GPT-6 Astra Explained: Pricing, Benchmarks, and What OpenAI Left Out

OpenAI released GPT-6 Astra to approved users on 3 September 2026 and to everyone else the following day. It replaces GPT-5.6 Sol as the flagship, and OpenAI's framing was not subtle — a generational leap in cybersecurity, professional work, software engineering, and science.

Before anything else, a naming note, because it has caused real confusion: the previous model was GPT-5.6 Sol; the new one is GPT-6 Astra. There is no "GPT-5.6 Astra". If you searched for that, this is the model you were looking for.

The interesting question is not whether Astra is capable — it obviously is. It is whether it is worth double what the model it replaced cost, and what OpenAI chose not to put in its own comparison table.

The specs, verified

Pricing, including the tier that catches people out

Verified 9 September 2026 against OpenAI's published API pricing and corroborating trackers.

Three things in that table deserve attention.

The 272K cliff. The context window is 1.05 million tokens, but the standard price only applies up to 272,000 input tokens. Cross that line and the entire request re-prices at $20 in and $75 out — not just the tokens above the threshold. A request at 280K input costs double a request at 270K. If you are feeding large document sets, that boundary is worth engineering around.

Caching is a 10x lever, not a 40x one. Cached input at $1/M is a 90% discount off the standard rate. Worth having — but note that Anthropic's competing Claude Fable 5.1 prices cache reads at $0.25/M against a $10 standard rate, a 97.5% discount. For agent workloads that re-read the same context on every turn, that gap compounds fast.

Batch is half price. If your work is not latency-sensitive — overnight content generation, bulk classification, evaluation runs — Batch and Flex halve the bill for the same model. This is the easiest saving available and most teams forget it exists.

What it costs in practice

Take a mid-sized agent handling 10 million input tokens and 2 million output tokens a day, with 70% of the input served from cache. On Astra that is 3M uncached input at $10/M ($30), 7M cached at $1/M ($7) and 2M output at $50/M ($100) — about $137 a day. The same workload on GPT-5.6 Sol, at $5 input, $0.50 cached and $30 output, comes to about $78.50 a day. That is the price difference showing up in your actual invoice rather than in a spec sheet.

The efficiency argument OpenAI makes is that Astra completes more tasks per attempt, so cost per finished job rises less than cost per token. That may well be true. It is also, at the time of writing, unproven — launch data is too sparse to show whether the efficiency offsets the premium.

How to actually get it

Note how restrictive that is. Astra is not a general consumer upgrade — it is a metered allowance on the most expensive tiers. Two hundred messages a week on a $200/month plan is roughly 28 a day. For comparison, Anthropic includes Claude Fable 5.1 on its $20/month Pro plan.

Benchmarks: the wins, and the omission

OpenAI's announcement is not short of numbers. These are the headline results, with GPT-5.6 Sol and Claude Fable 5.1 where published.

Computer use and agents

Coding

Reasoning and academics

Long context and safety

Some of those are genuinely striking. ARC-AGI-3 going from 7.8% to 99.9% is saturation of a benchmark designed to be hard. Computer-use safety failures dropping from 22% to 2.4% is the difference between an agent you can let touch a browser and one you cannot. The honeypot result — Astra never took the infrastructure shortcut that Sol took almost half the time — is the kind of alignment finding that matters more than a leaderboard position.

Now the caveats

Every score was run at maximum effort unless noted. Maximum effort lifts scores and inflates both latency and token spend. Your production configuration will not be running at max effort on every request, and your results will not match these.

The omitted row. Astra's 57.2% on Humanity's Last Exam with tools trails Claude Fable 5.1 (65.0%), Claude Fable 5 (63.8%), and Claude Opus 5 (63.6%). It is the only academic benchmark in the comparison set that Astra loses — and it does not appear in OpenAI's own announcement table.

The independent index disagrees. Artificial Analysis's aggregated intelligence index (v4.1.1) ranks Claude Fable 5.1 first at 65.7, Claude Opus 5 second at 63.1, and GPT-6 Astra third at 61.2. That directly contradicts the "world's most intelligent model" framing when you step outside vendor-run benchmarks.

FrontierMath has a disclosure. Epoch AI, which runs FrontierMath, notes that OpenAI funded its development and has exclusive access to part of it. The 97.6% is still a real result; it is not an independent one.

The DeepSWE chart is framed generously. OpenAI's chart uses a 67.4% figure for Fable 5.1 that makes the lead look larger than the broader result set supports — and Meta's Muse Spark 1.3 had already topped that benchmark at 75.4%.

One safety metric got worse. OpenAI disclosed that Astra's written reasoning is harder to monitor than Sol's. This is a direct consequence of the recurrent-depth architecture, and it is the honest counterweight to an otherwise strong safety card.

Astra or Fable 5.1?

They launched three days apart at similar frontier positioning, so the comparison is unavoidable.

Same headline token price. Astra takes computer use, cybersecurity, and mathematics by clear margins. Fable 5.1 takes the independent aggregate index, academic reasoning with tools, and — by a wide margin — the cost of running an agent that re-reads its context, thanks to that 4x cheaper cache read.

The consumer story is not close: $20/month against $200/month for frontier access.

Who should pay for it

Worth it if: you are doing security research or red-teaming (the ExploitBench and V8 vulnerability results are in a different class), building computer-use agents that click through real interfaces, working on hard mathematics, or running retrieval over genuinely enormous documents where the MRCR long-context recall numbers matter.

Not worth it if: you are writing content, running marketing workflows, summarising meetings, or building a customer-facing chatbot. None of Astra's wins live in that territory, and you would be paying roughly 1.75x Sol's running cost for capability you never invoke.

Check the boring option first: Batch and Flex at half price, and a lower effort setting, will get most teams a better cost-per-completed-task than reaching for the newest flagship at maximum effort. Measure cost per finished job, not cost per request — a cheaper call that needs three retries is not cheaper.

Related reading

• Claude Fable 5.1 Explained — Anthropic's frontier release from the same week
• GPT-5.6 Sol, Terra and Luna — the tier below Astra, and the cheaper option most workloads should start on
• How to Use OpenAI Codex — OpenAI's agentic coding system and which models it runs
• ChatGPT Review — our full review of the consumer product and its plans
• AI Tool Pricing Index — our verified, regularly re-checked pricing dataset

Pricing, specifications, and benchmark figures were verified on 9 September 2026 against OpenAI's published materials and independent trackers including Artificial Analysis. Frontier model pricing in this family has already changed once since launch — check OpenAI's own pricing page before budgeting against these numbers.