GPT-6 Astra: documented specifications, pricing, and evaluation checklist

Documented model specifications

OpenAI lists gpt-6-astra with a 1,050,000-token context window and up to 128,000 output tokens. The model supports text and image inputs and text outputs. Its supported reasoning effort settings are low, medium, high, xhigh, and max. These are documented limits, not guarantees of accurate retrieval or successful task completion. Source: OpenAI model documentation.

Pricing and a worked estimate

The model page lists standard text pricing per million tokens as $10 input, $1 cached input, and $50 output. Prompts above 272,000 input tokens use twice the input and cache rates and 1.5 times the output rate for the full request. Cache writes are priced separately. Tool charges and service tiers can change the total. Source: pricing section of the model page.

Illustrative calculation, not a measured bill: 10,000 uncached input tokens and 2,000 billable output tokens cost $0.10 + $0.10 = $0.20 at the listed standard rates, before tools or other charges. Check actual usage records, including billable reasoning tokens, when comparing costs.

How to evaluate it for your workflow

Start with a small set of tasks from your own work. Write the expected outcome before running the model. For coding, require the relevant tests to pass. For document extraction, compare each field with a checked answer. For computer use, record both task completion and unwanted actions.

1. Record the model ID, date, reasoning setting, prompt, tool permissions, and environment.
2. Use identical inputs and success criteria for each model. Repeat tasks to expose variable results.
3. Save outputs, failures, elapsed time, token usage, and total cost. Report the number of attempts.
4. Review incorrect outputs manually. A high aggregate score can hide failures that matter for your use case.
5. Publish the task set and evaluation procedure before calling the results reproducible.

What this guide does not establish

We have not reproduced OSWorld, ExploitBench, or long-context retrieval results for Astra. We do not infer a model-specific cybersecurity classification from a general safety page. We also do not infer internal neural architecture from available product features.

Before deployment, check account access, data handling requirements, tool permissions, and the current documentation. Run an evaluation that reflects the consequences of errors in your application.

Read our evidence and corrections standards · Report a correction