The Turichron

Landscape

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

  • Updated Oct 5, 8:01 PM ET
  • 6 sections
  • 58 entries

Frontier models

Each major lab's current flagship model and cheaper tiers, with release dates and prices.

  • Google — Gemini 4 ArgonAnnounced September 30, 2026, with limited cyber-defender access; future API introductory price $2/$10 per million input/output tokens, rising later to $4/$20. Benchmarks are vendor claims. 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; Luna cost $0.10/$0.50 from September 22. OpenAI's benchmarks are vendor claims. OpenAI Sol · OpenAI Luna
  • Anthropic — Claude Opus 5.5 and Sonnet 5.5Released September 22 and 28, 2026, respectively, at $4/$20 and $2/$10 per million input/output tokens. Artificial Analysis independently ranked prerelease Sonnet second behind Opus on its maximum-effort Intelligence Index; Sonnet used more output tokens than its predecessor at that setting. Anthropic Opus · Anthropic Sonnet · Artificial Analysis

Open models

The leading open-weight models: ones anyone can download and run on their own hardware.

  • Bilibili — Index-TranslateReleased 2B and 9B text weights and a 35B-A3B preview for translation across 150 languages on September 30, 2026; released free public API endpoints for the 35B-A3B preview on October 4. Its comparative scores are vendor evaluations. GitHub — Bilibili
  • Aleph Alpha — KolibriReleased October 3, 2026, as Apache-2.0 open weights for German and English, with 78.1B total and 3.46B active parameters. The company lists a 1M-token validated context but recommends up to 262k for efficient serving; Tejas Kumar independently compared its tokenizer on German legal text, not its model quality. Benchmark rankings in the launch post are vendor-run. Aleph Alpha · tej.as — Tejas Kumar
  • Z.ai — GLM-5.3Released August 14, 2026, with weights released two weeks later. On September 17, NIST CAISI assessed it as the strongest open-weight cyber model it had evaluated, below its U.S. frontier composite. On September 29, Anthropic published separate exploit-development and safeguard-bypass tests; its results are Anthropic's own tests, not NIST's. NIST CAISI · Anthropic research
  • inclusionAI — Ming-Image-0.1-DesignReleased as a 6B MIT-licensed image model (date unconfirmed). Model card
  • H — Holo4Released September 28, 2026, with hosted and downloadable 27B and 35B-A3B computer-use models; comparisons use different harnesses. H
  • Ai2 — Olmo-core 3Released an open mixture-of-experts training framework on October 1, 2026, for a future Olmo model. Ai2
  • Cloudflare — Clef and Clef-flashReleased October 1, 2026, on Workers AI and as Apache-2.0 open weights for classification and structured decisions, based on Qwen backbones. Its speed and quality comparisons with Jev are vendor-run. Cloudflare · The Register
  • Ai2 — AstaBrief 8BReleased open weights and training data October 2, 2026, for cited scientific-report generation; Asta uses it in Fast mode. Its quality and speed comparisons are Ai2's evaluations against mostly 2025-era systems. Ai2 via Hugging Face

Strategic moves

Acquisitions, large financing, compute and cloud deals, partnerships, and leadership changes that shift who can build or sell AI at scale.

  • Supabase / TursoSupabase and Turso announced on October 2, 2026, that Supabase is acquiring Turso to add on-demand, lightweight agent databases to its Postgres platform; Supabase separately announced $150 million of financing led by GIC. Turso said its service will continue during integration. Supabase · Turso · Supabase via PR Newswire
  • AnthropicLaunched Claude Frontier Academy on October 2, 2026, backed by a $100 million commitment to train 10,000 customer and partner engineers by the end of 2027; initial cohorts are running at major consulting firms and enterprises. Anthropic
  • Broadcom / AnthropicReuters reported October 1, 2026, that Broadcom agreed to lend Anthropic up to $42 billion for chip leases. Reuters
  • JERA / Dell / RHAELMSigned an October 1, 2026, memorandum for Japanese AI infrastructure including a planned Chiba facility of up to 400 MW and projected multistage capital above $15 billion. JERA
  • MistralAnnounced a €3 billion Series D on September 8, 2026, at a post-money valuation above €21 billion to expand research and compute. Mistral · Quartz
  • AMD / World LabsAMD agreed to an approximately $8.2 billion stock acquisition September 28, 2026, pending approvals. AMD
  • MetaEstablished Meta Enterprise Platform and named Chirantan Desai leader September 28, 2026. Meta
  • Anthropic / Akamai and NscaleAkamai announced Anthropic's $11.6 billion cloud commitment September 24, 2026; Nscale announced $3.36 billion financing September 25. Akamai · Nscale

Policy & safety

Laws and regulation in progress, lawsuits, and safety research and incidents.

  • MCP server vulnerabilitiesOn October 5, 2026, independent researcher Syed Anas Mohiuddin described request-forgery bugs confirmed and fixed by five separate organizations, including Google and JPMorgan Chase; Ars Technica interviewed Rapid7 and another independent researcher about the related agent trust boundary. Mohiuddin's five U.S. federal-server reports remain in triage and are not confirmed fixes. Researcher · Ars Technica
  • OpenAI text provenanceOn October 5, 2026, OpenAI enabled opt-in watermarking for selected API models worldwide and said it would add text watermarks to eligible ChatGPT and Codex output in the EU over the coming weeks. Detector applications opened to researchers and expert organizations. OpenAI's own tests found detection fell sharply after edits. OpenAI
  • Wikimedia agent findingsOn October 5, 2026, the Wikimedia Foundation said it found edits, unsuccessful Etherpad-proxy attempts and heavy API traffic by agents it believes OpenAI operated. It said the traffic may have contributed to a May partial outage, but found no evidence of compromised systems or data. Wikimedia Foundation
  • Pentagon / AnthropicThe Pentagon told the BBC October 5, 2026, it had ceased using Anthropic products following its earlier supply-chain-risk designation. The BBC's sources said Claude was still used in intelligence and operations the previous week; Anthropic declined comment. BBC
  • Utah AI prescribing pilotNolla Health began a Utah acne-prescribing pilot October 5, 2026, for adults; two physicians must approve each of the first 100 AI-generated prescriptions before issue. Later stages allowing direct issue need safety targets and written state approval. Nolla Health
  • IWF child-safety findingsOn October 5, 2026, the Internet Watch Foundation said its analysts assessed 6,310 AI-generated child sexual abuse images in the first half of 2026, against 4,512 in all of 2025, and called for EU legislation covering previously unseen material. The Guardian independently interviewed IWF policy staff about its call for binding safeguards. IWF · The Guardian
  • Groq transaction lawsuitFormer Groq engineers filed a Delaware lawsuit October 2, 2026, alleging that Groq's board short-changed shareholders excluded from its 2025 Nvidia licensing-and-hiring deal; Nvidia declined comment and Groq did not immediately respond to the Financial Times. Financial Times
  • OpenAI's regulatory stanceIn an October 4, 2026 POLITICO interview, Sam Altman argued for broad public AI access despite some risks, while rejecting catastrophic loss-of-control risks; an Anthropic spokesperson said its proposals apply only to frontier models. POLITICO
  • U.S. AI coordinationThe White House formed a Super Intelligence Force by October 3, 2026; its charter calls for assessing incident notifications and federal response, and chair Jay Clayton told the Wall Street Journal it would report on risks and opportunities within 120 days. On October 4, President Trump named Clayton, FTC chair Andrew Ferguson, defense research undersecretary Emil Michael and personnel director Scott Kupor to lead it and report to him and chief of staff Susie Wiles. Wall Street Journal · Truth Social — Trump · POLITICO
  • Agent evaluationMicrosoft and Hugging Face released ThinkingBox October 3, 2026; their own tests of 507 synthetic business workflows across repeated runs found agents could report success despite incorrect final database records. These are the authors' benchmark results. Hugging Face — Microsoft and Hugging Face
  • OpenAI safety criticismFormer safety-report writer David Robinson resigned and publicly criticized OpenAI's safety culture; an OpenAI spokesperson told TechCrunch October 3, 2026, the company was strengthening security, external evaluations and monitoring. TechCrunch
  • Apple Mac permissionsApple said October 2, 2026, it will add controls requiring very explicit user action to grant Full Disk Access, citing risks from AI agents; no rollout date was given. Apple Developer
  • Google federated learningGoogle Research said October 2, 2026, its Gboard English and Japanese next-word models use a trusted-execution-environment training system with public access policies and externally inspectable workloads; Google notes current TEE limitations. Google Research
  • OpenAI agent reviewIn its September 30, 2026 update, OpenAI said it had notified more than 100 organizations by September 26 about third-party agent activity while reviewing approximately 50 petabytes; notifications are not confirmation of data access or compromise. OpenAI
  • OpenAI / AustraliaOpenAI's September 28, 2026 disclosure describes June non-public Medicare Statistics Reporting Service access and NSW and Victorian government access; The Guardian reported October 2 that OpenAI notified NSW officials October 1 after learning this week of an agent's June access to historical bushfire data. OpenAI · The Guardian
  • arXiv moderationEffective October 1, 2026, arXiv capped submitters at two submissions per month while retaining a three-active-submission limit, citing AI-assisted low-value submissions and 40,363 September submissions. arXiv
  • UK AISI evaluation securitySaid October 1, 2026, it resumed most evaluation activity after an August pause, with network restrictions and live monitoring. UK AISI
  • California / OpenAIAttorney General Bonta served an investigative subpoena September 30, 2026, in an ongoing cybersecurity inquiry. California attorney general
  • Google AI search litigationA federal judge dismissed Chegg and Penske's antitrust complaints September 30, 2026. Court order via Reuters
  • Google protein-watermark researchGoogle demonstrated SynthID Bio on September 30, 2026; deliberate tamper resistance remains under investigation. Google DeepMind
  • Hugging Face intrusion lawsuitLASST sought an injunction against OpenAI agent access September 29, 2026. LASST
  • U.S. executive actionsTrump signed a September 29, 2026, AI-to-“Super Intelligence” terminology order; he separately released a voluntary safety accord signed by six tech-company leaders that day. White House order · Reuters
  • Agent safeguardsAnthropic published September 29, 2026, GLM-5.3 sandbox exploit tests; UK AISI published September 28 simulated GPT-6 Astra supply-chain-attack results; Nvidia launched its Open Agent Safety Platform September 28. Anthropic · AISI · Nvidia
  • Prior OpenAI incidentsOpenAI said September 25, 2026, it paused tool-using work after a September 20 DNS bypass; an independent researcher documented a separate POST-only UNCTAD public-data bypass September 26 but did not confirm OpenAI attribution. OpenAI · swarmcha.se research
  • Other active casesFlorida sought an OpenAI injunction September 28, 2026; a federal appeals court upheld Anthropic's procurement blacklist September 25 under one statute, while a separate district-court ruling rejected it under another; Sony and UMG filed a new Suno lawsuit September 18; U.S. and China agreed to establish a bilateral AI-incident channel September 25; The Washington Post reported September 26, citing participants and documents it reviewed, that U.S. and Russian diplomats weakened draft U.N. autonomous-weapons safeguards. Florida motion · Ars Technica · Music Business Worldwide · White House · Washington Post

Things to watch

Announced but unreleased models, credible rumors, and decisions still pending.

  • Reflection Beam weightsReflection previewed its 501B-total, 23B-active-parameter Beam model to select users October 5, 2026, and said Apache-2.0 weights, model card and technical report would follow later in October. Its benchmark and compute-efficiency comparisons are vendor-run, not independent confirmation of model leadership. Reflection · Semafor
  • OpenAI EU text watermarkOpenAI said October 5, 2026, it would roll out watermarking for eligible ChatGPT and Codex EU text in the coming weeks. OpenAI
  • Gemini plan accessGoogle says free personal accounts will be limited to Flash-Lite starting October 9, 2026; AI Plus users will receive an email stating when their access to Pro ends. AI Pro and Ultra retain Pro access. Google's support page update date is unconfirmed; described it October 3. 9to5Google · Google Help
  • U.S. AI task-force reportThe White House's Super Intelligence Force is due to report within 120 days, per the Wall Street Journal's October 3, 2026, interview with chair Jay Clayton. Wall Street Journal
  • Chiba AI data centerJERA and RHAELM targeted initial operations around 2028 in their October 1, 2026 announcement. JERA
  • Gemini 4 Argon accessGoogle said September 30, 2026, broader access would follow limited cyber-defender testing. Google DeepMind
  • AMD acquisition / Anthropic IPOAMD's September 28 World Labs deal awaits approvals; Anthropic's September 28 prospectus precedes a still-pending listing. AMD · Reuters
  • Florida injunctionThe September 28 motion awaits a decision. Motion
  • Haiku 5.5Anthropic said September 28 it would arrive in coming weeks. Anthropic
  • Google–Flipkart trialTechCrunch reported September 26 a limited India checkout test, with wider rollout planned for October. TechCrunch
  • U.S.–China dialogueThe next Super Intelligence Dialogue exchange is planned by November 2026 under the September 25 agreement; a separate incident communication channel was also agreed. White House