Architecture cost review
GPT-6 Astra Pricing: What a Real Workload Costs
Astra's headline rate is only half the story. Long prompts cross a 272,000-token boundary, output is expensive, and batch or caching changes the answer materially. Here is the bill before the benchmark debate begins.
Direct answer
GPT-6 Astra is a premium model, not a drop-in economy upgrade. Below 272K input tokens it costs $10 per million input, $1 per million cached input and $50 per million output. At 272K input or more, input doubles to $20 and output rises to $75. Use it where a better result can avoid a retry, a review cycle or a failed task.
Decision summary
| Decision area | What matters |
|---|---|
| Standard, under 272K input | $10/M input · $1/M cached input · $12.50/M cache write · $50/M output |
| Long context, 272K input or more | $20/M input · $2/M cached input · $25/M cache write · $75/M output |
| Batch | 50% of the applicable standard token rates |
| Web search | $10 per 1,000 calls, plus model-token charges |
The short version
Astra is priced for difficult work. A request with 1 million fresh input tokens and 200,000 output tokens costs $20 below the long-context threshold. For comparison, the same token shape costs $8 on GPT-5.6 Sol. Astra must therefore save more than $12 in retries, human review or downstream failure to win that simple comparison.
The threshold matters more than it first appears. Once a request contains at least 272,000 input tokens, OpenAI applies the long-context rates to that request: $20 per million input and $75 per million output. Do not estimate a large-repository or document workflow from the headline $10/$50 rates.
Worked cost examples
At standard rates, 100,000 fresh input tokens and 20,000 output tokens cost $2: $1 for input and $1 for output. Ten identical jobs cost $20 before tools or retries.
If those jobs reuse 80,000 cached input tokens and add 20,000 fresh tokens each, the token bill falls sharply after the cache is populated. But cache writes carry their own $12.50-per-million rate, so short-lived or constantly changing prompts may not recover the write cost.
- 1M input + 200K output, standard: $20
- 1M input + 200K output, long-context band: $35
- 1M cached input + 200K output, standard cache-read case: $11
- The standard batch equivalent of the $20 workload: $10
Why output control matters
Output is five times the standard input rate. An agent that narrates every step, dumps large files or retries verbose answers can spend more on completion tokens than on the context that started the task.
Use maximum-output limits and ask for compact machine-readable results where that fits the job. For coding and research agents, count tool traces and failed branches as part of the workload rather than pricing only the final answer.
When Astra can still be cheaper
A higher per-token model can be cheaper per accepted result. If Sol needs three attempts and Astra completes the same task once, Astra may win even at 2.5 times the worked-example token bill. That is a measurement question, not a launch-day assumption.
Start with a labelled sample of difficult tasks. Keep easy traffic on a cheaper model, send only the hard tail to Astra, and compare cost per accepted task rather than cost per request.
Context window is capacity, not a reason to fill it
OpenAI documents a 1.05-million-token context window, with up to 922,000 input and 128,000 output tokens. That capacity is useful for unusually large source packs, but it can also turn weak retrieval into a very expensive prompt.
Test retrieval, summarisation and stable-prefix caching before sending an entire corpus. The question is not whether Astra can read it; it is whether the extra context changes the accepted answer enough to pay for itself.
Forecast Astra as a variable-cost procurement item
Astra procurement should begin with workload forecasting rather than a model-name decision. The official short-context rate is $10 per million input tokens and $50 per million output tokens; requests above 272,000 input tokens use the higher long-context rates for the full request. Cached input, cache writes, batch processing and web search create separate cost lines. Forecast by workload class—interactive chat, document analysis, coding agents and asynchronous jobs—instead of multiplying one average token estimate across the organisation.
A useful forecast has at least four fields for every workflow: expected input tokens, expected output tokens, cache-read percentage and tool-call volume. A short-context request containing one million input tokens and 200,000 output tokens costs $20 before tools. The same volumes at long-context rates cost $35. A batch version of the short-context example costs $10, but only when delayed processing is acceptable. Capacity and rate limits also belong in planning; the monthly estimate alone does not tell an engineering team whether peak traffic will run.
Subscription access needs a separate forecast. ChatGPT Work and Codex share a plan allowance, and OpenAI says Astra can consume that allowance faster than GPT-5.6 Sol. The published Pro allowances are not unlimited. Keep seat cost, API consumption and purchased credits in separate budget lines, set a monthly ceiling, and require token and tool-usage reporting before expanding access. A subscription is not a fixed-cost substitute for an unbounded API workload.
Price failure by successful outcome, not by request
The useful procurement measure is cost per accepted result: total model and review spend divided by outputs that pass the organisation's quality check. Retries, human corrections, abandoned runs and tool-side failures belong in the numerator. This stops a low request price from looking efficient when a cheaper model needs repeated intervention.
Separate failure into three levels. A recoverable failure is an incorrect draft that a reviewer fixes. An operational failure is a broken code change, wasted long-running agent session or tool action that must be reversed. A consequential failure creates privacy, security, financial or compliance exposure. Astra's safety card reports stronger alignment and prompt-injection robustness than Sol in OpenAI's tests, but also lower adversarial monitorability. High-impact workflows should price approvals, sandboxing and review into deployment rather than assume capability removes those controls.
Turn the comparison into a break-even test. If one Astra task costs $20 and removes 40 minutes of work valued at $30 per hour, the labour value is also $20; at $100 per hour, it is about $66.67. That does not prove Astra is economical because the task may still need review. It defines the pilot data: accepted-result rate, retries, reviewer minutes, tool charges and time to completion. A premium is justified when those measures improve materially, not simply when the model is newer.
Key takeaways
- →GPT-6 Astra is a premium model, not a drop-in economy upgrade. Below 272K input tokens it costs $10 per million input, $1 per million cached input and $50 per million output. At 272K input or more, input doubles to $20 and output rises to $75. Use it where a better result can avoid a retry, a review cycle or a failed task.
- →Route Astra only to tasks with a measured quality or completion-rate advantage, cache stable context, cap output and use batch for delay-tolerant work.
- →OpenAI can change prices, plan limits and product access. Cached-input savings require repeatable prefix content; batch trades latency for a 50% discount.
How this page was prepared
This launch cluster separates OpenAI's product and benchmark claims from independent observations. Prices and limits come from official documentation checked on 14 September 2026. Comparisons use explicit token assumptions and do not claim first-hand testing.
Frequently asked questions
How much does GPT-6 Astra cost per million tokens?
Below 272K input, $10 per million input, $1 per million cached input and $50 per million output. At 272K input or more, the rates are $20, $2 and $75 respectively.
Does GPT-6 Astra have batch pricing?
Yes. OpenAI lists Batch API rates at 50% of the applicable standard token rates for delay-tolerant work.
How much is 1M input and 200K output?
$20 at the short-context rates: $10 input plus $10 output. In the long-context band the same volumes cost $35.
Is cached input always cheaper?
Cache reads are much cheaper, but creating the cache costs more than a normal input token. Caching pays when a stable prefix is reused enough times.