TL;DR — OpenAI released GPT-6 Astra on September 3, 2026 and opened the API as gpt-6-astra the next day: a new flagship that “finishes the work” — complex reasoning, coding, computer use, research and full document production. The sticker is $10 / $50 per million tokens with a 1.05M context window and 128K output, a 2.5× price jump over GPT-5.6 Sol. Self-reported benchmarks jump in every category that matters for agents: DeepSWE v1.1 at 74.1, OSWorld 2.0 at 72.6, BrowseComp at 91.5, ExploitBench at a perfect 100. Two reads frame the launch: the API quietly defaults reasoning.effort to low despite the “Highest reasoning” card, and the cyber results are gated behind Trusted Access for defenders. The rollout itself was the news: a staggered “limited customers first” release that looked like a leak, then a formal launch 24 hours later (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks).
Introduction
GPT-6 Astra is OpenAI’s answer to a crowded September: Anthropic’s Claude Fable 5.1, Google’s Gemini 3.8 Flash line, and a long tail of open-weight releases. What separates this launch is the positioning. OpenAI is not selling a chat model; it is selling an agent-grade worker — a model built around finishing end-to-end tasks: codebases, browser sessions, research loops, and the production of “finished documents, spreadsheets and presentations” (Source : DEV Community — GPT-6 Astra: OpenAI’s New Model Is Built to Finish the Work).
That framing matters for anyone building on the API today. The last frontier releases were priced and tuned as reasoning engines. Astra is the first OpenAI flagship that reads more like an autonomous operator: a 1.05M context window big enough to hold an entire codebase session, tool support that now includes hosted shell, apply patch, skills and computer use on the Responses API, and a benchmark table organized by job — coding, science, computer use, then cyber (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks).
The Positioning: “Most Capable” Meets a Lab Sentence
OpenAI calls Astra its most capable model “for the hardest end-to-end work.” LLM Stats is blunter: that is a lab sentence, and the numbers on the announcement page are OpenAI’s own, not independently verified (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks). Independent platforms rank it strongly but not unanimously: BenchLM places it #1 of 22 eligible models for reasoning and logic at 88.8/100, while OpenRouter lists it with an 1,050,000-token context and $10/$50 pricing backed by two providers (Source : BenchLM — GPT-6 Astra Benchmarks & Pricing) (Source : OpenRouter — GPT-6 Astra).
The model card adds two architectural tells:
reasoning.effortnow runs five levels — low, medium, high, xhigh, max — where max is the new “top of the ladder” (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks).- The API defaults effort to low. OpenAI markets “Highest reasoning” on the card, but every integration that does not explicitly set effort gets the cheapest thinking mode (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks).
That gap between marketing and default is the single most important implementation detail in the launch: two apps pointing at the same model id can behave very differently depending on whether they raise effort.
The Benchmark Table, Read by Job
OpenAI published a self-reported table. Treat it as directional, not verified — but the shape is consistent: Astra improves most where agents actually work.
Coding. Terminal-Bench 4.0 at 57.7, DeepSWE v1.1 at 74.1 — a small hop over GPT-5.6 Sol’s 72.7 on OpenAI’s own page — and FrontierCode 1.1 Extended at 64.5 (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks). The DeepSWE improvement over Sol is real but modest; this is not the blowout that the “GPT-6” label implies on its own.
Science and research. Terminal-Bench-Science 0.1 at 64.6 versus Claude Fable 5.1’s 52.6 is the comparison pair OpenAI chose to highlight. FrontierMath T4 v2 hits 97.6; GPQA Diamond at 96.0 is effectively saturated and should be read as a ceiling row, not a differentiator (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks).
Computer use and agents. BrowseComp at 91.5, OSWorld 2.0 at 72.6 (offline partial), Agents’ Last Exam at 59.3, AutomationBench at 41.4, ScreenSpot-Pro at 92.7 with no tools and BenchCAD at 95.9 with Python (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks). This is the category that justifies the “finishes the work” pitch: browser navigation, tool orchestration and long-horizon autonomy are where the gains concentrate.
Cyber. ExploitBench at 100% — exploit development from known vulnerabilities — with ExploitGym at 42.4 and SEC-Bench Pro at 85.4 (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks). A perfect cyber score is exactly the kind of number that gets a model policy attention rather than applause.
The Cyber Story: Capability and Guardrails in the Same Release
The most consequential part of the launch is not in the benchmark table. OpenAI’s own Path to Astra document designates Astra at the Critical cybersecurity capability threshold under its Preparedness Framework, effective September 1, 2026. In practice, the most advanced cyber workflows are limited: advanced defender access routes through Trusted Access and the Daybreak Blue program, and the default product is not unrestricted dual-use (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks). The Path document also states Astra was not involved in the Hugging Face incident, a distinction OpenAI is now actively drawing in public (Source : OpenAI — Path to Astra).
This is the new normal for frontier releases: the capability and the restriction ship together. For enterprises evaluating Astra, the security question is no longer “how capable is it at offensive tasks” but “which tier of access will my use case land in.”
The Rollout That Looked Like a Leak
Astra also made news for how it shipped. The September 3 release went to limited customers first — a staggered rollout that looked, from the outside, like an accidental early exposure, and prompted coverage of a “fake release” before OpenAI confirmed and formalized the launch the next day with the API opening (Source : ai.rs — GPT-6 Astra Benchmarks: What the 98.6% on ARC-AGI-3 Hides) (Source : Codersera — GPT-6 Astra vs GPT-5.6 Sol: Should You Upgrade?).
Read generously, the staggered rollout is OpenAI managing load and risk on a model with a Critical cyber designation. Read cynically, it is launch theater. Either way it signals something real about OpenAI’s rhythm in 2026: no more monolithic release days — flagship capability now arrives in waves, and the “release date” you hear on day one is rarely the date the API actually opens for everyone.
Pricing: The Same Sticker, a Different Cost Story
Astra’s Standard price is $10 / $50 per million input/output tokens — the same sticker class as Claude Fable 5.1, but 2.5× GPT-5.6 Sol’s $4 / $20 (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks).
Three pricing details matter more than the headline:
- Cached input at $1 (0.1×) — but Fable 5.1 undercuts to $0.25, and OpenAI has not published cache-hit ratios for Astra (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks).
- Cache writes at $12.50 — a 1.25× surcharge on uncached input, not a discount. Long agent loops that rewrite their context pay for it (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks).
- A 272K-token cliff: prompts over 272K input tokens bill at 2× input and cache rates and 1.5× output for the full request (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks).
The cost per task may still fall — OpenAI argues lower token use from better reasoning can offset the higher sticker — but only for workloads that genuinely finish faster. Batch and Flex run at 50% of Standard; Fast mode is 2×.
What This Means for Builders
For agent builders, Astra changes three calculations:
- Context is now a real product dimension. A 1.05M window plus effort levels up to max means long-horizon jobs that previously required external memory systems can stay in-context — at a price.
- Effort is a dial, not a default. Integrations that never set
reasoning.effortare silently buying the low-effort model. The card’s “Highest” is opt-in. - Cyber capability has an access layer. If your workflow touches security tooling, plan for Trusted Access review, not instant API keys (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks).
FAQ
When was GPT-6 Astra released? OpenAI announced it September 3, 2026 and opened the API model gpt-6-astra September 4. The free API tier is unsupported; access depends on account tier and rollout limits (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks).
What does GPT-6 Astra cost? $10 per million input tokens, $50 per million output (Standard). Cached input is $1; cache writes are $12.50. Prompts over 272K input tokens incur 2× input/cache and 1.5× output. Batch and Flex are 50% off Standard (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks).
What context window does it support? 1,050,000 tokens of context with up to 128,000 output tokens. Knowledge cutoff is April 30, 2026 (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks).
Is GPT-6 Astra better than GPT-5.6 Sol? On OpenAI’s own table, yes across the board — but modestly in places (DeepSWE 74.1 vs 72.7) and at 2.5× the token price. The bigger jump is architectural: five effort levels, 1.05M context, and a tool stack built for autonomous work (Source : Codersera — GPT-6 Astra vs GPT-5.6 Sol: Should You Upgrade?).
Can I fine-tune GPT-6 Astra? No. Fine-tuning is not supported on this card; the model is available through Chat Completions, Responses and Batch only (Source : LLM Stats — GPT-6 Astra: Released Flagship, API Pricing and Benchmarks).