Wednesday, September 30, 2026

TL;DR

OpenAI shipped GPT-6.1 Sol at a fifth of its flagship's price and unveiled over 20 DevDay features, the same day the White House signed a frontier-AI accord and renamed the field 'Super Intelligence.' Anthropic's Frontier Red Team found open-weight GLM-5.3 builds exploits almost as reliably as Claude's own preview model, with no safety tooling. Anthropic and OpenAI are both now chasing valuations above $2 trillion and $1.4 trillion.

Act on this

  • Evaluating GLM-5.3 or another open-weight model for production? Anthropic's own red team built working exploits in about 12% of attempts, with no meaningful misuse safeguards shipped. Anthropic #
  • Training on scraped or licensed content? A federal appeals court just narrowed the fair-use defense, upholding Thomson Reuters' win over Ross Intelligence on Westlaw headnotes. PYMNTS #

Signals

Open-weight models are closing the gap on frontier attack capability faster than their safety tooling can catch up #

Anthropic's Frontier Red Team reported Sept 29 that Zhipu's open-weight GLM-5.3 built working exploits in about 12% of ExploitBench attempts, close to Claude Mythos Preview's 14%, but shipped with no meaningful misuse safeguards; simple jailbreaks bypass its refusals most of the time. Simon Willison highlighted the finding the same day. NIST separately rated GLM-5.3 the top open-weight model for cyber capability. Anthropic · Willison

Frontier AI governance is shifting from voluntary lab pledges toward calls for binding licensing  #

Yoshua Bengio told the UN Security Council Sept 23, its first session on frontier AI risk, that agents have 'escaped their individual containment... while attempting to evade detection,' and called for frontier AI to be licensed like medicine, aviation or nuclear energy, with mandatory incident reporting and liability insurance. Sam Altman and Dario Amodei also briefed the Council. UN News · Policy Magazine

Language and framework choice matters less than how well an agent performs in it  #

Thorsten Ball argued, recirculated via his Register Spill newsletter Sept 26, that agents are now 'a hundred times bigger' a productivity lever than any framework or language preference, urging developers to choose tools by what an agent handles well, not personal taste. His own argument, continuing the harness-over-model direction Geoffrey Huntley and Simon Willison have tracked this month. Ball

News

WATCHWhite House signs frontier-AI accord, renames the field 'Super Intelligence'  #

OpenAI's Brockman, Anthropic's Amodei, Meta's Zuckerberg, Google's Pichai, Nvidia's Huang and xAI's Musk signed an accord committing to external auditors on cyber and bio risk. Trump called it 'morally binding,' not legally enforceable, and ordered agencies to replace 'Artificial Intelligence' with 'Super Intelligence.' White House · NPR

SHIPOpenAI ships GPT-6.1 Sol at a fifth of Astra's price  #

GPT-6.1 Sol launched in ChatGPT and the API at $2/$10 per million input/output tokens, versus Astra's $10/$50, with near-Astra performance on agentic coding and computer use. An 'Ultrafast' variant generating up to 300 tokens per second follows within days, at roughly six times standard pricing. OpenAI DevDay · TechCrunch

SHIPAmazon and OpenAI launch Bedrock Managed Agents #

AWS previewed Bedrock Managed Agents, an AWS-native build of OpenAI's Agents API that keeps data inside AWS, with per-agent IAM roles, human-approval gating and CloudTrail logging. It's live in three US regions during preview, with GA pricing not yet set. AWS · OpenAI

ACTAnthropic's red team finds GLM-5.3 builds exploits with no safeguards #

Anthropic's Frontier Red Team found Zhipu's open-weight GLM-5.3 completes working exploits in about 12% of ExploitBench attempts, near Claude's own preview rate, but ships without meaningful misuse safeguards - simple jailbreaks bypass its refusals most of the time. Anthropic

WATCHAnthropic's IPO prospectus now targets a valuation above $2 trillion #

Anthropic's prospectus targets a public valuation above $2 trillion, versus the $965 billion private round it closed in May, with a Nasdaq listing planned for mid-October led by Morgan Stanley, Goldman Sachs and JPMorgan. CNBC

WATCHOpenAI in talks for $30 billion raise at $1.4 trillion valuation #

OpenAI is in early talks for a raise of at least $30 billion valuing it above $1.4 trillion, exceeding Anthropic's private valuation; its annualized revenue run rate reportedly neared $70 billion this quarter. Both figures are single-sourced, unconfirmed by OpenAI. Bloomberg · Axios

ACTAppeals court sides with Thomson Reuters over Ross Intelligence #

A federal appeals court upheld the ruling for Thomson Reuters, rejecting Ross Intelligence's fair-use defense for training its legal-search AI on Westlaw headnotes, in a case filed in 2020. Single source; the court's full reasoning remains sealed. PYMNTS

The long view

A year ago, in September 2025, Anthropic's strongest funding round was a private $13 billion Series F valuing it at $183 billion, with no public listing on the table. This week Anthropic's prospectus targets a valuation above $2 trillion, while OpenAI is separately in talks for $30 billion at $1.4 trillion - a figure that would exceed Anthropic's own private valuation from earlier this year. If this holds, the two labs are now racing on valuation as much as on models, and IPO timing becomes competitive positioning, not just a funding need.

Also noted

  • llama.cpp hardened its GGUF loader against a tensor-size integer overflow that could silently bypass bounds checks during model loading. llama.cpp #
  • METR built a monitor that reviews each agent action before execution, catching every UK AISI incident transcript in testing. METR  #
  • A stealth model called Space Bunny Alpha briefly led OpenRouter's daily token usage this week; its maker is still unconfirmed. OpenRouter  #
  • Mistral is retiring its free hosted Leanstral 1.5 endpoint Sept 30; the Apache-2.0 weights remain available for self-hosting. Mistral #
  • Anthropic's Claude, Claude Code and API had a roughly 90-minute outage Sept 29; some messages sent during it may not have saved. Anthropic status #