GPT-6 Astra API Pricing: $10 Input, $50 Output

OpenAI has published GPT-6 Astra as a billable API model. The official model ID is gpt-6-astra, and Standard short-context pricing is $10 per million input tokens and $50 per million output tokens. Cached input costs $1 per million and cache writes cost $12.50 per million.

The rollout began September 3 for enterprises in OpenAI’s Trusted Access Program. OpenAI says access through the API and its Plus, Pro, Business, and Enterprise plans will expand over the following days. That makes Astra a priced product, but access remains staged rather than generally available to every account at once.

GPT-6 Astra pricing

Processing and context tier	Input	Cached input	Cache write	Output
Standard, up to 272K input	$10	$1	$12.50	$50
Standard, above 272K input	$20	$2	$25	$75
Batch or Flex, up to 272K input	$5	$0.50	$6.25	$25
Fast mode, up to 272K input	$20	$2	$25	$100

All figures are USD per 1 million tokens. The long-context tier applies to the entire request when input exceeds 272,000 tokens. Batch and Flex are 50% of Standard. Fast mode is twice the applicable Standard rate, and OpenAI says Astra Fast mode is unavailable with EU data residency.

Astra has a 1.05 million-token context window, a 922,000-token maximum input, and a 128,000-token maximum output. It accepts text and images and returns text. The Responses and Chat Completions endpoints are supported, although OpenAI recommends Responses for tool use.

How Astra compares with GPT-5.6

At short context, Astra costs 2.5 times GPT-5.6 Sol’s current $4 input rate and 2.5 times Sol’s $20 output rate. Astra is 50 times GPT-5.6 Luna on both uncached input and output. The premium only pays when Astra’s higher task completion rate, lower retry count, or reduced human review offsets it.

For a workload with 100 million uncached input tokens and 20 million output tokens, Astra Standard costs $2,000. GPT-5.6 Sol costs $800 on the same raw token mix, while GPT-5.6 Luna costs $44. This comparison holds token volume constant; OpenAI says Astra often uses fewer output tokens on difficult work, so teams should compare cost per accepted result rather than list rates alone.

ARC-AGI-3 result and cost

ARC Prize reports that GPT-6 Astra reached 62.7% for $26,098 on the ARC-AGI-3 Semi-Private set with its Standard harness at max reasoning. With OpenAI’s Provider Adapter, which preserves opaque reasoning state between requests and compacts longer conversations, Astra reached 99.9% for $18,817 at high reasoning.

Those scores do not show that one prompt setting universally raises quality while lowering cost. The harnesses expose different state-management capabilities. Across the 167 game-reasoning pairs both harnesses solved, ARC Prize reports the Provider Adapter used 49% fewer total tokens and was about 3.66 times faster by aggregate recorded elapsed time.

ARC Prize also says Astra used fewer actions than its human baseline on 96% of completed levels and 51.7% fewer actions per level on average. This is strong evidence for agentic efficiency inside ARC-AGI-3. It is not a general office-work, coding, or revenue benchmark.

Labs coverage decision

The canonical pricing table now includes Astra, but the Cost-per-Task Leaderboard does not yet show an Astra result. Our 49 deterministic tasks run through a stable routed snapshot with complete token, retry, latency, correctness, and billed-cost records. Astra’s initial access is controlled, and the current benchmark route used by Labs does not expose an authorized gpt-6-astra endpoint.

ARC Prize’s results stay in our benchmark coverage notes because the Standard and Provider Adapter harnesses are stateful game evaluations, not interchangeable with our single-turn task set. We will add Astra after the exact route is available to the harness and a fresh 49-task run passes the normal reproducibility and spend gates.

What buyers should do

1. Confirm that gpt-6-astra is enabled for the exact organization and region before planning a migration.
2. Log input, cached input, cache writes, output, tool calls, retries, wall time, and reviewer minutes.
3. Test short and long requests separately because crossing 272K input changes every token rate for the request.
4. Compare Standard with Batch or Flex when latency is not the constraint; both halve Astra’s token price.
5. Use accepted-task cost as the decision metric. Astra must reduce failures or human work enough to justify its higher unit rate.

Use the maintained OpenAI pricing page for current structured rates, then model your real token mix in the AI token calculator. Teams comparing frontier alternatives can also review Anthropic pricing and the OpenAI API pricing guide.

Teams that need a lower-cost coding control while Astra access expands can benchmark the Z.ai coding plan on the same accepted-task rubric.

Affiliate disclosure: we may earn a commission from the sponsored link above. It does not affect this analysis.

Bottom line

GPT-6 Astra is now priceable: $10 input and $50 output per million short-context tokens, with a 2x input/cache and 1.5x output tier above 272K input. The model is expensive by OpenAI’s current ladder, so its business case depends on fewer failed runs and less human rework.

ARC Prize’s 62.7% Standard and 99.9% Provider Adapter results show why harness design can dominate both score and cost. Buyers should reproduce that lesson on their own workflows instead of treating either percentage as a direct cost forecast.

Sources: OpenAI’s GPT-6 Astra model page, API pricing, Astra developer guide, Path to Astra, and ARC Prize’s GPT-6 Astra on ARC-AGI-3. Pricing and availability checked September 3, 2026.