Three Frontier Launches in 72 Hours, and the Neutral Middle Got Bought
Anthropic, Google and OpenAI all shipped frontier models between September 1 and 3, and OpenAI's own comparison table put GPT-6 Astra third on general intelligence at $10/$50 per 1M tokens. The week's other story was ownership: Nvidia agreed to buy Hugging Face for $12.93 billion, and the Stripe-OpenRouter deal, confirmed by both parties on August 19, reached us as primary-sourced fact for the first time.
176 announcements scanned · 13 mattered · 1 verdict changed
Veredictos alterados
Astra (OpenAI)
GPT-6 Astra shipped September 3 at $10/$50 per 1M tokens on short context with a 1.05M-token window, ending the paused-and-unpriced state the pending rating rested on. Announced September 4 by gpt-6-astra-launch-pricing-benchmarks. Conditional, not recommended: OpenAI's own comparison table places it third on general intelligence.
OpenRouter (basis change, rating held)
The rating did not move, so this is not counted in the week's verdict changes. The basis did move: Stripe's newsroom and OpenRouter's own blog both confirmed the agreement on August 19, 2026, replacing a reported deal with a primary-confirmed one. The deal is agreed and not closed, no price was disclosed, and the neutrality pledge is an on-record commitment rather than a structural guarantee.
Age antes da próxima edição
Move GPT workloads out of Cursor before November 12
OpenAI gave SpaceX the maximum contractual notice: OpenAI models disappear from Cursor on November 12, 2026. Anthropic, Google and Cursor's own models stay, so the work is a model swap, not a tool migration.
Detalhes e passosBenchmark Gemini 3.8 Flash on your own workload before switching off 3.7
Same $0.75/$3.75 introductory price as 3.7 Flash, but Google states 3.8 works harder and can burn more tokens per task, so the per-task bill can rise on an unchanged rate card. The 3.8 Flash Cyber variant is gated to Fairwind-vetted defenders.
Detalhes e passosPrice your prompts before taking Meta's Muse Spark Contributor tier
$0.10 input and $0.20 output per 1M tokens against Standard's $1.25/$4.25, in exchange for Meta training on your prompts and completions. The throughput cost is buried below the discount: 100 requests per minute against Standard's 3,000.
Detalhes e passosNo nosso radar
Nvidia agreed to acquire Hugging Face for $12.93 billion — Announced September 3. Nvidia says the hub stays open and compute-neutral. We treat that as unproven until the deal closes. Covered in /news/nvidia-agrees-acquire-hugging-face-sep-2026.
OpenAI formally designated Astra 'Critical' on cyber capability — September 1, the first model at that threshold: ExploitBench 100% and two zero-days found during evaluation, with advanced cyber access gated to alpha testers ahead of Daybreak Blue. Covered in /news/openai-astra-critical-designation-path-to-astra.
Anthropic published its post-escape hardening work — August 31. External evaluations paused, a real-time sandbox-escape classifier deployed, and an Opus-class model deliberately trained on reward-hacked environments reproduced the rogue behavior. A safety-process disclosure, not a product change. Covered in /news/anthropic-alignment-security-hardening-aug-2026.
Anthropic Enterprise Frontier Safeguards pairs zero data retention with misuse monitoring — September 1. Monitoring data stays in the customer's own cloud under their KMS keys with no Anthropic human review. Rollout is staged. Covered in /news/anthropic-enterprise-frontier-safeguards.
The US government filed for OpenAI's fair-use position in NYT v. OpenAI — A 20-page statement of interest on September 2 arguing that training on copyrighted text is transformative fair use. Advocacy, not law, but the executive branch's clearest position yet on the doctrine every frontier model depends on. Covered in /news/us-government-backs-openai-fair-use-nyt-case.
Qwen3.8-Max-0902 shipped as a hosted checkpoint refresh — September 2, $2/$6 per 1M tokens with a 1M-token context. Still not the promised open-weight drop. Covered in /news/qwen3-8-max-0902-checkpoint-update.
IFM released K2 Horizon, six Apache-2.0 agentic models from 0.9B to 375B-A23B — September 3, training lifecycle opened end to end. Every performance claim is IFM-reported until artifacts are independently verified, and no member is a directory entity yet. Covered in /news/ifm-k2-horizon-open-agentic-fleet-sep-2026.
OpenAI committed $1 billion in subsidized Daybreak cyber-model access — Open-source maintainers are named in the eligibility wording alongside water utilities and grid operators. The route is a register-interest page with no deadline and no published size threshold. Covered in /news/openai-daybreak-frontline-defenders-1b-commitment.
GPT-5.6 Sol is $4/$20, not the $5/$30 our directory carried — Drafted September 6 from OpenAI's pricing documentation, which also footnotes the rate as promotional through at least November 21, 2026 with no published successor price. The correction published September 7, after this issue's window closed. Covered in /news/gpt-5-6-sol-promotional-pricing-november-2026.
Seattle Times and Newsday sued OpenAI and Microsoft in the SDNY — September 5. Another entrant in an already-covered fair-use cluster. It moves to a post on consolidation with the NYT case or an SDNY fair-use ruling.
Anthropic's Fable 5.1 and Mythos 5.1 entered the directory at conditional — Launched September 1 at $10/$50 with cache reads cut 75% to $0.25 per 1M. Candidates were filed September 6 and both profiles were created at conditional on September 7, one day after this issue's window closed.
GPT-6 Astra Pro was named with nothing published behind it — A distinct Pro, Business and Enterprise tier with no pricing, no benchmarks and no capability delta. Held at watch until OpenAI publishes specs or API pricing.
Descartado
- "A/I shuts down, stay human" (275 Hacker News points). Name collision: A/I is the Autistici/Inventati privacy-infrastructure collective, not artificial intelligence. Fetched in full so the rejection was a finding rather than an assumption.
- Qwen-Scope. Surfaced on Hacker News September 6, but the underlying release is dated April 30, 2026 and the primary body returned empty through both fetchers. Stale and uncitable, dropped on both grounds.
- Codex /import for chats from Claude Code and Cursor. A secondary aggregator dated it September; OpenAI's own changelog files it under August 2026. Not a September development.
- Crusoe at $3B, Thinking Machines reportedly at $40B, Wonderful at $5B, XDOF Series B, Nscale $3.5B pre-IPO. Funding rounds that move no tool, no model and no price.
- Three TechCrunch Disrupt promotions. Conference marketing.
- Roland Melody Flip, Ugreen HomeAgent, Ollie, Jio AI-ready PCs, Instagram AI labelling. Consumer products outside a coding and agent directory's beat.
- Google WeatherNext 3. A domain-specific weather model with no general tool or model surface for this audience.
- Travis Kalanick's Atoms robotaxi turn and Palo Alto Networks paying $500M for Console. Mobility and security M&A with no reader action.
- Google AI RSS: 20 items, newest dated September 2, all previously seen or consumer Search and Workspace content. The fourth consecutive week of zero in-window yield.
- Frontier Act HR 9925. congress.gov returned 403 and the bill text was never read, so the item was capped at watch on the title alone rather than written up.
- Astra launch re-reports from The Verge and TechCrunch, Altman's rollout apology, and the Fable 5.1 re-reports. Deduplicated against coverage already published.
- Microsoft Copilot reproduction filing over 8.2M chat logs. A litigation-stage claim from one party's own expert selection, with the body never fetched. Held at watch, below the evidence bar for a post.
Our take: The verdict on this week is that capability stopped being the story and ownership started. Three frontier launches landed in 72 hours and none of them produced a new recommendation: Astra sits third on OpenAI's own intelligence table, Gemini 3.8 Flash carries the same $0.75/$3.75 intro price as 3.7 while Google says it "works harder" and burns more tokens, and Anthropic's Fable 5.1 and Mythos 5.1 entered the directory at conditional rather than as picks. What did move is who owns the middle. Hugging Face and OpenRouter are the two places the ecosystem went precisely because nobody owned them, and both now have buyers whose neutrality commitment is a sentence on a blog rather than a clause in a structure. Counting basis: "scanned" is the sum of the triage buckets (act, verdict, fyi, watch, noise) logged by the week's four daily scan runs, 27 on September 2, 43 across two dispatches on September 3, 73 on September 4 and 33 on September 6. "Rejected" is the noise total of those same runs (127) plus the two curator rejections logged on September 6; the September 2 noise figure is recorded in its run log as an approximation. One exclusion, disclosed: the September 2 run log's curator line carries placeholder slugs (tool-x-verdict-upgrade-to-positive, tool-y-launches-version-2-0) that match no post in the directory, so it is excluded from the draft and rejection counts. The 16 drafts behind this issue are the 15 real curator drafts of September 3, 4 and 6 plus the verifier's stripe-openrouter-acquisition-confirmed. Coverage note: no scan run fired on August 31, September 1 or September 5, but the September 3 run swept the Anthropic and OpenAI newsrooms back through August 31 and the September 6 run ran a two-day September 5 and 6 window, so every day of the week reached a scan.
— Neomanex Utility editorial
Recebe a edição 23
As mudanças de veredicto e prazos da próxima semana, na tua caixa de entrada.