GPT-6 Astra

GPT-6 Astra launched September 3, 2026, per OpenAI's own announcement blog, safety overview, and System Card — all three independently corroborated by Bloomberg, CNBC, 9to5Mac, and Axios coverage the same day. OpenAI frames it as "the world's most intelligent and aligned model" and "the most capable model we have ever broadly deployed," positioned above the existing GPT-5.6 line (GPT-5.6 Sol, Terra, and Luna) rather than replacing it: none of OpenAI's own model or deprecation pages describes Astra as a successor to any GPT-5.6 model, and all three remain sold at unchanged prices with no retirement date — unlike this site's established supersession cases (Claude Fable 5 to Fable 5.1, Gemini 3.7 Flash to 3.8 Flash), where the vendor's own comparison table used explicit "successor to" language. Astra ships as a single model with five reasoning-effort settings (low/medium/high/xhigh/max) rather than as a three-tier family the way GPT-5.6 did. OpenAI's own launch page reinforces the same positioning without using supersession language: its benchmark tables compare Astra exclusively against GPT-5.6 Sol, and its own "choosing a model" guidance now recommends Astra first and omits Sol from that sentence entirely (while still selling it) — informal signals Astra has taken the flagship slot, not a formal retirement.

API pricing is $10 per million input tokens and $50 per million output tokens for prompts up to 272,000 tokens, per OpenAI's own pricing page — the same 272K threshold GPT-5.6 Sol and Luna already carry on this site. Above that line, the entire request bills at double the input rate and 1.5 times the output rate ($20/$75), a rule OpenAI states applies to the full request, not just the tokens past the line. Context window is 1,050,000 tokens, matching Sol and Luna's own figure exactly; max output is 128,000 tokens. Knowledge cutoff is April 30, 2026, both stated in plain text on OpenAI's own model page. No parameter count is disclosed, consistent with every other OpenAI model tracked here.

One day after launch, five of the eleven benchmarks this site tracks carry an independently-run Astra score: Humanity's Last Exam (54.68% at max effort, Artificial Analysis), GPQA Diamond (96.06% at max effort, also Artificial Analysis — this benchmark is saturated on this site and every comparison on it is tainted regardless of gap size), Terminal-Bench 2.1 (87.4% at high effort, tbench.ai's own canonical board, rank 1 of 18), DeepSWE (74% at xhigh effort, tied for the board's top spot), and ARC-AGI-2 (95.0% at Max tier, arcprize.org). Six others — LiveCodeBench, SWE-bench Verified, Toolathlon-Verified, Agents' Last Exam, LiveBench, and HMMT Feb 2026 — had not added Astra as of 2026-09-04, checked live on each; all six boards' most recent update predates the model's release.

Only one of those five produces a clean, same-tier, same-harness comparison against a tracked GPT-5.6 model: Humanity's Last Exam, where Astra's 54.68% clears GPT-5.6 Sol's own 49.49% (both at max effort, both Artificial Analysis) by 5.19 points — a real gap against this benchmark's 2-point noise floor. ARC-AGI-2 is also same-tier and same-harness (both Max, both arcprize.org), but the 2.5-point gap over Sol's 92.5% sits inside that benchmark's 9.2-point margin — a tie, not the confident jump the raw scores alone would suggest. DeepSWE compares Astra's own best config (74% at xhigh) against Sol's own best (73% at max) — a mismatched-tier comparison by construction, though checking Astra at the same max tier Sol uses (73%, per deepswe.datacurve.ai's full effort-level table) shows the two are functionally tied regardless of which tier is used. Terminal-Bench 2.1 can't be compared at all: Astra's only tracked score comes from tbench.ai's canonical board, while Sol's comes from Artificial Analysis's own run — two different harnesses on the same benchmark, which this site records as Setup-dependent rather than guessing which one would win on the other's turf.

OpenAI's own System Card discloses that Astra is the first model to cross the "Critical" cybersecurity capability threshold under its Preparedness Framework — confirmed on both the pre-launch "Path to Astra" post (2026-09-01) and the System Card itself (2026-09-03), not just press paraphrase. The evidence OpenAI cites: 100% on its internal ExploitBench (versus 78.5% for Sol), including two previously-unknown zero-days found on a version of that benchmark built entirely from post-cutoff CVEs to rule out contamination; and internal red-team exercises where Astra, given only a starting goal, built a full sandbox-escape exploit chain against a hardened browser and a privilege-escalation chain against a hardened OS kernel — both previously undisclosed vulnerabilities, both since reported to maintainers. Disclosed mitigations include a refusal boundary tuned to block proof-of-concept exploit requests by default (91.5% refusal rate on an internal cyber-jailbreak set, versus 59% for Sol) and a production monitoring layer watching tool-using sessions across ChatGPT, Codex, and the API. OpenAI also discloses, without spinning it as solved, that Astra's own monitorability has gone down relative to Sol: in adversarial tests explicitly instructing it to evade detection, it can sometimes do so undetected on sabotage-style tasks.

A more capable, safety-loosened sibling exists for vetted cyber-defense customers, but it is not Astra: OpenAI's "Daybreak Red" access tier — the one that unlocks proof-of-concept exploit generation — runs on a separate, earlier model called GPT-5.6-Cyber, not on Astra. A lighter tier for Astra itself, "Daybreak Blue," is not yet broadly available either: OpenAI's own Help Center states reduced refusals aren't available on Astra for most Daybreak customers, and the benchmark figures OpenAI cites for Astra under Daybreak Blue (a 92% proof-of-concept-exploit completion rate, versus 2.4% under standard safeguards) came from a special internal test configuration OpenAI itself labels not the default production setup — not from general access. OpenAI says it plans to expand Daybreak access for Astra in the coming weeks, without committing to a specific Astra-based Daybreak Red tier. Where this site has recorded a similar-looking restricted sibling for other vendors (Claude Mythos 5.1, Gemini 3.8 Flash Cyber), those are the same-generation model with fewer guardrails; Astra's cyber-capable counterpart today is a different, older model entirely, and Astra's own reduced-safeguard access is still rolling out.

Astra does not appear under its own name inside ChatGPT's model picker — it ships there as "GPT-6 Pro," a separate option from "GPT-5.6 Sol Pro" with its own usage caps that vary by plan, per OpenAI's own Help Center article. Two of OpenAI's own pages disagree on what to call the initial enterprise rollout program ("Trusted Access Program" on one, "Daybreak Access Program" on another) — a naming inconsistency this site is disclosing rather than resolving, since OpenAI itself hasn't. The same two pages also disagree on ChatGPT Plus: the announcement says Astra reaches all Plus users "over the coming days," while the Help Center states GPT-6 Pro is not included with Plus in Chat — reported here as the conflict it is.