Release and availability: GPT-6 Astra launched on September 3, 2026
OpenAI announced GPT-6 Astra on September 3, 2026. At launch it was available to a limited set of organizations, with broader access to paid ChatGPT plans and the API promised over the following days. Reuters and Microsoft independently confirmed the staged rollout. [1][7][8]
Technical specifications: 1.05M context and 128K output
| Specification | GPT-6 Astra |
|---|---|
| Context window | 1,050,000 tokens |
| Max output | 128,000 tokens |
| Knowledge cutoff | Apr 30, 2026 |
| Input | text + images |
| Output | text |
| Reasoning | low / medium / high / xhigh / max |
The API documentation lists a 1,050,000-token context window, 128,000 maximum output tokens, and an April 30, 2026 knowledge cutoff. The model accepts text and image input and produces text output. [2][11]
API pricing: $10 input, $50 output, and the important 272K threshold
| Specification | Price / MTok |
|---|---|
| Input | $10.00 |
| Cached input | $1.00 |
| Cache write | $12.50 |
| Output | $50.00 |
| >272K input | 2× input/cache; 1.5× output |
| Batch / Flex | 50% of Standard |
| Fast | 2× applicable rates |
OpenAI API Standard rates are $10/MTok input, $1/MTok cached input, $12.50/MTok cache writes, and $50/MTok output. Above 272K input tokens, the entire request is billed at 2× input/cache rates and 1.5× output. Batch and Flex are 50% of Standard rates; Fast is 2×. [2]
Reasoning levels: low, medium, high, xhigh, and max
`reasoning.effort` supports `low`, `medium`, `high`, `xhigh`, and `max`. Astra does not support `none`; when migrating from `none` or `minimal`, OpenAI recommends starting with `low` and testing the result. [2][3]
Migrating from GPT-5.6 is more than a model-ID swap
The model ID is `gpt-6-astra`, but migration requires integration changes. Tool calling with Astra requires the Responses API even though the model also supports Chat Completions. OpenAI also instructs developers to remove unsupported parameters including `temperature`, `top_p`, and `top_logprobs`, plus `logprobs` in Chat Completions. [3]
- Set the model ID to `gpt-6-astra`.
- Use the Responses API for tool calling.
- Remove `temperature`, `top_p`, and `top_logprobs`; remove `logprobs` in Chat Completions.
- If you used `none` or `minimal`, start testing with `low`.
- Treat Fast mode separately: it is unavailable for Astra with EU data residency.
Computer use: Agents’ Last Exam, OSWorld, and ScreenSpot-Pro
OpenAI reports 59.3% on Agents’ Last Exam, 72.6% on OSWorld 2.0 offline partial, and 92.7% on ScreenSpot-Pro without tools. It also says Astra reached higher OSWorld performance in simulations while taking about 47% less time per task than GPT-5.6 Sol. [1]
Professional work: AutomationBench, BenchCAD, and BrowseComp
| Benchmark | GPT-6 Astra | GPT-5.6 Sol | Claude Fable 5.1 | Claude Opus 5 |
|---|---|---|---|---|
| Agents' Last Exam | 59.3% | 53.6% | — | 55.5% |
| OSWorld 2.0 offline partial | 72.6% | 65.7% | — | 70.2% |
| AutomationBench | 41.4% | 18.1% | 31.4% | 26.9% |
| BenchCAD | 95.9% | 83.3% | 84.3% | 82.1% |
| Terminal-Bench 4.0 | 57.9% | 37.3% | 55.8% | 52.3% |
| DeepSWE v1.1 | 74.1% | 72.7% | 67.4% | 73.7% |
| Terminal-Bench Science 0.1 | 64.6% | 22.4% | 52.6% | 30.0% |
| FrontierMath Tier 4 (v2) | 97.6% | 83.0% | 87.8% | 73.2% |
| GPQA Diamond | 96.0% | 94.6% | 93.7% | 93.7% |
| HLE with tools | 57.2% | — | 65.0% | 63.6% |
Astra scores 41.4% on AutomationBench versus 18.1% for GPT-5.6 Sol and 31.4% for Claude Fable 5.1. It reaches 95.9% on BenchCAD and 91.5% on BrowseComp. These are launch configurations reported by OpenAI, not a guarantee of superior performance in every business workflow. [1]
Coding: Terminal-Bench 4.0, DeepSWE, and FrontierCode
Astra reaches 57.9% on Terminal-Bench 4.0, 74.1% on DeepSWE v1.1, and 64.5% on FrontierCode 1.1 Extended. The coding picture is not one-sided: several competing models remain close or higher on individual coding evaluations. [1]
Science and math: very strong results without universal dominance
Astra records 64.6% on Terminal-Bench Science 0.1, 97.6% on FrontierMath Tier 4 (v2), and 96.0% on GPQA Diamond. Yet on Humanity’s Last Exam with tools, OpenAI reports 57.2%, while the same table lists Claude Fable 5.1 at 65.0% and Claude Opus 5 at 63.6%. [1]
Long context: MRCR up to 1M tokens
On OpenAI MRCR v2, Astra scores 100.0% in the 256K–512K range and 96.3% at 512K–1M. That is promising for very long inputs, but production economics must account for the higher pricing multiplier once input exceeds 272K tokens. [1][2]
ARC-AGI-3: why the 99.9% score needs methodology context
OpenAI shows 99.9% on ARC-AGI-3, but that score uses a Provider Adapter that preserves reasoning state and supports compaction. ARC Prize reports 62.7% for the same model under its Standard harness and 99.9% with the Provider Adapter. It also reports that Astra used fewer actions than the median tested human on 96% of levels. [1][5]
Independent Artificial Analysis testing: stronger coding, higher price
Artificial Analysis scores Astra at 61 on its Intelligence Index at `max` and 67 on its Coding Agent Index. Its independent testing finds a large token-efficiency gain for coding, while the higher token price makes Astra more expensive per task than GPT-5.6 Sol at max effort on the general Intelligence Index. [6]
Cybersecurity: OpenAI’s first Critical-capability model
OpenAI classifies Astra as its first model to reach the Critical cybersecurity capability level under the Preparedness Framework. Without production safeguards, Astra scored 100% on ExploitBench, 42.4% on ExploitGym, and 88.0% on SRE-Bench in one attempt; the system card also describes advanced zero-day discovery and exploitation capability in expert testing. [1][4]
Safety and monitorability: better behavior, harder reasoning oversight
The system card presents a dual result: Astra is better aligned overall, yet its written chain of thought is harder to monitor than GPT-5.6 Sol in evaluations that explicitly pressure the model to evade monitoring. OpenAI says these were adversarial conditions and that overall alignment evaluations still favor Astra. [4]
Prompt-injection robustness and task-boundary adherence
OpenAI reports stronger prompt-injection robustness and better adherence to authorized task boundaries. In a simulation of more than 54,000 internal Codex tasks, Astra received roughly half as many flags for higher-severity misaligned behavior as GPT-5.6 Sol. [4]
New API capabilities: async tools and mid-turn steering
Astra introduces async tool calling, mid-turn steering, and mid-conversation reasoning changes through `configuration_update` without rewriting the original prompt prefix. These features matter for agents that wait on slow tools or must absorb new instructions without restarting the entire task. [3]
Codex and working memory across context windows
In Codex, OpenAI is testing notes that persist across context-window boundaries. Earlier windows remain searchable, allowing Astra to recover requirements or test results that were not captured in a compacted summary. The feature is experimental and is planned to become the Astra default later. [1]
Data retention: ZDR and Private Safety Processing
OpenAI says Astra supports Zero Data Retention for eligible API customers. It is also testing Private Safety Processing, designed to detect safety patterns across related interactions without exposing underlying customer content to OpenAI personnel; the feature remains in early-customer testing. [1][9]
Rollout across ChatGPT, API, Azure, and Bedrock
OpenAI announced Astra for ChatGPT Plus, Pro, Business, and Enterprise, the API, Microsoft Azure, and Amazon Bedrock. Pro, Business, and Enterprise users are also promised GPT-6 Astra Pro, but the launch materials do not publish a separate public API model ID or API price for that Pro variant. Enterprise administrators must enable Astra because it is off by default at launch. [1][8]
GPT-6 Astra vs GPT-5.6 Sol and Claude: where the advantage is real
Astra’s clearest launch gains over GPT-5.6 Sol appear in agentic workflows, computer use, Terminal-Bench Science, and several professional evaluations. It does not dominate Claude across the board: Fable 5.1 is higher on HLE with tools, and OpenAI’s own comparison table lists a higher Artificial Analysis Intelligence Index score for Fable 5.1. Model selection should therefore be workload-specific. [1][6]
What actually changed from GPT-5.6 Sol?
Compared with GPT-5.6 Sol, Astra changes more than quality: it has different reasoning settings, higher per-token pricing, new asynchronous tool and steering capabilities, stronger long-context results, and a much higher cyber capability level. A production migration should retest prompts, tools, costs, and safety behavior rather than only swapping the model name. [2][3][4][11]
