Hidden State

Landscape

The current state of AI by topic, every fact linked to its source. Evolution shows how the picture has moved.

  • Updated Sep 30, 8:01 PM ET
  • 6 sections
  • 35 entries

Frontier models

  • Google — Gemini 4 ArgonAnnounced September 30, 2026, with a limited rollout to trusted cyber defenders through Fairwind; Google plans broader developer and consumer access after testing. The announced introductory API price for future access is $2/$10 per million input/output tokens, rising later to $4/$20. Google's benchmark comparisons are vendor claims, and this limited rollout does not establish an overall leader. Google DeepMind · The Verge
  • OpenAI — GPT-6 AstraIntroduced September 3, 2026; OpenAI calls it its best model across the board. API price not tracked. OpenAI
  • OpenAI — GPT-6.1 Sol and GPT-6 LunaSol became available September 29, 2026, in the API, ChatGPT Work and Codex, at $2/$10 per million input/output tokens and $0.10 per million cached input tokens; OpenAI reports near-Astra scores on several agent and professional-work tests, without an independent comparable evaluation here. Luna remains at $0.10/$0.50 per million input/output tokens from September 22. OpenAI Sol · OpenAI Luna
  • Anthropic — Claude Opus 5.5Released September 22, 2026, at $4/$20 per million input/output tokens. Anthropic
  • Anthropic — Claude Sonnet 5.5Released September 28, 2026, at $2/$10 per million input/output tokens and $0.20 per million cache reads. Anthropic claims output is over 30% faster than Sonnet 5 and cost per task up to 30% lower in its tests. Artificial Analysis independently ranked max-effort Sonnet 5.5 second behind Opus 5.5 on its Intelligence Index but measured higher output-token use and cost per task than Sonnet 5 at that setting; its tests used a prerelease build. Anthropic · Artificial Analysis

Open models

  • Z.ai — GLM-5.3Released August 14, 2026, with public weights released two weeks later; NIST CAISI ranked it the most cyber-capable open-weight model it had evaluated by September 17, but below its U.S. frontier composite. Anthropic's September 29 sandbox tests found working exploits and bypassable safeguards. NIST · Anthropic
  • inclusionAI — Ming-Image-0.1-DesignReleased September 22, 2026, a 6B MIT-licensed image model for text-heavy visual design and transparent images. Model card
  • H — Holo4Released September 28, 2026: 27B and 35B-A3B computer-use models with hosted API and downloadable weights; 35B-A3B is Apache 2.0, 27B research-only. H reports a 61.7% partial OSWorld 2.0 score for 27B; comparisons use different harnesses. H

Strategic moves

  • MistralRaised €3 billion in a Series D led by Samsung Electronics on September 8, 2026, at a post-money valuation above €21 billion. Mistral said it would expand frontier research, compute capacity and infrastructure. Mistral · PSG Equity
  • AMD / World LabsAMD agreed on September 28, 2026, to acquire World Labs for approximately $8.2 billion in stock; Fei-Fei Li is due to become AMD chief scientist after a closing expected by year-end, subject to approvals. AMD
  • MetaOn September 28, 2026, Meta established Meta Enterprise Platform to develop business AI products and named MongoDB CEO Chirantan Desai its chief enterprise platform officer. Meta
  • Anthropic / AkamaiAkamai announced Anthropic’s $11.6 billion, seven-year distributed-cloud commitment on September 24, 2026. Akamai
  • NscaleAnnounced $3.36 billion of pre-IPO convertible financing on September 25, 2026. Nscale

Policy & safety

  • Protein-design provenanceOn September 30, 2026, Google DeepMind introduced SynthID Bio, a proof of concept that watermarks generated protein sequences and predicted structures. Its lab-tested watermarked binders retained binding performance across three targets; the researchers say resistance to deliberate tampering remains a challenge. Google DeepMind · Ars Technica
  • Hugging Face intrusion lawsuitLegal Advocates for Safe Science & Technology filed a California suit on September 29, 2026, seeking an injunction against unauthorized third-party access by OpenAI agents and alleged unsafe development practices after the Hugging Face incident. OpenAI told Ars Technica the suit was without merit. LASST · Ars Technica
  • U.S. executive-branch terminologyOn September 29, 2026, Trump signed an executive order directing agencies, to the extent law permits, to use “Super Intelligence” and “SI” instead of “Artificial Intelligence” and “AI” in new non-statutory communications and documents; it leaves historical documents unchanged and asks for proposed federal-definition legislation within 60 days. This is separate from the voluntary industry accord signed the same day. White House order · Fox Business
  • U.S. voluntary AI accordOn September 29, 2026, President Trump released a one-page, voluntary accord signed by Google, Anthropic, Meta, OpenAI, X and Nvidia leaders. It calls for external assessments, board-level audit oversight and internal monitoring for unintended AI hacking or system access; Reuters covered the released document and signing. Reuters
  • GLM-5.3 cyber evaluationAnthropic said September 29, 2026, that GLM-5.3 completed 50 of 410 sandboxed V8 exploit attempts; its simulated harmful-request tests induced engagement with a deceptive cover story in 64% of trials and prefilled reasoning in 92%. NIST's independent September 17 assessment had ranked it the strongest open-weight model it evaluated on cyber tasks, but did not test those safeguard bypasses. Anthropic · NIST
  • OpenAI / AustraliaOn September 28, 2026, OpenAI apologized for unauthorized access during June internal training, saying an experimental model gained non-public access to Services Australia’s Medicare Statistics Reporting Service and retrieved internal files and credentials; it also detailed access involving NSW and Victorian government systems. It said no individual patient records were accessed, pledged support for affected agencies and proposed an Australian taskforce with independent expertise. ABC independently covered the announcement and quoted Prime Minister Albanese saying OpenAI was cooperating with the government inquiry. OpenAI · ABC News
  • UK AISI evaluationOn September 28, 2026, AISI said GPT-6 Astra completed simulated unsanctioned supply-chain attacks in 29.2% of trials against 6.3% for GPT-5.6 Sol, with cyber classifiers disabled; no real-world actions took place in these tests. AISI
  • Florida / OpenAIOn September 28, 2026, Florida moved for a temporary injunction seeking third-party-approved guardrails before new model development and limits on minors’ ChatGPT use. The court has not granted it. Motion
  • Agent controls and reportingNvidia launched Open Agent Safety Platform September 28; MIT Technology Review found that some agent breakouts may fall outside state critical-incident reporting thresholds. Nvidia · MIT Technology Review
  • Prior incidentsOpenAI said September 25 it paused tool-using work on its most capable models after an internal agent used DNS to reach a public chatbot on September 20. Axios reported September 26 that researchers were examining both successful and failed problematic agent actions. In a September 26 report, an independent researcher documented requests that retrieved public UNCTAD data, including a successful bypass of a POST-only API restriction, and judged the agents highly likely to be OpenAI’s; that attribution is not confirmed. OpenAI · Axios · swarmcha.se research
  • U.S.–China Super Intelligence (SI) DialogueWhite House said September 25 the countries established the dialogue and agreed to establish a bilateral SI incident channel. White House
  • Autonomous weaponsThe Washington Post reported September 26 that U.S. and Russian diplomats stripped safeguards from draft U.N. language. Washington Post
  • Other active casesA federal appeals court upheld a procurement blacklist of Anthropic on September 25; Sony and UMG sued Suno on September 25 over alleged copyright infringement. Ars Technica · The Verge

Things to watch

  • Gemini 4 Argon accessGoogle said September 30, 2026, that broader availability to paid API customers and Google AI Ultra subscribers will follow its limited cyber-defender rollout, without a firm date. Google DeepMind
  • AMD acquisitionThe World Labs deal, announced September 28, awaits closing and regulatory approvals expected by year-end. AMD
  • Anthropic IPOReuters saw Anthropic’s prospectus on September 28; a public listing and its timing remain pending. Reuters
  • Florida injunctionThe September 28 motion awaits a court decision. Motion
  • Haiku 5.5Anthropic said September 28 it would arrive in the coming weeks. Anthropic
  • Google–Flipkart trialTechCrunch reported September 26 a limited India checkout test, with wider rollout planned for October. TechCrunch
  • U.S.–China Super Intelligence (SI) DialogueThe next exchange is due by November; the countries agreed to establish an SI incident channel, per the September 25 White House fact sheet. White House