GPT-6 Astra Pricing: The 272K-Token Cost Cliff

GPT-6 Astra is OpenAI's flagship model, released in early September 2026 with a context window of 1,050,000 tokens. GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens on the standard tier. Any prompt over 272,000 input tokens reprices the entire request to $20 input and $75 output, per OpenAI's GPT-6 Astra model page. The headline rate is not the number that decides your bill. For any agent or retrieval workload that accumulates context, the 272K threshold is where the cost math changes.

This post gives the full rate card, a worked cost for a real large request, and a side-by-side with Claude Opus 5, Claude Sonnet 5, and Gemini 3.1 Pro. If you are choosing a model for long-context work, the comparison table below is the part to keep. Every rate here was fact-checked against each vendor's published pricing page on the day of publication.

What does GPT-6 Astra cost per token?

GPT-6 Astra has two pricing tiers, split at 272,000 input tokens. The standard tier covers any request with 272,000 or fewer input tokens. The long-context tier covers any request above that. Long-context pricing is the term OpenAI uses for the higher rate, and it reprices the full request rather than only the tokens past the threshold, per CloudZero's pricing breakdown.

Rate (per 1M tokens)	Standard tier (272K input or less)	Long-context tier (over 272K input)
Input	$10.00	$20.00
Cached input	$1.00	$2.00
Cache writes	$12.50	$25.00
Output	$50.00	$75.00

Source: OpenAI's GPT-6 Astra model page and CloudZero's pricing breakdown.

Service tiers then multiply those rates. Batch and Flex processing run at 50 percent of the applicable rate. Fast mode runs at 2 times the applicable rate, and Fast mode is not available with EU data residency, per the OpenRouter model listing. A context window is the maximum number of tokens a model can process in one request, and for GPT-6 Astra that limit is 1,050,000 tokens.

How much does the 272K cliff actually add?

Crossing 272,000 input tokens by a single token nearly doubles the cost of that request. Because the long-context tier reprices the whole request, the jump is a step, not a slope.

The table below prices the same request just under and just over the threshold. Each request has a 20,000-token output and no caching.

Request size	Tier applied	Input cost	Output cost	Total
272,000 input tokens	Standard	$2.72	$1.00	$3.72
273,000 input tokens	Long-context	$5.46	$1.50	$6.96

Source: calculated from GPT-6 Astra published rates.

One extra thousand tokens raises the bill by 87 percent. For an agent that grows its context turn by turn, that transition is easy to hit without noticing. Consider a support agent that appends each ticket, reply, and knowledge-base article to the prompt. It can pass 272,000 tokens mid-conversation, and every later turn then bills at the higher rate. The fix in practice is to cap retrieved context below 272,000 tokens, or to summarize and truncate history before the request is sent rather than after the cost lands.

GPT-6 Astra versus Claude and Gemini on a large request

For a large prompt, Claude Opus 5 and Claude Sonnet 5 are cheaper than GPT-6 Astra, and the gap widens once GPT-6 Astra enters its long-context tier. Claude Opus 5 is Anthropic's flagship model, priced at $5 per million input tokens and $25 per million output tokens with no context-size surcharge, per Anthropic's pricing page. Claude Sonnet 5 is Anthropic's mid-tier model at $2 input and $10 output, after the September 2026 price increase was cancelled.

The table prices one request of 500,000 input tokens and 20,000 output tokens, uncached, on each model's standard published rates. Capability is not equal across these models, so read the table as a cost floor, not a ranking.

Model	Input rate	Output rate	Request cost (500K in, 20K out)
Claude Sonnet 5	$2.00	$10.00	$1.20
Google Gemini 3.1 Pro	$4.00 (over 200K)	$18.00 (over 200K)	$2.36
Claude Opus 5	$5.00	$25.00	$3.00
GPT-6 Astra	$20.00 (over 272K)	$75.00 (over 272K)	$11.50

Sources: OpenAI, Anthropic, and Google published API rates, cross-checked against the llm-stats model profile.

At this request size GPT-6 Astra costs about 3.8 times Claude Opus 5 and nearly 10 times Claude Sonnet 5. If your workload is long-context summarization or retrieval rather than frontier reasoning, that difference is large enough to test a cheaper model first. Our breakdown in "Claude Opus 5 returns 400 on temperature. What to do" covers a migration gotcha worth knowing before you switch.

Is the context cliff unique to OpenAI?

No. Google uses the same structure on Gemini 3.1 Pro, with the threshold set lower at 200,000 tokens. Below 200,000 input tokens, Gemini 3.1 Pro costs $2 input and $12 output per million. Above it, the rate rises to $4 input and $18 output, per Google's Gemini API pricing.

Anthropic is the outlier. Claude Opus 5 and Claude Sonnet 5 charge one flat rate regardless of prompt size, per Anthropic's pricing page. That makes Claude pricing easier to forecast for workloads with variable context length, such as chat assistants or coding agents, because there is no threshold to model. GPT-5.6 also matters here as a cheaper OpenAI option, and our post "Who can actually run GPT-5.6-Cyber, and what it costs" covers where that model fits.

Which GPT-6 Astra tier should you budget for?

Budget for the long-context tier if your typical request carries more than about 200,000 input tokens, because you will cross 272,000 often enough that the higher rate is your real average. Budget for the standard tier only if you can hard-cap context below the threshold and enforce that cap in code.

For most retrieval and agent workloads, three settings control the bill. First, cap retrieved context to keep requests on the standard tier where possible. Second, use prompt caching for any stable system prompt or document set, since cached input at $1 per million is a tenth of fresh input. Third, route batch-safe work through the Batch tier at half price. The same time-of-day and tier logic shows up elsewhere in the market, as covered in "DeepSeek V4 pricing now depends on what time you run it".