GPT-6 Astra is OpenAI's flagship language model, released September 3, 2026. It takes text and image input, returns text, and carries a 1,050,000-token context window with 128,000-token maximum output. It costs $10 per million input tokens and $50 per million output tokens, with cached input at $1 per million. OpenAI positions it for complex reasoning, agentic coding, computer use, research and document creation, and supports reasoning effort levels from low through max. The knowledge cutoff is April 30, 2026.

How much does GPT-6 Astra cost?

Standard API pricing is $10 per million input tokens and $50 per million output tokens as of September 2026. Cached input reads cost $1 per million, and cache writes are reported at $12.50 per million. Batch and flex processing run at 50% of standard rates. Prompts above 272,000 tokens are billed at a long-context premium, roughly doubling the effective input rate. That makes Astra one of the most expensive models on the market per token — its cost advantage comes from token efficiency, not price.

Is GPT-6 Astra available to everyone?

Not at launch. OpenAI began the September 3, 2026 rollout through its enterprise Trusted Access Program, with API access and ChatGPT Plus, Pro, Business and Enterprise plans following in the days after, plus Amazon Web Services availability. There is no realtime audio support and no fine-tuning at launch. The lineup is GPT-6 Astra and GPT-6 Astra Pro — OpenAI did not ship Luna, Terra and Sol tier variants the way it did for GPT-5.6.

Is GPT-6 Astra actually AGI?

No, and the claim should be read as marketing. OpenAI president Greg Brockman suggested in a press briefing that GPT-6 Astra could be 'artificial general intelligence' and OpenAI framed the launch as the 'AGI era.' Independent measurement is more restrained: Artificial Analysis scored Astra at 61.1 on its Intelligence Index at max effort, behind Claude Fable 5.1 at 65.6 and Claude Opus 5 at 63.0. Astra does lead decisively on ARC-AGI-3 abstract reasoning at 62.7% against 30.2% for Opus 5, and it leads every model on token efficiency.

Quick Answer

What Is GPT-6 Astra? OpenAI's New Flagship Explained

Published: September 4, 2026

The Short Answer

GPT-6 Astra is OpenAI’s flagship model, released September 3, 2026. It costs $10 per million input tokens and $50 per million output tokens, carries a 1,050,000-token context window, and takes text and image input.

The three things that actually matter:

It is not cheap per token, but it is cheap per task. Astra emits far fewer output tokens than any rival, which flips the cost comparison against nominally cheaper models.
Access is staged. Enterprise Trusted Access first, then API and ChatGPT paid plans.
The “AGI” framing is marketing. Independent scoring puts it behind Claude Fable 5.1 on raw intelligence.

Last verified: September 4, 2026.

Specifications

	GPT-6 Astra
Released	September 3, 2026
Input / MTok	$10.00
Cached input / MTok	$1.00 (writes ~$12.50)
Output / MTok	$50.00
Context window	1,050,000 tokens
Max output	128,000 tokens
Input modalities	Text, image
Output modalities	Text
Knowledge cutoff	April 30, 2026
Reasoning effort	low → medium → high → xhigh → max
Batch / flex	50% of standard rates
Long-context threshold	272,000 tokens (premium above)
Not available	Realtime audio, fine-tuning

The lineup is GPT-6 Astra and GPT-6 Astra Pro. Unlike GPT-5.6, OpenAI did not announce Luna, Terra and Sol tier variants — so there is no cheap GPT-6 tier yet, and the GPT-5.6 family remains the budget option inside OpenAI’s own catalogue.

What It Is Actually For

OpenAI positions Astra for complex reasoning, agentic coding, computer use, research and document creation. Tool use spans web search, file search, code interpretation, hosted shell and computer use.

The reasoning-effort dial is the important operational detail. Astra runs from low to max effort, and OpenAI notes that unless stated otherwise its published evaluations ran at maximum effort. That improves benchmark results while increasing latency and token consumption — so a benchmark table and your production bill describe different configurations of the same model. Pick the effort level deliberately rather than inheriting the default.

The Benchmark Picture, Honestly

Two layers exist here, and mixing them is how people get the story wrong.

Layer one: the leaked launch table. A comparison table circulating with the rollout showed Astra ahead of Claude Fable 5.1, Fable 5 and Opus 5 on every comparable benchmark — including a claimed 97.6% on FrontierMath Tier 4 v2 against 87.8% for Fable 5.1, 96.0% on GPQA Diamond, 64.6% on Terminal-Bench Science 0.1, 74.1% on DeepSWE v1.1 and 41.4% on AutomationBench. These figures are unofficial and largely unverified.

Layer two: independent measurement. Artificial Analysis published its verdict on September 3, 2026 and it is more mixed:

Metric	GPT-6 Astra	Claude Fable 5.1	Claude Opus 5
AA Intelligence Index	61.1 (max)	65.6	63.0
AA Coding Agent Index	67.0 (xhigh)	70.4 (max)	68.1 (xhigh)
Cost per coding task	$3.27 (xhigh)	$9.18	$8.17
ARC-AGI-3 (standard harness)	62.7%	not scored	30.2%

So: Astra does not lead the raw intelligence leaderboard. Fable 5.1 does. What Astra leads is abstract reasoning — its 62.7% on ARC Prize’s official standard ARC-AGI-3 harness more than doubles Opus 5’s 30.2% and dwarfs GPT-5.6 Sol’s 7.8% — and efficiency, where it is genuinely in a class of its own.

The Token-Efficiency Story

This is the part most launch coverage buries, and it is the actual reason to care about Astra.

Artificial Analysis measures tokens consumed per task, not just price per token. Running the full AA Intelligence Index takes Astra (high) about 16 million output tokens versus about 123 million for Gemini 3.8 Flash (high) — a 7.7x gap that erases Flash’s roughly 13x token-price advantage.

Per task, Astra emits roughly 2,200 to 14,000 output tokens depending on effort level, while Claude Fable 5.1 needs 14,500 to 45,000 and Gemini 3.8 Flash about 48,000. Verbose models also re-read more context and make more tool calls, so their bills compound.

The concrete result: in the AA coding-agent harness, Codex running Astra (low) costs $1.41 per task and scores 62.6, beating Opencode with Gemini 3.8 Flash at $2.04 and 61.1 — cheaper and better, despite a 13x higher sticker price per token.

The honest caveat: the cheapest per-task option outright is still GPT-5.6 Luna at $0.29 (index 57.1), and AA notes Astra sits behind its own cheaper predecessor GPT-5.6 Sol on intelligence-per-dollar because Astra’s prices are 2.5x higher.

The Limits Worth Knowing Before You Commit

Prompts above 272,000 tokens reprice. Long-context requests are billed at a premium — reported as roughly 2x input, which pushes effective long-context rates to around $20 per million input tokens. Filling the 1M window is genuinely expensive, on the order of $10 or more for a single call.
No fine-tuning at launch, so domain adaptation means prompting, retrieval or caching — not weights.
No realtime audio. If you built voice on GPT-5.6’s realtime stack, Astra is not a drop-in.
Staged access means you cannot plan a September migration around it with confidence until your account actually has it.
Benchmarks ran at max effort unless noted. Your production effort setting will score lower and cost differently.

Should You Switch?

Switch for hard reasoning work — research synthesis, abstract problem solving, long-horizon analysis — where the ARC-AGI-3 gap and the token parsimony compound in your favour.

Do not switch high-volume simple workloads. GPT-5.6 Luna at $0.20/$1.20 or Gemini 3.8 Flash at $0.75/$3.75 remain the right answer for extraction, classification and routing. Astra’s efficiency advantage shows up on hard tasks, not easy ones.

Do not switch coding agents reflexively. Claude Fable 5.1 still leads the Coding Agent Index at 70.4, and its $0.25 cache reads cut long agent runs by up to roughly 45%. Astra’s pitch there is comparable quality at a third of the per-task cost, which is a real argument — but it is a cost argument, not a capability one.

Measure before you migrate. Run a hundred representative tasks, compare total spend and completion rate, and set the effort level explicitly. Cost per completed task is the only number that matters.