Overview
GPT-6 Astra is OpenAI's September 3, 2026 flagship model for complex professional work and autonomous agents. It is rolling out across ChatGPT, the OpenAI API, and Amazon Bedrock, with a focus on computer use, coding, science, long-context reasoning, and workflows that call tools over extended periods.
This is a substantial model upgrade rather than a routine endpoint refresh. Astra introduces asynchronous tool calling, lets developers steer an agent while it is still working, and supports changing reasoning effort between responses via configuration_update without rewriting the cached prompt prefix. Those capabilities make it relevant to teams comparing AI agent platforms, but they also create migration work and a significantly higher standard API price than GPT-5.6 Sol.
The model is not a new standalone ToolWorthy product. It is the current model release for ChatGPT, while the previous GPT-5.6 release remains the better reference for teams deciding whether Astra's gains justify the cost and compatibility changes.
What's New
Asynchronous tool calling for longer-running agents
GPT-6 Astra can initiate new tool calls without waiting for every earlier call to finish. That lets an agent continue useful work while slower operations are still running, which can shorten the critical path for research, coding, browser automation, and multi-system workflows.
This changes orchestration assumptions. Applications should be prepared for concurrent tool activity, out-of-order completion, cancellation, and partial failure. A sequential tool loop can still work, but it will not capture the main latency benefit of the new model.
Mid-turn steering and dynamic reasoning effort
Developers can send steering instructions over a WebSocket connection while Astra is working. The Responses API also supports configuration_update, allowing reasoning effort to change between responses in standard single-agent requests while preserving the cached prompt prefix.
Together, these controls support a practical pattern: begin with lower effort, inspect intermediate progress, then increase effort or redirect the agent only when the task warrants it. This can reduce wasted reasoning on straightforward work while retaining a path for difficult branches.
Stronger computer use and professional work
OpenAI positions Astra for browser and desktop interaction, software engineering, documents, spreadsheets, presentations, scientific work, and security operations. The official model page lists computer use, hosted shell, code interpreter, web search, file search, image generation, MCP, tool search, apply patch, and skills among supported Responses API tools.
The model accepts text and image input and produces text output. Audio and video are not supported model inputs, so multimodal applications that depend on those formats still need transcription, frame extraction, or a separate model path.
Same context and output limits
The API model retains the 1,050,000-token context window and 128,000-token maximum output supported by GPT-5.6 Sol, with a newer April 30, 2026 knowledge cutoff. That is useful for large repositories, long evidence sets, and persistent agent state, but prompts above 272,000 input tokens use higher long-context pricing for the entire request.
Higher cybersecurity capability with tighter controls
OpenAI classifies GPT-6 Astra as its first model to reach the Critical cybersecurity capability threshold. The company reports stronger exploit discovery, incident response, and security engineering performance, alongside stricter safeguards, monitoring, and refusal behavior for advanced offensive tasks.
The safety tradeoff is important: OpenAI also reports reduced written-reasoning monitorability compared with GPT-5.6 Sol. Teams should not treat hidden or abbreviated reasoning as an audit trail; use observable tool logs, scoped credentials, approval gates, and independent output checks.
Compared With Previous Version
GPT-6 Astra succeeds GPT-5.6 Sol as OpenAI's flagship path for the hardest work. The key difference is not only benchmark performance. Astra is designed for agents that operate for longer, interact with multiple tools, and accept steering while a task is in progress.
| Decision Area | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|
| Primary role | Flagship agentic and professional model | Previous flagship reasoning model |
| API model ID | gpt-6-astra |
GPT-5.6 Sol model identifier used by the existing integration |
| Context / max output | 1.05M / 128K tokens | 1.05M / 128K tokens |
| Tool orchestration | Asynchronous calls and mid-turn steering | Programmatic tool calling and beta multi-agent support |
| Reasoning control | Low through max, including mid-turn updates | Existing configurable reasoning effort |
| Standard API price | $10 input / $50 output per 1M tokens | $4 input / $20 output per 1M tokens (current promotion, available at least through November 21, 2026) |
| Best fit | Long-running agents, computer use, complex coding and science | Cost-sensitive flagship workloads already tuned for GPT-5.6 |
The price increase is material: compared with GPT-5.6 Sol's current promotional $4 input / $20 output per 1M tokens, available at least through November 21, 2026, Astra costs 2.5 times as much per input and output token. Upgrade decisions should therefore use cost per successful task, including retries and human review, rather than benchmark scores or token price alone.
Performance Benchmarks
OpenAI's launch materials report broad gains over GPT-5.6 Sol. These are vendor-published results using OpenAI's stated evaluation setups, so they are useful for shortlisting but do not replace private evaluations.
| Benchmark | GPT-6 Astra | GPT-5.6 Sol | Decision Signal |
|---|---|---|---|
| OSWorld 2.0 | 72.6% | 65.7% | Computer-use task completion |
| Agents' Last Exam | 59.3% | 53.6% | Long-horizon agent work |
| ScreenSpot Pro | 92.7% | 76.9% | Visual interface grounding |
| AutomationBench | 41.4% | 18.1% | End-to-end software automation |
| Terminal-Bench 4.0 | 57.9% | 37.3% | Command-line software engineering |
| Terminal-Bench Science | 64.6% | 22.4% | Scientific tool use |
| MRCR, 512K-1M | 96.3% | 73.8% | Very-long-context retrieval |
| ExploitBench | 100% | 78.5% | Advanced cybersecurity capability |
OpenAI also says Astra completed OSWorld 2.0 tasks in roughly 40 minutes versus 75 minutes for GPT-5.6 Sol. Separately, OpenAI reports 1.9 times faster Mind2Web task completion when Astra is combined with an updated Codex harness. Test those latency claims against your tool stack because network calls, browser state, retries, and review gates can dominate end-to-end time.
Migration Guide
For an existing GPT-5.6 integration, start with a controlled model-routing change rather than moving all production traffic at once.
- Change the model to
gpt-6-astrain a staging or evaluation environment. - Keep the current reasoning level when migrating an already reasoning-heavy workflow. For workloads that used no or minimal reasoning, start Astra at
lowbecausenoneis not supported. - Move tool-calling applications to the Responses API. Chat Completions can use the model, but OpenAI says Astra tool calling requires Responses.
- Remove unsupported sampling controls:
temperature,top_p, andtop_logprobs. Also removelogprobsfrom Chat Completions and themessage.output_text.logprobsinclude from Responses requests. - Update orchestration for asynchronous tool calls before enabling that behavior. Add correlation IDs, idempotency, cancellation, timeouts, and handling for out-of-order results.
- Use
configuration_updateto change reasoning effort between responses while preserving the cached prompt prefix. - If migrating from GPT-5.5 or earlier, replace
prompt_cache_retentionwithprompt_cache_options.ttland use the documented30mvalue where appropriate. - Run private evaluations for completion rate, tool accuracy, latency, output-token use, safety behavior, and total cost before changing the default model.
Compatibility Notes
- API surface: Tool calling requires the Responses API even though the model is also available through Chat Completions.
- Reasoning: Supported levels are
low,medium,high,xhigh, andmax;noneis not supported. - Sampling parameters:
temperature,top_p, andtop_logprobsare unsupported and should be removed rather than silently retained. - Modalities: Text and image input are supported. Audio and video input are not supported, and output is text only.
- Fine-tuning: The official model page does not list fine-tuning support.
- Data residency: OpenAI's migration guide says EU data-residency traffic uses Standard processing; Fast and Priority processing are unavailable there.
- Fast mode: Fast processing has no latency service-level agreement and costs twice the applicable Standard rate.
- Long context: Requests above 272K input tokens are billed at twice the input and cache rate and 1.5 times the output rate for the full request.
Availability & Access
OpenAI is using a staged launch rather than instant universal access. The launch announcement described future access across Plus, Pro, Business, and Enterprise, but the current Help Center narrows Chat access: Astra is rolling out there as GPT-6 Pro for Pro $100, Pro $200, Business, and Enterprise and is not included with Plus in Chat. Enterprise access also depends on workspace model-access permissions.
GPT-6 Pro Chat usage is included within eligible plan allowances, with plan-specific limits. The current Help Center lists 50 messages per week for Pro $100, 200 per week for Pro $200, 15 per month for Business Standard, and 50 per week for Business Premium. Free, Go, and Plus users should not assume Chat access based on general ChatGPT availability; ChatGPT Work and Codex use separate allowances.
Developers can use the gpt-6-astra API model where enabled, and AWS customers can access Astra through Amazon Bedrock. Rollout timing, regional processing rules, rate limits, and account eligibility can differ, so check the live model catalog before scheduling a production migration.
Pricing & Plans
Standard OpenAI API pricing for GPT-6 Astra is:
| Item | Price per 1M tokens |
|---|---|
| Input | $10.00 |
| Cached input read | $1.00 |
| Cache write | $12.50 |
| Output | $50.00 |
Batch and Flex processing are listed at 50% of Standard rates. Fast processing can provide up to twice the speed at twice the applicable Standard price. Long-context requests above 272K input tokens incur the higher multipliers described in Compatibility Notes.
ChatGPT subscriptions and API usage are separate billing surfaces. A Plus or Pro subscription does not include API spend, and plan allowances do not translate directly into token quotas. For document or data-heavy workflows, compare Astra with specialized AI spreadsheet tools when deterministic editing and lower-cost repeatability matter more than general reasoning.
Safety & Limitations
Astra's Critical cybersecurity classification means it requires stricter operational controls, not less oversight. OpenAI says advanced cyber requests may be refused and describes additional safeguards and monitoring around high-risk use. Daybreak already exists for approved defenders; OpenAI says it plans to expand access and roll out less restrictive safeguards in the coming weeks.
The model can still hallucinate, misread interfaces, call the wrong tool, or produce plausible but incorrect code. Its April 2026 knowledge cutoff also means current facts require retrieval. For production agents, use least-privilege credentials, sandboxed execution, allowlisted actions, complete tool logging, human approval for irreversible operations, and independent verification of high-impact outputs.
Who Should Upgrade / Who Should Wait
Upgrade to GPT-6 Astra if:
- Your agents spend significant time waiting on sequential tools and can benefit from asynchronous execution.
- Computer use, repository-scale coding, scientific workflows, or very long context are core requirements.
- Mid-turn steering and dynamic reasoning controls can reduce retries or manual restarts.
- Private evaluations show that higher completion rates offset the increased token price.
Wait or keep GPT-5.6 Sol if:
- Your current workflow already meets quality targets and is sensitive to token cost.
- You depend on unsupported sampling parameters, Chat Completions tool calling, audio/video input, or fine-tuning.
- Your tool layer cannot yet handle concurrent calls, cancellation, and out-of-order completion safely.
- Enterprise administrators have not enabled access or your required region does not support the processing mode you need.
FAQ
What is the API model ID for GPT-6 Astra?
Use gpt-6-astra. Availability can depend on account, region, and rollout stage.
Is GPT-6 Astra available in ChatGPT?
Yes, OpenAI announced a staged rollout to ChatGPT Plus, Pro, Business, and Enterprise. Enterprise administrators must enable it, and Astra Pro is limited to Pro, Business, and Enterprise plans.
How much does GPT-6 Astra cost in the API?
Standard pricing is $10 per 1M input tokens, $1 per 1M cached-input reads, $12.50 per 1M cache-write tokens, and $50 per 1M output tokens. Long-context and Fast processing can increase those rates.
Does GPT-6 Astra support a one-million-token context window?
Yes. The official model page lists a 1,050,000-token context window and a 128,000-token maximum output. Requests above 272K input tokens use higher long-context pricing.
Is GPT-6 Astra backward compatible with GPT-5.6 tool workflows?
Not completely. Existing prompts are a useful starting point, but tool calling requires Responses, several sampling parameters are unsupported, and asynchronous calls require concurrency-safe orchestration.
Should every GPT-5.6 workload move to Astra?
No. Astra is most compelling when agent completion, computer use, coding, or long-context quality offsets its higher price. Stable, cost-sensitive workloads should remain on GPT-5.6 until private evaluations show a clear gain.
