GPT Astra pricing: what GPT-6 Astra costs to run
Updated October 22, 2026 · first published September 23, 2026
GPT Astra refers here to GPT-6 Astra, OpenAI's model for demanding reasoning, coding and research work. Its standard API rates are $10 per million ordinary input tokens, $1 per million cached-input tokens and $50 per million output tokens. Cache writes, processing mode and long context change the bill. Rates below were checked against official documentation on 22 October 2026.
Is GPT Astra a ChatGPT plan or an API model?
gpt-6-astra is the API model identifier. A model name identifies the system handling a request; a ChatGPT plan identifies a subscription with its own access and usage rules. Availability in a ChatGPT workspace does not establish your API access or pricing.
OpenAI documents separate billing systems for ChatGPT and the API platform. Budget API calls separately from subscriptions, and check access in the account you will actually use. This guide calculates API token costs; it does not promise a particular ChatGPT plan entitlement.
What are GPT-6 Astra's API prices?
All amounts are USD per one million tokens at standard rates, before any applicable regional uplift, tax, contract discount or tool charge.
| Token category | Standard rate |
|---|---|
| Ordinary input | $10.00 |
| Cached input read | $1.00 |
| 30-minute cache write | $12.50 |
| 12-hour cache write | $25.00 |
| Output | $50.00 |
Sources: OpenAI API pricing and the Astra model page. The model page documents a 1,050,000-token context window, up to 128,000 output tokens, and reasoning efforts low, medium, high, xhigh and max; it does not list a none setting.
How do processing mode and long context change the rates?
| Condition | Effect |
|---|---|
| Batch or Flex | 50% of standard rates; ordinary input/output become $5/$25 per million |
| Fast mode | 2× applicable rates; short-context ordinary input/output become $20/$100 |
| More than 272,000 input tokens | 2× input and cache rates, 1.5× output rates for the full request |
The long-context threshold applies to the full request, not just the tokens above 272,000. For standard processing above that threshold, ordinary input is $20, cached reads $2, 30-minute writes $25, 12-hour writes $50 and output $75 per million. Record the applicable context band and processing mode when estimating a bill.
The pricing page also lists Ultrafast separately. If you enable it, use its published rates instead of treating it as Fast mode. Keep the returned service tier with the usage record so an unexpected tier change is visible.
How should you calculate cache costs without double counting?
In OpenAI's prompt-caching documentation, input categories are mutually exclusive: a token uses the ordinary-input, cache-read or cache-write rate. A cache-write rate replaces the ordinary-input rate for those tokens; it is not an extra charge on top of it.
ordinary_input = total_input - cache_reads - writes_30m - writes_12h
cost_usd = (
ordinary_input * 10
+ cache_reads * 1
+ writes_30m * 12.5
+ writes_12h * 25
+ billed_output * 50
) / 1_000_000This formula uses standard short-context rates. Substitute the appropriate rates for other modes or long context. Use the provider's billed output count, including billable reasoning where reported; do not add a reasoning subset again if it is already included.
What does one Astra agent step cost?
Illustrative workload: one standard request with 200,000 total input tokens and 20,000 billed output tokens. Keep 20,000 input tokens ordinary; vary the treatment of the remaining 180,000-token prefix.
| Prefix treatment | Calculation | Request cost |
|---|---|---|
| Warm cache read | $0.20 ordinary + $0.18 read + $1 output | $1.38 |
| New 30-minute write | $0.20 ordinary + $2.25 write + $1 output | $3.45 |
| New 12-hour write | $0.20 ordinary + $4.50 write + $1 output | $5.70 |
These are calculated scenarios, not measured average task costs. They exclude tools, infrastructure and additional attempts. A workflow with five identical warm-cache steps costs $6.90 in model tokens, rather than $1.38 for the whole task.
The context threshold matters too. With 90% cached input, no writes and the same 20,000 output tokens, 270,000 input tokens cost $1.513. At 300,000 input tokens, long-context rates apply to the full request and the cost is $2.64.
When should you route a task to Astra?
Measure completion quality and total task cost on a fixed sample before choosing a model or raising its reasoning effort. Sum every step, failed attempt, tool fee and retry. Divide by the tasks that pass the same acceptance check. A cheaper individual request can still produce a more expensive completed task.
Our cross-model cost-per-task analysis preserves a dated benchmark comparison; it is not a current guarantee for your workload. Re-run your own evaluation when models, prices or defaults change. The practical agent workflow guide shows where to enforce budgets and collect the records needed for that evaluation.
What should your Astra cost log contain?
- Workflow, step, attempt and provider request identifiers.
- Requested and returned model and service tier, reasoning effort, and applicable context band.
- Total input, cache reads, writes split by retention, and billed output.
- Estimated cost, currency, price-table version, outcome and any unresolved usage.
Reconcile estimates against provider usage and invoices. If a timeout leaves billing unknown, mark the amount pending instead of recording zero. A budget should reserve the maximum permitted next attempt before dispatch, including a possible higher context band.
Sources and update scope
Official API pricing, model specifications, caching rules and ChatGPT/API billing guidance linked above were checked on 22 October 2026. This update expands the English guide. Existing translations retain their earlier text; consult the dated official sources before using any historical comparison for purchasing.
Frequently asked questions
What is GPT Astra?
This guide uses GPT Astra to mean GPT-6 Astra, API model gpt-6-astra. ChatGPT subscriptions have separate access rules and billing.
How much does GPT-6 Astra cost?
At standard short-context API rates, ordinary input is $10, cached reads $1 and output $50 per million tokens. Cache writes cost $12.50 for 30 minutes or $25 for 12 hours per million tokens.
Is an Astra cache write charged on top of input?
No. OpenAI bills each input token at its applicable ordinary-input, cache-read or cache-write rate. Use disjoint token categories to avoid double counting.
What happens above 272,000 input tokens?
Astra applies 2× input and cache rates and 1.5× output rates to the full request, not just the excess tokens.
Does a ChatGPT subscription cover API calls?
API usage is billed separately from ChatGPT subscriptions. Check API billing and model access in the organization making the requests.
Related
- GPT-6 pricing: Astra, Sol and Luna
- GPT-6 Sol cost model
- GPT-6 Luna cost model
- Claude Fable 5.1 cost model
- Opus 5.5 cost by effort level
- Practical agent workflow architecture
Want this applied to your stack? Bring the provider bills, gateway logs, and top workflows; we will map the cost drivers and savings path. Book a free audit →