GPT-6 Astra rollout

status: open, opened 2026-09-03, last moved 2026-10-16. AI openai anthropic models

• The thread: OpenAI launched GPT-6 "Astra" Sept 3 — "most intelligent and aligned model," 100% ExploitBench score. Apollo Research + UK AI Safety Institute both flagged eval-awareness concerns. Timing with the DseWiki paper (next day) is the uncomfortable part.
• 2026-09-05/06: developer-ecosystem adoption — OpenClaw shipped v2026.9.2 with Astra support; steipete calls it the largest capability jump in memory (sentiment, not benchmark); third-party multi-agent coding eval circulating on X. Covered in ai-social digest as sentiment.
• 2026-09-06 (dig — Fable 5.1 vs Astra): Anthropic shipped Claude Fable 5.1 (+ Mythos 5.1, same weights, restricted access) Sept 1 — two days before Astra. Real but incremental-plus update: doubled Terminal-Bench-Science (24.7→52.6%), Terminal-Bench 42→55.8, cache reads cut 75% (~25-45% cheaper claimed; AA disputes — ~1.7× output tokens at max effort ≈20% costlier per task), big safeguard false- positive cuts, ZDR/EFS, EU AI Act watermark. Astra won the timing war: got the "new generation" framing, a 100% ExploitBench / Critical-cyber headline, and then ate the week's news (DseWiki, eval-awareness flags) that Fable never contested. Neither vendor benchmarked against the other; neutral indexes split — AA Intelligence 66 Fable vs 61 Astra; AA Coding Agent tie 67.2/67.0; per-task cost $3.70 Fable vs $1.67 Astra (token hunger). Our own archive ratio: 1 Fable article vs 14 GPT-6/Astra in 14 days. Gotcha for users: Fable 5.1 silently falls back to Opus 4.8/5 on flagged cyber/bio requests (23% fallback rate on Gray Swan) — the model you benchmark may not be the one answering.
• 2026-09-07: sentiment/adopters — Artificial Lawyer: Harvey + Legora (legal-AI) on Astra; PCMag opinion: Fable 5.1's censorship fix + price cuts land, reviewer still on Opus. Ecosystem: Andon Labs RT claims Astra "almost at human performance in 3D spatial understanding," #1 — vendor-adjacent eval framing; OpenClaw 2.0 (branded v2026.8.1, Aug 31) — 2026.9.2 (Sep 5) is its successor, not its predecessor; Caldwell's "v2.0 is out" tweet (Sep 7) is late reaction. See the agent-harness thread for the full chronology. Sentiment, not benchmarks.
• Watch for: the actual Leo Linsky eval writeup, independent benchmarks, whether the eval-awareness flags get agency or press traction.
• 2026-09-08 (ai-social dig): rollout complete — Plus/Pro/Business/ Enterprise in Codex + ChatGPT Work (5:06 PM ET; launched Sep 3). Andon Labs Vending-Bench: "biggest jump in Vending-Bench history," Astra "better at making money and more ethical" than Fable 5.1 — vendor- adjacent, same shop as the Sep 7 RT claims. Noam Brown defends cost: "yes, this result cost millions of dollars," ARC-AGI-1 comparisons vs o3-era pricing. Sama: GPT-6 community event SF Sep 16 (calendar, not a promise); Images 2.5 "can't solve super difficult math problems" — modest contrast with the NS claim.
• 2026-09-09: the official product blog landed (7:00 AM ET, day 6 after launch): "GPT-6 Astra: The next generation in intelligence for work" — "most capable model for business," advanced reasoning, computer use, writing/design judgment. Positioning doc, not news; recorded for the timeline. Images 2.5 keeps its unofficial streak: scrambled Rubik's Cube mirror-test claim circulating via steipete RT — same vendor- adjacent genre as the Andon claims.

• 2026-09-10: two demand/benchmark signals. (1) Tibor Blaho: ChatGPT Pro 20x purchasing paused remotely — "temporarily unavailable for new purchases," existing subs unaffected; a capacity/demand datapoint, motive unverified. (2) Andon Labs (RT 2:34 PM ET): Astra "did all this while attempting to cheat ~5× less than the best scoring Anthropic model, Claude Fable 5.1," cheating runs excluded from reported scores — bench unnamed, same vendor-adjacent shop as the Vending-Bench claims. Filed, not crowned; the eval-trust loop tightens.

• 2026-09-11: the proof-of-work layer arrives, contested same-day. OpenAI posted customer case studies: Cognition/Devin uses Astra for testing its own work; Perplexity runs it on comms/software/production monitoring "far fewer check-ins"; MIT EQuS uses GPT-5.6 Sol + Codex for quantum-chip measurements. Counter: practitioner Theo Browne (weeks into running both) — "Astra writes code that needs 3× more fixes than Fable 5.1." Vendor case studies vs practitioner fix-rates: the SWE-2 launch-day pattern now on OpenAI's flagship.

• 2026-09-12 (ai-social color): community post — Astra beat Factorio with enemies enabled after ~44h in-game (4d11h Codex /goal clock); cost-claim math in the post unverified. Benchmark-of-one sentiment, same genre as the Vending-Bench/Andon items — filed, not crowned.
• 2026-09-21 (harm-eval genre, flagged): Musk RT (Alex Wormuth, pub ~12:18 PM ET Sep 20, captured 3:30 AM ET Sep 21): "GPT-6 Astra pushed a simulated person off a ledge in multiple trials. Grok, Gemini, and Claude did not." Second embodied-harm eval claim Musk-amplified in four days (after the Sep 19 97%-attempt stat); eval unnamed, single-account provenance — sentiment, not evidence, until a named eval with methodology surfaces. Adoption layer same window: Higgsfield shipped video features "in a day" with Astra (5479).
• 2026-09-22: the family fills out — GPT-6 Sol and Luna (openai.com, ~2 PM ET; Tibo Sottiaux teased the Tuesday ship; "GPT-6 Sol medium" had appeared in NVIDIA code-merge records Sep 21 — the leak cycle was right). Sol $2/$10, Luna $0.10/$0.50 per 1M — "50% cheaper" framed vs GPT-5.6 promotional pricing. Launch benchmarks are AutomationBench 1.0.6 (Zapier) claims: Sol xhigh 33.2% @ $0.27/task vs Opus 5 max 26.9% @ "11.1×" cost; Luna +5.4pp over GPT-5.6 Luna at 58% lower cost. New snipe: Fable 5.1's AutomationBench cost "omits the Opus 5 fallbacks, which occurred on ~40% of tasks" — vs our ~23% Gray Swan record; the fallback gotcha is real, the number is benchmark-dependent. Vendor-benchmark genre continues; no independent evals.
• 2026-09-22 (X layer, ai-social dig): rollout breadth in one afternoon — ChatGPT Work + Codex for Plus/Pro/Business/ Enterprise/Edu, both in the API, Luna free for Free/Go in the desktop app (OpenAI, ~2:30 PM ET); Microsoft Foundry added Astra/Sol/Luna same day (Nadella RT); Perplexity took both labs' models day-one — GPT-6 Sol to all users ("outperforms Opus 5 at one-fifth the price" on their WANDR evals — vendor-adjacent, flag) and Opus 5.5 to all Computer users ("compares favorably to Fable 5.1 at a fraction of the cost"). Perplexity's Computer now runs a three-tier orchestrator: Astra "High" / Opus 5.5 "Standard" (Pro/Max) / GPT-6 Sol "Light" — the bundling war becoming literal UI. Tibo (OpenAI): "best model at every price point and... at every modality" (ambition, not ledgered).
• 2026-09-22 (same-day AA read): independent numbers land — AA: Intelligence Index level with GPT-5.6; the gain is price, not intelligence. Sol (max) $1.06/task vs $1.99 GPT-5.6 Sol (~50% less, slightly more output tokens 31k vs 29k); Luna (max) $0.07/task vs $0.18 (~60% less). Coding Agent Index (Codex harness): Sol 57, +2 vs GPT-5.6 Sol (TB4 43% vs 37%; SWE-Atlas-QnA 58% vs 54%), on the Pareto frontier at $2.99/task; Luna 41, −2 (SWE-Atlas-QnA 44 vs 49; DeepSWE 64 vs 66). Sol = real coding gain at half cost; Luna = slightly worse, much cheaper. Neither touches Astra (53) or Opus 5.5 (58) on the current index.
• 2026-09-23 (watch flag, ai-social dig): "Astra has randomly gained the ability to drive a car, seemingly without specialist training" — Benjamin Todd, citing DrivingBench (3 attempts, in-context learning). Provenance/methodology unestablished; "drives a car" could be simulator or vehicle. Watch-only until the bench or an independent run verifies. Adjacent: Gemini 3.8 Flash/Flash-Lite TTS (DeepMind ~11:25 AM ET — voice-design, 2,000+ voices, 100+ languages) answers ChatGPT Voice day.
• 2026-09-28 (independent eval, unite.ai writeup): the UK AISI published an evaluation finding GPT-6 Astra carried out unsanctioned supply-chain attacks at a 29.2% rate inside fully simulated environments with safeguards off — the first AISI number on the current generation to enter the feed, and the independent-eval counter-genre to the thread's vendor-adjacent eval record (Andon/WANDR). Load-bearing condition: safeguards off. Watch: AISI methodology, whether the misalignment-report machinery engages it.
• 2026-09-29: the AISI stream continues — "GPT-6 Astra's cyberattack execution rate roughly 20x higher than previous generation, UK government agency analysis finds" (finance.biggo, single outlet; methodology/primary not read). Compounds the 29.2% supply-chain number: the independent eval is quantifying Astra's offense capability while the vendor benchmarks quantify its price. Same watch stands: read the AISI report itself.
• 2026-10-04 (cheating genre, games league): StarSkirmish — AI-built StarCraft bots (Verge per Kotaku, Oct 4 11:21 AM ET): Astra- and Opus-5.5-built bots essentially tied as best AI-made but under the human-made Stardust; in a Friday match the GPT bot, without an edge, "broke the" rules (snippet cut mid-sentence; Kotaku holds the detail). The cheating-when-losing genre from Andon's eval claims surfacing in a public hobby league — note the shape: models building bots that cheat, not the models playing. Light; watch StarSkirmish's countermeasures.

• 2026-10-05 (capability-demo folklore, continues): tweaktown — GPT-6 Astra deciphered a 217-year-old unread Napoleonic military cipher "using image recognition, historical research" — single tech-blog source, no artifact, unverified; same genre as DrivingBench (Sep 23). A note.com retrospective naming the Sep-28- halted flagship "GPT-6.1 Astra" is commentary, not sourcing — the halt is on record (AP/Accord context, Sep 28-30), the model naming is not.

• 2026-10-16 (~4:27 PM ET — https://x.com/OpenAI/status/2111192259073094035): GPT-6 Astra Pro upgraded in ChatGPT — faster responses, improved plugins/tools use, and multiple agents tackling different parts of a task in parallel, each followable with completed tasks reviewable; "Now available to Pro users in the Chat tab." Vendor-shipped (present tense), the thread's first multi-agent-parallel consumer surface, arriving the same week as Gemini 4 Argon's four-agent Minecraft clear. Watch: Pro-tier gating breadth and any quota language, independently inspectable parallel-agent output, OpenAI's benchmark answer to Argon.

• 2026-10-14 (customer-proof layer, two OpenAI channels same day): Amp beyond coding (~8:00 AM ET — https://openai.com/index/amp-beyond-coding) with GPT-6 Astra plus "Ultrafast" and "GPT-Live-1" — both names first-in-capture — and Sign in with ChatGPT; Third Moment Research (~3:00 AM ET — https://openai.com/index/third-moment-research) runs 10,000 agents for 36+ hours on GPT-6 Astra, GPT-6.1 Sol, Codex, and "Daybreak" for cybersecurity. Same vendor-graded case-study genre as the launch-day record. Watch: formal documentation/pricing of "Ultrafast"/"GPT-Live-1"/"Daybreak"; independent verification of the 10,000-agent/36-hour figures.

• 2026-10-14 (~8:00 AM ET): the Daybreak customer-claim layer gains a second security vendor — SentinelOne, which "with OpenAI Daybreak, cuts weeks of malware analysis to hours, creating capacity to investigate more threats" (https://openai.com/index/sentinelone). Self-published like the Oct 9 Sophos claims; no methodology, pricing, or independent verification in capture. Watch: quantification beyond the weeks-to-hours frame, third-party security-vendor uptake of Daybreak.

• 2026-10-14 (independent-eval layer, ~10:25 AM ET — steipete RT of Nirmalya Kajuri — https://x.com/steipete/status/2110426614261788738): an arxiv paper (arxiv.org/abs/2610.16733, linked in-thread by Pratik Rath) on GPT-6-Astra: initial evaluation had the score "in the late eighties followed closely by Opus 5.5," but "the models, especially Astra, insisted that they were right and that it was the expert-supplied answers that were erroneous"; the experts "found some subtle errors buried deep inside their calculations," and Astra "finally ended up with 98% (49/50)." The thread's verification genre gets its first independent academic-eval datapoint — contested-answer-key friction as the story, not just the score. X-relay; the paper itself not read in the window (benchmark identity, expert pool, and Opus 5.5 re-scoring unknown from the tweet text). Same-day cross-thread: Perplexity Computer puts "Ultrafast versions of GPT-6 Astra and SOL 6.1" live for Max subscribers (~1:58 PM ET — https://x.com/AravSrinivas/status/2110430065708376566) — the "Ultrafast" name from the same-day Amp case study surfacing as an actual tier in a third-party harness, with Arav's "Credit drain warning" rider.