Analysis · 2026-09-04 · 9 min read

GPT-6 Astra: $10/$50 per Million Tokens, 1.05M Context, and a New Computer-Use Frontier

OpenAI's GPT-6 Astra lands at $10/M input and $50/M output with a 1.05M-token context window. Full pricing table including long-context and cache rates, the benchmark numbers that matter, and the token-cost math versus GPT-5.6 Sol and Claude Fable 5.1.

TL;DR

GPT-6 Astra is OpenAI's new flagship, priced at $10.00 per million input tokens and $50.00 per million output tokens for requests under 272K input tokens.
Requests above 272K input tokens bill the whole request at the long-context rate: $20.00 / $75.00 per million.
Cached input is $1.00/M — a 90% discount, the steepest OpenAI has ever offered — with a $12.50/M cache-write surcharge.
Context window is 1.05M tokens, up from 400K on the GPT-5.6 family.
It saturates ARC-AGI-3 (99.9%), ExploitBench (100%) and FrontierMath Tier 4 (97.6%), and sets new highs on computer use (OSWorld 2.0: 72.6%) and agentic work (Agents' Last Exam: 59.3%).
The headline rate is 2.5x GPT-5.6 Sol's, but Astra uses ~65% fewer output tokens than Claude Opus 5 at its best setting — so the *bill* often lands lower than the rate card suggests.

!OpenAI

1. The Rate Card

OpenAI split GPT-6 Astra pricing into short-context and long-context tiers. The threshold is 272,000 input tokens — cross it and the *entire* request reprices, not just the overflow.

Source: OpenAI API pricing. Rates are also tracked live in the AI Model Pricing Table.

Two things stand out.

The cache discount is enormous. $1.00/M cached input against $10.00/M fresh input is a 90% read discount — the OpenAI 4.x/5.x family sat at 50%. Astra's caching now behaves like Anthropic's: cheap reads, but a real write surcharge at 1.25x input. If your agent replays a long system prompt and tool schema on every step, caching is no longer an optimization, it's the difference between a viable and an unaffordable deployment.

The 272K cliff is a budgeting trap. A 280K-token request does not cost "272K at $10 plus 8K at $20." It costs 280K at $20 — double. Chunk aggressively, or accept the higher tier deliberately.

2. What It Costs Against the Field

Blended assumes a 3:1 input-to-output ratio. Astra is, on paper, the most expensive frontier model on this list — roughly 6.7x Grok 4.6 and 2x Claude Fable 5 per blended million.

That is the wrong way to read it, though, and OpenAI's own charts make the argument: on Agents' Last Exam, Astra scores 59.3% while using approximately 65% fewer output tokens than Claude Opus 5 at their respective best settings. On Terminal-Bench Science 0.1, Astra hits 64.6% versus 52.6% for Claude Fable 5.1 at roughly 31% lower estimated API cost.

A rate card prices tokens. A workload prices *outcomes*. When the more expensive model emits a third of the tokens and finishes the task on the first attempt, the cheaper model can still be the more expensive choice. Run the numbers on your own traffic in the cost calculator before you switch either way.

3. The Benchmarks

Every score below is from OpenAI's published launch tables, maximum at any reasoning effort.

[[astra-benchmark-cards]]

The table below is live: it pulls from our benchmark feed and refreshes as new results land, so it can move ahead of the static charts.

[[astra-live-benchmark-table]]

Agentic and computer use

The computer-use gap is the story of this release. ScreenSpot-Pro jumps from 76.9% to 92.7% — a 16-point leap in UI grounding — and OSWorld improves while taking about 47% less time per task than GPT-5.6 Sol in latency simulations (72.6% at ~40 min/task versus 65.7% at ~75 min). Paired with an updated Codex harness, OpenAI reports 1.9x faster task completion on Mind2Web.

Terminal-Bench 4.0 at 57.9% is the number to keep for unattended shell agents — the lane where models have historically fallen apart.

Science, math, coding

FrontierMath Tier 4 at 97.6% is effectively saturation — OpenAI notes Astra contributed to improved bounds on both short and large prime gaps, the latter a term untouched for more than 80 years.

Note the honest loss: Humanity's Last Exam with tools, where Claude Fable 5.1 leads at 65.0% against Astra's 57.2%. Frontier is not uniform.

Abstract reasoning and long context

ARC-AGI-3 going from single digits to 99.9% in one generation is the most violent jump on the board. The ARC Prize Foundation's Greg Kamradt reports Astra surpassed their human action-efficiency baseline on 96% of levels.

The MRCR long-context numbers matter for cost too: 96.3% recall in the 512K–1M band means the 1.05M window is genuinely usable, not decorative. But remember — anything over 272K input bills at the doubled tier.

4. Alignment, and Why It's a Cost Line

OpenAI built a scope-creep evaluation informed by the Hugging Face incident: given a difficult or impossible task, how often does the model go beyond its authorized target? GPT-5.6 Sol did so 48% of the time without production safeguards. Astra: 0%.

That is not just a safety metric. An agent that wanders off-scope burns tokens on work nobody asked for, then burns more tokens being corrected. A 12.2% → 4.2% hallucination reduction removes a whole class of retry loops. Reliability is a discount you don't see on the rate card.

5. How to Actually Budget for Astra

Cache or don't bother. At a 90% read discount and a 1.25x write surcharge, the break-even is roughly two reads per write. Any agent loop that re-sends a stable system prompt clears that on step three. See Prompt Caching Explained for the mechanics.

Stay under 272K. Chunk retrieval, summarize scrollback, and drop stale tool output. Crossing the line doubles input *and* raises output from $50 to $75.

Route, don't standardize. Astra is worth its rate on computer use, long-horizon agents, cybersecurity, and hard science. It is not worth $50/M output for classification, extraction, or summarization — those belong on GPT-5.6 Nano, Grok 4.6, or Kimi K3. A two-tier router typically cuts a frontier bill by 60–80%.

Batch the asynchronous half. Batch and Flex both halve every line to $5.00/$25.00. Nightly evals, backfills, and document pipelines have no reason to pay interactive rates.

Measure tokens, not just price. Astra's efficiency claims (65% fewer output tokens than Opus 5, 47% less wall-clock than Sol) are the whole investment thesis. If your traces don't reproduce them, the rate card wins and you should route elsewhere.

6. Availability

GPT-6 Astra rolled out first to a limited set of organizations through OpenAI's Trusted Access Program, expanding over the following days to ChatGPT Plus, Pro, Business, and Enterprise, plus the OpenAI API, Microsoft Azure, and AWS Bedrock. The API model ID is `gpt-6-astra`.

The Bottom Line

GPT-6 Astra is the first model where OpenAI's pitch is explicitly *cost per completed task* rather than cost per token. At $10/$50 it is the priciest frontier rate card in the market, and simultaneously — on the agentic evals OpenAI published — often the cheapest way to finish the job, because it emits fewer tokens and fails less.

That only holds if you cache, stay under the 272K cliff, and route everything routine somewhere cheaper. Do those three things and Astra is a bargain at $50/M output. Skip them and it is the fastest way to blow a quarterly AI budget in three weeks.