GPT-6 Astra: Specs, Pricing, Context Window and Availability Facts

GPT-6 Astra is a proprietary large language model from OpenAI, released on 3 September 2026, which OpenAI's models page describes as "our most capable model for the most demanding work". This profile collects the published facts about it in one place: API identifier, context window, output limit, knowledge cutoff, prices for every service tier, tools, cloud availability, safety ratings from its system card and deprecation status. It is written for developers, architects, buyers and governance teams who need checked numbers rather than launch coverage. Every fact comes from OpenAI's own documentation and system card, listed under Standards and references. Where OpenAI publishes nothing on a point, the page says so.

Last checked: 2026-11-15.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses OpenAI's public documentation as of November 2026. Model details change, so confirm against OpenAI's current documentation.

Key takeaways

• Released 3 September 2026. The OpenAI API changelog records GPT-6 Astra (API ID gpt-6-astra) on that date, and it is listed first among OpenAI's flagship models.
• 1.05M tokens in, 128K out. The model page lists a 1,050,000-token context window, 128,000 maximum output tokens and a knowledge cutoff of 30 April 2026.
• $10 input and $50 output per million tokens. Standard pricing, with cached input at $1.00, Batch and Flex at 50% of Standard, Fast mode at 2x and an Ultrafast tier at $60 and $300.
• Critical for cybersecurity under the Preparedness Framework. OpenAI's system card calls it the company's first model to reach the Critical level of cybersecurity capability, and treats it as High in the biological and chemical category.
• No deprecation date. OpenAI's deprecations page lists no notice or shutdown date for gpt-6-astra as of 15 November 2026.

What is GPT-6 Astra?

GPT-6 Astra is the top tier of OpenAI's GPT-6 generation. OpenAI's 2026 naming uses four tiers: Astra, Sol, Terra and Luna. The models page positions Astra for "the most demanding work", and the model page lists complex reasoning, coding, computer use, research and document creation as intended uses. It is a reasoning model: it produces reasoning tokens before its answer, and the depth is set with the reasoning effort parameter.

The API changelog entry for 3 September 2026 says OpenAI "Released GPT-6 Astra, our most capable model, built for the hardest end-to-end work", and on the same date added new controls for long-running work in the Responses API. Astra accepts text and image input and returns text. It is closed weight and proprietary, and OpenAI does not state a parameter count. For the full OpenAI line-up, see OpenAI Models.

Key facts at a glance

As of 2026-11-15. Source: OpenAI model page for gpt-6-astra, API changelog and system card.

Release and version history

As of 2026-11-15. Dates from the OpenAI API changelog and the system card change log.

The model page lists one snapshot, gpt-6-astra, with no dated variants. OpenAI's pricing page also lists gpt-6-astra-law, a legal-domain variant under specialised models, priced at the same Standard rates with promotional pricing stated through at least 31 December 2026. OpenAI does not publish a separate model page for that variant on the pages checked, so this profile covers gpt-6-astra only.

How much does GPT-6 Astra cost?

As of 2026-11-15. USD per million text tokens. Source: OpenAI pricing page and the gpt-6-astra model page.

• Long prompts: prompts over 272K input tokens are charged at 2x input and cache rates and 1.5x output, for the whole request (model page).
• Cache writes: a 30-minute cache write costs 1.25x the uncached input rate. When migrating from GPT-5.5 or earlier, OpenAI's GPT-6 guide says to replace prompt_cache_retention with prompt_cache_options.ttl.
• Regional processing: the pricing page states a 10% uplift on data residency endpoints for models released on or after 5 March 2026, and the same uplift on FedRAMP endpoints.
• Tool fees: web search is $10.00 per 1,000 calls plus search content tokens at model rates; other tools may carry per-call fees listed on the pricing page.

Worked example at Standard rates: a request with 100,000 input tokens and 10,000 output tokens costs $1.00 for input plus $0.50 for output, $1.50 in total, before any tool fees.

Context window and output limits

The context window is 1,050,000 tokens and the maximum output is 128,000 tokens per response. Because requests above 272K input tokens are billed at higher rates, the effective price per token changes at that threshold: plan retrieval and context packing with the 272K boundary in mind, not only the 1.05M ceiling.

Modalities and tools

Astra accepts text and images and returns text. Audio and video are not supported as input or output on this model; OpenAI offers separate realtime, transcription and image models for those.

• Features: streaming, function calling and Structured Outputs are supported; fine-tuning is not.
• Responses API tools: web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search. OpenAI's GPT-6 guide states that Astra, Sol and Terra require the Responses API for tool calling.
• Reasoning effort: low, medium, high, xhigh and max, set with reasoning.effort (Responses) or reasoning_effort (Chat Completions). Astra does not support none; the guide recommends low instead.
• Unsupported sampling parameters: temperature, top_p and top_logprobs must be removed, and logprobs in Chat Completions.
• Long-running work controls: asynchronous tool calling (async: true, results returned later by call_id), mid-turn steering over WebSocket, and configuration_update items that change reasoning effort mid-conversation without rewriting the prompt prefix.
• Prompting notes from OpenAI: the guide says Astra asks clarifying questions more often than earlier models, favours formatted responses and may write broader tests than small coding changes need, and it recommends auditing skill and instruction files such as AGENTS.md.

Hosted shell, computer use and MCP give the model reach into systems. Treat each connected tool as a credentialed integration. See the Securing OpenAI Agents SDK and Responses API Guide.

Where is GPT-6 Astra available?

As of 2026-11-15. Source: gpt-6-astra model page.

OpenAI's GPT-6 guide states that Astra's Ultrafast mode supports US data residency and global processing, and that Astra's Fast mode has no latency SLA. Region lists for other data residency options are in OpenAI's separate data residency guide.

Access restrictions

Astra is not a restricted-access model in the API: it is listed as a flagship model with published rate limits for the Build, Launch and Grow tiers, and it is not supported on the Free tier. OpenAI's restricted cyber models sit separately under OpenAI Daybreak trusted access (gpt-6-cyber, gpt-5.6-cyber). The system card states that Astra had a restricted initial deployment period, that users flagged as potentially high risk meet a more conservative refusal boundary covering a wider range of dual-use risks, and that trusted access applies to cyber work.

Benchmarks (provider-reported, dated)

As of 2026-11-15. All figures are provider-reported, from the GPT-6 Astra system card (published 3 September 2026) or, where marked, the GPT-6.1 Sol system card addendum (29 September 2026). Independent leaderboard results are not included because none were checked against their own published pages for this profile.

For attack success rates and flag counts, lower is better; for the other scores, higher is better. Cyber capability scores are reported to explain the Preparedness rating, not as product features.

Model card and system card

OpenAI published the GPT-6 Astra system card on its Deployment Safety Hub on 3 September 2026, with a change log running to 30 October 2026, and a shorter safety overview on openai.com. What the card states:

• Preparedness ratings: Astra is "our first model to reach the Critical level of cybersecurity capability under our Preparedness Framework"; OpenAI is "treating Astra as High capability in the biological and chemical domain"; and Astra "does not reach our High threshold" for AI self-improvement.
• Safeguards in development: stricter isolation, checkpoint encryption, monitoring of full trajectories including chains of thought during internal use, and a blocking alignment evaluation before internal use.
• Safeguards in deployment: misalignment monitoring on all tool-using inference in external deployment, robustness training against jailbreaks, regression testing against earlier jailbreaks, automated red teaming and age-specific protections for users under 18.
• Monitorability: the card says monitorability decreased relative to GPT-5.6 Sol; under adversarial prompting Astra could underperform strategically without detection and sometimes evaded internal monitors on certain sabotage tasks, and OpenAI reports no evidence of steganographic reasoning.
• Agentic behaviour: in realistic work environments the overall misaligned outcome rate was 3.4% for Astra versus 18.8% for GPT-5.6 Sol without a confirmation policy, and 3.0% versus 8.0% with one.

For how labs structure these ratings, see the Frontier AI Model Risk Frameworks Guide.

Data handling

OpenAI's API data controls page applies to Astra as to other API models. There is no Astra-specific data policy on the pages checked, apart from the misalignment monitoring described in the GPT-6 guide and system card.

• Training use: data sent to the OpenAI API is not used to train or improve OpenAI models unless the customer explicitly opts in.
• Abuse monitoring logs: kept for up to 30 days by default, unless longer retention is legally required or needed to protect the services.
• Zero Data Retention and Modified Abuse Monitoring: available only with prior OpenAI approval; ZDR forces store to false on Responses and Chat Completions. OpenAI states it may make certain models ineligible for these controls for specific customers with notice.
• Application state: Responses API state is kept for at least 30 days by default when store is true; prompt cache data for up to 24 hours (12-hour cache writes are priced separately).
• Misalignment monitoring: the GPT-6 guide says systems review Astra interactions asynchronously and raise alerts when needed.

API keys and service accounts that call Astra are non-human identities with direct spend and data exposure. See the OpenAI API Keys Security Guide and OpenAI Platform Organisation and Project Security Guide.

Deprecation status

As of 15 November 2026, OpenAI's deprecations page lists no notice date, shutdown date or replacement for gpt-6-astra. Astra does not appear as the replacement for any retiring model; the GPT-5 and o3 retirements on 11 December 2026 point to gpt-6.1-sol instead. OpenAI's practice, shown on the same page, is to publish a notice date, a shutdown date and a recommended replacement, so monitor that page and the changelog rather than relying on this snapshot.

How GPT-6 Astra differs from other models in its family

As of 2026-11-15. Source: OpenAI models page and per-model pages. Models listed by release date.

All three share the 1,050,000-token context and 128,000-token output limit. The documented differences are price (Astra input and output cost five times GPT-6.1 Sol at Standard rates), cyber evaluation scores in the GPT-6.1 Sol system card, and Astra's misalignment monitoring and Ultrafast data residency options. GPT-6.1 Terra and GPT-6 Luna are lower-cost tiers. See GPT-6 Astra vs Sol vs Luna vs Terra and the GPT-6.1 Sol profile.

Practitioner checklist

• Pin gpt-6-astra explicitly in code and record the date you adopted it, since only one snapshot exists.
• Replace any none or minimal reasoning effort with low, and remove temperature, top_p and logprobs parameters before migrating.
• Move tool calling to the Responses API; Astra requires it.
• Model cost with the 272K input threshold and reasoning tokens included, and set per-project budgets before enabling Fast or Ultrafast.
• Test the same workload on GPT-6.1 Sol at matched reasoning effort; use Astra where the measured difference justifies the price.
• Scope hosted shell, computer use and MCP tools to least privilege, with human approval for destructive actions.
• Confirm whether ZDR or Modified Abuse Monitoring is approved for your organisation before sending regulated data.
• Store the API keys used for Astra in a secrets manager, separate projects by environment, and rotate keys on staff changes.
• Watch the deprecations page and changelog monthly for changes to gpt-6-astra.

Standards and references

• OpenAI: GPT-6 Astra model page (accessed 2026-11-15)
• OpenAI Deployment Safety Hub: GPT-6 Astra System Card (accessed 2026-11-15)
• OpenAI: GPT-6 Astra safety overview (accessed 2026-11-15)
• OpenAI: API pricing (accessed 2026-11-15)
• OpenAI: API changelog (accessed 2026-11-15)
• OpenAI: Using GPT-6 models guide (accessed 2026-11-15)
• OpenAI: Deprecations (accessed 2026-11-15)
• OpenAI: Data controls in the OpenAI platform (accessed 2026-11-15)

Related NHI Mgmt Group resources: OpenAI Models, GPT-6 Astra vs Sol vs Luna vs Terra, GPT-6.1 Sol, ChatGPT Enterprise Security Guide, How to audit OpenAI and Anthropic organisation API keys.