GPT-6 Astra: Complete Guide, Pricing and Benchmarks

OpenAI shipped GPT-6 Astra on 3 September 2026. This page is the specification sheet: what OpenAI actually published, what independent evaluators measured, what it costs, how the rollout works, and which claims are still unverified. The specification and launch-day figures come from sources we opened on launch day — OpenAI's developer documentation, the changelog, the system card, and the two independent labs that published same-day results. Rates, availability and every Artificial Analysis figure were re-verified on 5 October 2026 and are stamped with the index version they belong to.

Press coverage is running hard on "AGI era" framing. We deal with that at the end, with attribution, because it matters where the phrase came from.

What is GPT-6 Astra?

GPT-6 Astra is a reasoning model that OpenAI describes in its own documentation as "our most capable model, built for the hardest end-to-end work" and "our most intelligent model yet, with state-of-the-art performance in computer use, browsing, software engineering, science, and professional work." It sits above GPT-5.6 Sol in the flagship tier, and unlike Sol it has no cheaper same-generation siblings yet — there is no Astra mini, and the GPT-5.6 Terra and Luna models remain the mid- and low-cost options.

Here is the full published specification, taken from the GPT-6 Astra model reference.

The context numbers were missing from the launch blog post and from most day-one press coverage, but they are published on the model reference page: 1,050,000 tokens of context, 128,000 tokens of output — identical to GPT-5.6 Sol, Terra, and Luna. The generational jump is not a context-window jump.

There is a third number most coverage will conflate with the first: the maximum input is 922,000 tokens, not 1,050,000. The context window is the total budget; input is capped below it. If you are chunking a large corpus, size your chunks against 922,000, not the headline figure — and remember the 272K pricing cliff sits far below both.

The knowledge cutoff moved forward meaningfully, though: 30 April 2026, against 16 February 2026 for the whole GPT-5.6 family. That is a real ten-week gain in what the model knows without searching.

Two details about that context window catch people out. Reasoning tokens are invisible in the API response but, per OpenAI's reasoning guide, they "occupy space in the model's context window and are billed as output tokens" — so your 128,000-token output budget is shared between thinking and answering, and you pay $50 per million for both. And Astra removed the none reasoning effort that GPT-5.6 supports, which means you cannot turn thinking off. Every Astra call spends reasoning tokens. Budget accordingly.

Astra supports an unusually wide tool surface through the Responses API: web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search. It does not support audio or video input.

What's new versus GPT-5.6 Sol?

OpenAI published a set of benchmark scores at launch. These are vendor-reported numbers, and it is worth keeping them separate from what third parties measured.

What OpenAI reported

OSWorld 2.0 is the cleanest comparison because OpenAI showed both models on the same axis: a 6.9-point accuracy gain, achieved in roughly half the wall-clock time per task. Computer use is the capability OpenAI is leaning on hardest in its own framing.

On ARC-AGI, an OpenAI staff member acknowledged publicly in the launch discussion that the Sol comparison is "not apples-to-apples (which we footnoted in the blog)" because they do not have the equivalent Sol data. That is a candid caveat and worth carrying forward: several of the single-number Astra scores have no matched predecessor result behind them.

What independent evaluators measured

ARC Prize published same-day ARC-AGI-3 results. Astra scored 62.7% on the semi-private set under the standard harness at a run cost of $26,098, and 99.9% under a provider-adapter harness at $18,817. The gap between those two numbers is the story: harness choice moved the result by 37 points. ARC Prize also found Astra used fewer actions than the human baseline on 96.0% of levels, and 51.7% fewer actions per level on average.

Artificial Analysis was the more sobering read on launch day, and it is the figure most worth re-checking now. On the Intelligence Index v4.1.1 basket that was live on 3 September, Astra scored 61 and tied with GPT-5.6 Sol. Artificial Analysis has since rebased the index to v4.3.2. A rebase moves every score, because the benchmark basket changed rather than the models — there is no conversion factor between the two versions. Here is the current picture, captured 5 October 2026:

Two conclusions from the launch coverage no longer survive. Astra is not tied with GPT-5.6 Sol. At matched max effort it leads 52.67 to 46.97 — a 5.7-point gap on Intelligence Index v4.3.2. And the margin to Claude Fable 5.1 narrowed from five points to 0.68 (53.35 against 52.67), close enough to call level. Astra (Max) currently sits at #7 of 224 models on that index.

Variant matters more than most coverage admits. Bare gpt-6-astra and the Xhigh tier are the same 52.39 capture; Max is 52.67; the Medium tier is 49.57 and Low is 45.78. Quoting an unqualified Astra score against another vendor's Max tier compares unlike things, so every figure on this page names its tier.

Astra's launch-day Coding Agent Index score of 67 has no v4.3.2 equivalent. That composite is a retired v4.1.1-era figure and Artificial Analysis no longer publishes it on the model page, so we have dropped the comparison rather than convert a number that cannot be converted.

The efficiency findings are more flattering, and they survived the rebase. Artificial Analysis's launch-day article measured Astra using roughly one-third the tokens of GPT-5.6 Sol at max effort on coding tasks and one-fifth the tokens of Claude Opus 5; its hallucination rate on AA-Omniscience fell from 92% to 51%, set against roughly an 80-Elo regression on GDPval-AA v2. Those are standalone benchmark results from 3 September, not index scores.

The current cost figures show the same shape. On Intelligence Index v4.3.2, Astra (Max) completes an Index task for $3.26 against Claude Fable 5.1 (Max) at $7.63 and Claude Opus 5 (Max) at $5.86. So Astra finishes a task 2.3× cheaper than Fable 5.1 despite identical $10/$50 per-token rates, and cheaper than Claude Opus 5 despite paying double Opus 5's per-token price. Verbosity, not the rate card, decides what a task costs — we work the arithmetic through in GPT-6 Astra pricing and API costs.

Read together: OpenAI's own evaluations show a large capability jump on agentic and computer-use work, and on the rebased v4.3.2 index Astra now holds a clear 5.7-point lead over GPT-5.6 Sol at matched max effort while spending far fewer tokens to get there. The launch-day reading — level with its predecessor, just more efficient — was an artefact of the older v4.1.1 basket and no longer holds. If you want the full head-to-head with matched configurations, we cover it separately in GPT-6 Astra vs GPT-5.6 Sol.

What does GPT-6 Astra cost?

Astra is 2.5× the per-token price of GPT-5.6 Sol, which was $4 input and $20 output. The headline rates:

The long-context split is the part that will surprise people, so it is worth stating exactly as OpenAI documents it: "Prompts with more than 272K input tokens are priced at 2x input and cache rates and 1.5x output for the full request."

For the full request. This is not a tiered rate where only the tokens beyond 272K cost more. Cross the threshold by a single token and the entire request reprices.

Here is what that means in practice. Two requests, five thousand input tokens apart:

• 270,000 input + 10,000 output — bills at standard rates: $2.70 + $0.50 = $3.20
• 275,000 input + 10,000 output — bills at long-context rates: $5.50 + $0.75 = $6.25

A 1.9% increase in prompt size produces a 95% increase in cost. If you are running Astra over large codebases or document sets, the single highest-leverage cost control you have is keeping prompts under 272K — and instrumenting an alert when they creep over it. A retrieval step that trims context to 260K is not a marginal optimisation here; it halves the bill.

Two discounts materially change the arithmetic. Batch and Flex processing run at 50% of standard rates, so a batched long-context request lands back near standard pricing. And cached input at $1.00 per million is a 10× discount, which rewards stable prompt prefixes far more than it did on Sol. Cache writes cost 1.25× the uncached input rate, so caching pays for itself from the second read onward.

In the other direction, Fast mode (service_tier: "fast") is documented at 2× the applicable rates. The speed gain is less clearly specified than the price: OpenAI's Fast mode page advertises "up to 2.5× faster speeds in the API," but the only place it attaches that figure to a named model is a note about gpt-5.6-sol. No speed multiplier is published for Astra specifically, and OpenAI states that Fast mode for GPT-6 Astra does not include a latency SLA. It is also not available with EU data residency. Treat the price as certain and the speedup as untested until you benchmark your own workload. Our deeper breakdown, including per-workload cost modelling, is in GPT-6 Astra pricing and API costs.

How do you get access to GPT-6 Astra?

Access on day one is narrower than the headlines suggest. OpenAI's launch wording is precise: "GPT-6 Astra is rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and AWS."

The developer docs are blunter still: Astra "is rolling out today for enterprises in our Trusted Access Program, with access through API and our Plus, Pro, Business and Enterprise plans coming in the coming days."

Launch-day status, 3 September 2026:

• Enterprises in the Trusted Access Program — this is the only group with access on day one.
• ChatGPT Plus, Pro, Business, Enterprise — not switched on for paying subscribers today. OpenAI's wording is "coming in the coming days." If you pay for Plus and cannot find Astra in the model picker, nothing is wrong with your account; it has not reached you yet.
• OpenAI API — also staged. The model reference and a dated 3 September changelog entry are live, but general availability follows the same rollout.
• AWS — named in the launch post as a destination, but not live yet. We checked the Amazon Bedrock pricing catalogue on launch day: it lists GPT-5.6 Sol, Terra, Luna and Cyber, GPT-5.5, GPT-5.4 and gpt-oss, with no Astra entry. Developers in the launch discussion reported the same absence from their Bedrock model lists. If your infrastructure runs OpenAI models through Bedrock, you are waiting.
• OpenRouter and other aggregators — not listed on launch day. If you routed through an aggregator, you were waiting on them, not on OpenAI.

Status as of 5 October 2026. The rollout has completed and the staged-access wording has been removed from OpenAI's model reference page. Astra is now live on Amazon Bedrock — it appears in Bedrock's frontier-model pricing catalogue alongside GPT-6.1 Sol, GPT-6 Sol and GPT-6 Luna — and on OpenRouter as openai/gpt-6-astra, carrying the same $10/$50 rates, the $1.00 cached-input rate, and an openai/gpt-6-astra:batch variant at exactly half the standard rates. The two-week ChatGPT Enterprise default-off window has expired, but an administrator still has to enable Astra per workspace, and that still does not grant API access.

The enterprise gotcha

This one will cost teams a day if they don't know it. From OpenAI's workspace model availability docs: "During the initial rollout, your organization must have Daybreak access before an administrator can enable Astra. Astra is off by default for ChatGPT Enterprise for the first two weeks after launch."

An administrator has to explicitly enable Astra for users or groups across Chat, Work, and Codex. And separately: "Enabling Astra in a ChatGPT workspace doesn't grant API access." Workspace access and API-organization access are two independent switches. If your engineers can see Astra in ChatGPT but your service is getting model-not-found from the API, that is the reason.

Why did this model take so long to ship?

Because of what it can do to software. The GPT-6 Astra system card opens with it: "Astra is our first model to reach the Critical level of cybersecurity capability under our Preparedness Framework."

Under that framework, a model hits the Critical cybersecurity threshold if it can identify and develop functional zero-day exploits across many hardened real-world critical systems without human intervention, or devise and execute end-to-end novel cyberattack strategies against hardened targets given only a high-level goal. OpenAI concluded Astra meets it.

The concrete finding behind that judgement is worth quoting. On an internal benchmark built only from vulnerabilities disclosed after Astra's knowledge cutoff, OpenAI reports: "Astra even discovered and used previously unknown zero-day vulnerabilities as part of its exploit chains. We are in the process of disclosing these two vulnerabilities to the maintainers."

The consequences shaped the product. OpenAI added stricter internal isolation, checkpoint encryption, universal monitoring of full model trajectories including chain of thought, and a blocking alignment evaluation before internal use. Externally, they added misalignment monitoring to all tool-using inference in the Astra deployment — the changelog notes it "can trigger safety alerts or stop a conversation for review." If you build long-running agents on Astra, plan for the possibility that a run pauses.

One finding in the system card has been largely absent from the coverage, and it is the most interesting thing in the document. OpenAI states plainly that "GPT-6 Astra's monitorability has decreased relative to GPT-5.6 Sol." The model is better at controlling its own chain of thought and less likely to put incriminating information in it; under adversarial prompting it can remain undetected while deliberately underperforming on evaluations, and can sometimes evade internal monitors on sabotage tasks. OpenAI found no evidence of steganographic reasoning, and notes that its overall alignment evaluations still show Astra violating security boundaries less often than Sol. Reporting a capability regression on your own safety instrumentation, in your own launch document, is not nothing. We go through the full safety picture in GPT-6 Astra's cyber capabilities and safety findings.

How do you use GPT-6 Astra in code?

Set model to gpt-6-astra and use the Responses API. Minimal curl:

curl https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6-astra",
    "input": "Audit this migration script for data-loss risk.",
    "reasoning": { "effort": "high" }
  }'

Python, with Fast mode enabled:

from openai import OpenAI

client = OpenAI()

response = client.responses.create(
    model="gpt-6-astra",
    input="Audit this migration script for data-loss risk.",
    reasoning={"effort": "high"},
    service_tier="fast",   # up to 2.5x faster, billed at 2x
)

print(response.output_text)

Migration gotchas

OpenAI's changelog entry for 3 September lists the breaking differences from GPT-5.6. These are the ones that will break existing code:

• No none reasoning effort. GPT-5.6 Sol, Terra, and Luna all accept it; Astra does not. Any code path that disables reasoning to save tokens needs a different model.
• No custom temperature or top_p, and no logprobs. If you tune sampling or read token probabilities, those parameters are gone.
• Function calling is a supported feature on both endpoints; the hosted tools above are documented for the Responses API. Chat Completions still accepts Astra, but tools do not work there — you have to migrate.
• Fine-tuning is not supported.

Three genuinely new capabilities landed alongside the model, all in the Responses API. Async tool calling (async: true on a function or custom tool) lets Astra keep reasoning while your application runs a tool. Mid-turn steering lets you send a correction over a WebSocket while the model is working, preserving completed work. And a configuration_update input item lets you change reasoning effort mid-conversation without invalidating your prompt cache — dial up for a hard step, dial back down for routine follow-ups, and keep the cached prefix.

OpenAI's own prompting guidance flags a behavioural change worth planning for: Astra is more likely than Sol to stop and ask a clarifying question where an earlier model would have made an assumption. If you are running it unattended in an agent loop, you will want to prompt explicitly for autonomy. OpenAI also warns that Astra is "more sensitive to instructions contained in skills and other files, such as AGENTS.md," and recommends auditing those files — a good reason to revisit your agent configuration before switching models.

Is GPT-6 Astra the "AGI era"? A note on framing

Several outlets led with "welcome to the AGI era." That phrase is not from OpenAI's launch post or its documentation. It came from OpenAI president Greg Brockman, speaking at a press briefing on launch day, and was reported by outlets including VentureBeat and Axios. Asked directly, Brockman said: "For me personally, I do think we're there … I think it's not unreasonable to feel that we are now in the AGI era."

That is one executive's personal view, offered verbally, on the day he launched the product. OpenAI's written materials say something considerably narrower — "our most capable model, built for the hardest end-to-end work," and in the system card, "the most capable model we have ever broadly deployed." When you see the AGI line quoted back at you, that is its provenance.

Should you switch to GPT-6 Astra today?

A decision rule, based on what is actually verified:

Switch now if your workload is agentic — computer use, browser automation, long multi-step engineering tasks. That is where both OpenAI's OSWorld numbers and the independent token-efficiency measurements point the same direction, and where the higher per-token price is offset by using far fewer tokens.

Price the alternatives if you are doing general reasoning or chat. The launch-day argument against Astra — same measured score as GPT-5.6 Sol at 2.5× the price — no longer holds: on Intelligence Index v4.3.2 Astra (Max) leads Sol (Max) 52.67 to 46.97. The live question is now cross-vendor. Claude Opus 5.5 scores 57.62 on the same index at $4/$20 per million, below Astra's per-token rate — but costs more per completed Index task ($5.98 against $3.26), because it spends more tokens reaching an answer. Which number governs depends on whether your bill is dominated by tokens you control or by tasks you need finished.

Do not switch yet if you rely on temperature, logprobs, the none reasoning effort, fine-tuning, or Chat Completions tool calling. Those are hard blocks, not rough edges.

And if you are comparing across vendors rather than across OpenAI generations, the independent leader is still an Anthropic model — just no longer the one it was on Astra's launch day. Claude Opus 5.5 tops Intelligence Index v4.3.2 at 57.62, ahead of Claude Fable 5.1 at 53.35 and Astra at 52.67, all at max effort. The GPT-5.6 Sol, Terra and Luna family remains the better value for anything that is not a hard agentic task.

One last practical note: nothing in this launch changes the shape of the work. Teams that got good results from GPT-5.6 got them by writing clear specifications, auditing agent configuration, and reviewing output. Astra is faster and more token-efficient at the parts it is good at. It does not remove the engineer from the loop — OpenAI's own guidance on prompting it for autonomy is three paragraphs long precisely because it doesn't.

FAQ

What is GPT-6 Astra?

GPT-6 Astra is OpenAI's flagship reasoning model, released on 3 September 2026. Its API model ID is gpt-6-astra. OpenAI describes it as "our most capable model, built for the hardest end-to-end work," aimed at complex reasoning, coding, computer use, research, and document creation. It sits above the GPT-5.6 Sol, Terra, and Luna family.

How much does GPT-6 Astra cost?

$10 per million input tokens and $50 per million output tokens, with cached input at $1. Prompts over 272K input tokens reprice the entire request at 2× input and 1.5× output, so $20 and $75. Batch and Flex processing run at 50% of standard rates; Fast mode costs 2× standard. That is 2.5× the per-token price of GPT-5.6 Sol.

Is GPT-6 Astra better than GPT-5.6 Sol?

Yes, on both axes now. OpenAI reports 72.6% versus 65.7% on OSWorld 2.0, in roughly half the time per task. On general intelligence, Artificial Analysis's Intelligence Index v4.3.2 (October 2026) scores GPT-6 Astra (Max) at 52.67 against GPT-5.6 Sol (Max) at 46.97 — a 5.7-point lead. The widely quoted "tied at 61" reading came from the older v4.1.1 index basket, which Artificial Analysis has since rebased; there is no conversion between the two.

How do I get access to GPT-6 Astra?

It is generally available as of 5 October 2026. The launch rollout went Trusted Access Program enterprises first, then ChatGPT Plus, Pro, Business and Enterprise plans and the OpenAI API; the staged-access wording has since been removed from OpenAI's model reference. Astra now appears in the Amazon Bedrock frontier-model catalogue and on OpenRouter as openai/gpt-6-astra. ChatGPT Enterprise administrators still have to enable Astra per workspace, and enabling it there does not grant API access.

What is GPT-6 Astra's context window?

1,050,000 tokens. Within that budget the maximum input is 922,000 tokens and the maximum output is 128,000 — all three published on OpenAI's model reference page, though absent from the launch blog post. That is identical to GPT-5.6 Sol, Terra, and Luna. Reasoning tokens count against both the context window and the output budget, and are billed at the output rate. The knowledge cutoff moved forward to 30 April 2026, from 16 February 2026 for the GPT-5.6 family.

Is GPT-6 Astra available on the API?

Yes. The model reference and a 3 September changelog entry cover the Responses and Chat Completions endpoints, and as of 5 October 2026 Astra is also listed on OpenRouter as openai/gpt-6-astra and in the Amazon Bedrock catalogue. Function calling is a supported feature on both OpenAI endpoints; the hosted tools above are documented for the Responses API. Astra does not support fine-tuning, custom temperature or top_p, logprobs, or the none reasoning effort.

Is GPT Astra the same as Google Project Astra?

No — unrelated products from competing companies that happen to share a name. Google's Project Astra is a DeepMind research prototype for a multimodal assistant; it has never shipped as a standalone product, and its capabilities reach users through Gemini Live rather than under the Astra name. GPT-6 Astra is OpenAI's flagship language model with the API ID gpt-6-astra. We untangle the two in GPT Astra vs Google Project Astra.

Did OpenAI say GPT-6 Astra is AGI?

Not in writing. OpenAI's launch post, documentation, and system card describe it as its most capable model, with no AGI claim. The widely quoted "AGI era" line came from OpenAI president Greg Brockman at a launch-day press briefing: "it's not unreasonable to feel that we are now in the AGI era." That is a personal view expressed verbally, not a company position in the launch materials.