment, with a shared localhost API.

Why it matters: Teams that want to test local models or run agent-written code can isolate execution from the host machine, reducing the damage a mistake or prompt injection could cause. The network allow-list and credential substitution address two common risks in agent workflows.

1 source
ReportedChips & Infrastructure ·

Optical computing firms aim for 10- to 80-fold energy gains over GPUs in AI inference

IEEE Spectrum reports that optical computing companies aim for 10 to 80 times better energy efficiency than today's GPUs, a goal not yet shown at scale.

Why it matters: Buyers planning inference capacity should treat optical efficiency claims as unproven. The article notes that energy gains depend on system size, encoding costs, and how much of the workload optics can actually handle.

1 source
ReportedModels ·

Tencent releases Hy4 Flash, a 224B multimodal model with 12B active parameters

Tencent's multimodal Hy4 Flash has 224B total parameters, activates 12B per token, and is priced at $0.070 and $0.281 per million input and output tokens.

Why it matters: A low-cost sparse model with a 1M-token context could reduce the cost of long-document workloads. Independent benchmark results are not given in the source.

1 source
OfficialProducts ·

OpenCode 1 Desktop now offers an upgrade to OpenCode 2.0

OpenCode 1 Desktop users can now upgrade to OpenCode 2.0, which downloads in the background and installs only when the user chooses.

Why it matters: Teams that depend on OpenCode plugins will need to update them as part of the move to OpenCode 2.0. Sessions, projects and settings are said to carry over.

1 source
OfficialProducts ·

OpenCode 1.19.6 caps image history and retries oversized request errors

OpenCode 1.19.6 replaces the oldest images in a session once its history passes 25 MiB and retries provider errors for oversized requests.

Why it matters: Long agent sessions with many screenshots or images are less likely to stop with a request-size error. Teams should expect the oldest images in such sessions to appear as placeholders.

1 source
ReportedPolicy & Regulation ·

OpenAI and the Government of Chile announce a partnership on responsible AI adoption

OpenAI and Chile's government will work together on responsible AI adoption, public-sector skills and AI applications from local startups.

Why it matters: This is a government-level arrangement with OpenAI and may shape AI procurement and skills programmes in Chile. Companies that sell to Chilean public bodies or to local startups should expect the terms to be clarified.

1 source
OfficialProducts ·

OpenAI adds two- and eight-hour background completion windows to Responses API beta

OpenAI's Responses API public beta lets developers choose a two-hour or eight-hour background completion window, with lower token rates for each.

Why it matters: Teams that can wait for results could cut token costs by accepting a longer completion window rather than paying for immediate responses. The discount applies to the windows OpenAI lists, so the saving depends on how long a job can tolerate waiting.

1 source
OfficialProducts ·

Claude Code 2.1.327 tightens permission checks on paths, symlinks and PowerShell

Anthropic's 2.1.327 changelog adds stricter path, symlink and shell checks, so more commands now ask first or are refused.

Why it matters: Unattended or automated Claude Code runs in auto or bypass modes may now pause or be refused where they previously ran. Pipelines that relied on earlier permission behaviour could stall.

1 source
OfficialProducts ·

Claude Code 2.1.327 changes artifact approval rules for managed organizations

The release refuses some new artifacts in sessions with nobody to approve them and stops the Artifact tool when Artifacts are turned off.

Why it matters: Organizations that require approval for each artifact may see refusals in background and cloud sessions. Logs now hold less detail on elicitation and hook answers, which affects audit trails.

1 source
OfficialProducts ·

Claude Code 2.1.327 adds stdin input for MCP servers and a billing group setting

Claude Code 2.1.327 lets users pass MCP server JSON through stdin and adds a setting to name a session's billing group, alongside many fixes.

Why it matters: Teams that script MCP server setup or manage billing through gateways get safer secret handling and clearer cost attribution. The .mcp.json fix removes a risk of silent configuration loss.

1 source
ReportedProducts ·

Google's Gemini API adds voice-driven editing to Gemini 3.5 Transcribe models

Gemini 3.5 Transcribe and its live variant now support voice-driven editing, rewriting and formatting of dictated speech or existing text.

Why it matters: Teams building dictation or transcription products may be able to handle spoken edits inside the transcription model, instead of in a separate post-processing step.

1 source