GPT-6 Astra: Launch Details, Pricing, Codex Computer Use and How It Compares With Claude Opus

GPT-6 Astra launch and availability

GPT-6 Astra’s API model ID is gpt-6-astra. Its launch covered both developer workflows and professional work: software installation and testing, frontend QA, online research, form completion, CRM updates, and the creation of documents, spreadsheets, and presentations that match existing templates.

Access was rolling out in stages. The OpenAI launch announcement described initial availability for selected organizations, followed over the coming days by ChatGPT Plus, Pro, Business, and Enterprise, the OpenAI API, Microsoft Azure, and Amazon Bedrock. As of September 5, eligibility did not establish that every account already had access.

ChatGPT plans and missing model access

The announced ChatGPT plans include Astra usage within subscription allowances, with additional credits available for purchase. OpenAI also announced GPT-6 Astra Pro for Pro, Business, and Enterprise.

Enterprise access was off by default and required administrator activation. A missing model picker entry therefore needs to be interpreted in context: rollout timing matters, Enterprise settings matter, and Codex had its own picker changes during the launch period.

OpenAI classified Astra as its first broadly deployed model to reach the “Critical level of cybersecurity capability” under its Preparedness Framework. The stated production safeguards include monitoring tool-using inference for misaligned behavior. That classification describes OpenAI’s capability assessment and deployment safeguards; it is not a measure of coding quality.

GPT 6 Astra pricing and context limits

The confirmed launch prices are straightforward for Standard processing. Fast mode was announced with up to twice Standard speed at twice Standard price.

Astra’s Standard input and output token prices are twice those of Opus 5. That comparison does not establish twice the cost per completed task: the price table does not tell you how many tokens each model will need to finish your work.

For a developer evaluation, compare both token spending and whether the resulting change completes the task. Fast mode’s “up to” wording also matters when setting expectations; the announcement does not promise that every request finishes twice as quickly.

Caching, long prompts, and output limits

Exact historical pricing requires more care than the Standard rates. The research brief could not establish that the live model documentation’s cached-input prices, cache-write prices, long-prompt surcharges, or Batch and Flex terms were published by September 5. Those figures are therefore excluded from this launch-date pricing guide.

The same limitation applies to Astra’s exact context and maximum-output specification block. Its historical publication date was not established, so those numbers should not be presented as confirmed September 5 specifications.

Opus 5’s dated release details do establish a default and maximum context of 1,000,000 tokens and a maximum output of 128,000 tokens. This leaves an evidence gap for an exact launch-date context comparison with Astra, rather than grounds for declaring either model the context-window winner.

For the separate question of Claude Code costs, see the Claude Code pricing guide.

API migration: tools require Responses

The September 3 API changelog identifies several requirements that matter when moving an existing integration to gpt-6-astra.

Astra does not support reasoning effort none, custom temperature, custom top_p, or logprobs. Tool calling requires the Responses API, so an integration using tools through Chat Completions must migrate.

Review request parameters before switching models

Changing only the model ID is insufficient if your requests use any of those unsupported settings. Review shared request builders as well as model-specific configuration: an existing default for temperature or reasoning effort can become incompatible with Astra.

The model documentation lists reasoning effort values of low, medium, high, xhigh, and max. Keep the distinction between selecting an effort level and disabling reasoning: Astra does not support the none setting.

The September 3 announcement also introduced asynchronous tool calling, mid-turn steering over WebSockets, and changing reasoning effort during a conversation while preserving the cached prompt prefix. These are relevant integration capabilities, but they do not remove the Responses API requirement for tools.

For a migration review, the immediate checks are therefore the endpoint used for tool calls, unsupported request fields, and the reasoning configuration. Resolve those before using task results to judge the model.

GPT-6 Astra Codex changes and computer use

The Astra launch paired the model with an updated Codex harness. OpenAI reported 1.9× faster task completion on Mind2Web compared with the existing GPT-5.6 Sol experience.

That result measures Astra together with the updated harness on a particular benchmark. It is a vendor-reported result, not a universal speed guarantee or an isolated measurement of the model’s contribution.

Astra’s computer-use examples include software installation and testing, frontend QA, form completion, CRM updates, and online research. For developers, installation, testing, and frontend QA are useful task categories to include in an evaluation because they exercise the announced capabilities.

What changed across the launch-period CLI releases?

The release sequence explains why Astra could appear differently across Codex installations:

• September 3, version 0.153.1: added Astra configuration through the API without making it the default or showing it in the model picker.
• September 3, version 0.153.2: changed the Fast-tier label to “2x speed, increased usage.” This was a displayed-text change only.
• September 4, version 0.153.3: added Astra to Amazon Bedrock model pickers for Mantle and Runtime global/US routes.
• September 4, version 0.153.4: fixed bundled picker visibility and made Astra the bundled default when no model was explicitly configured.

The Codex 0.153.4 release also specifies that asynchronous questions are used only when the session provides the tool.

To reproduce that release version:

npm install -g @openai/codex@0.153.4

A bundled default does not establish account eligibility. It also has a specific condition: no model is explicitly configured.

Enable experimental context management

The launch described experimental context management that can preserve notes across context windows and search earlier messages and tool outputs. Default activation “in the coming weeks” was a launch plan, not evidence that it was already enabled by September 5.

The September 4 configuration schema verifies this Boolean setting:

[features.context_management]
experimental_mode = true

The schema confirms the configuration key and its purpose. It does not establish account eligibility or prove that setting it successfully activates the feature for every account.

For the broader workflow comparison, see the Claude Code versus Codex guide.

GPT-6 Astra vs Claude Opus 5

Claude Opus 5 is the contemporary Opus comparison for this publication date. It launched on July 24, 2026, with API ID claude-opus-5.

The Claude release notes establish its pricing, context limits, thinking controls, and the August release of generally available Claude computer use.

Coding and computer-use benchmarks

OpenAI’s launch table reported the following comparisons:

Astra scored higher on these three selected OpenAI-reported comparisons. OpenAI’s evaluation caveat says the scores are maxima across effort settings and that production behavior can differ.

These results provide evidence for evaluating Astra on agent, computer-use, and terminal tasks. They do not establish that it will outperform Opus 5 on every repository or justify its higher token prices for every workload.

Thinking controls and Claude computer use

Opus 5 supports effort levels low, medium, high, xhigh, and max. Setting thinking: {"type": "disabled"} is allowed only at high or below; combining disabled thinking with xhigh or max returns HTTP 400.

Claude computer use was already generally available by the cutoff. The August 19 release introduced computer_toolset_20260801, requiring no beta header, supporting batch actions, enabling zoom by default, and using configs for member configuration.

Claude Fable 5.1 is another contemporary comparison, having launched September 1 with $10/$50 input/output pricing per million tokens—the same headline rates as Astra Standard. Its API ID is claude-fable-5-1; it has 1M default context, 128K maximum output, and always-on adaptive thinking. See the Claude Fable 5.1 model guide for that separate comparison.

Key takeaways

• GPT-6 Astra launched September 3 with API ID gpt-6-astra; staged availability means eligible accounts were not guaranteed access by September 5.
• Confirmed Standard API pricing is $10 per million input tokens and $50 per million output tokens, twice Opus 5’s token rates.
• Astra tool calling requires the Responses API, and several existing request settings are unsupported.
• Codex 0.153.4 fixed Astra’s bundled picker visibility and made it the default when no model was explicitly configured.
• Experimental context management has a verified configuration key, but eligibility and successful activation are not established by the schema.
• OpenAI’s selected benchmarks favor Astra over Opus 5, with evaluation caveats that limit conclusions about production workloads.

FAQ

What is GPT-6 Astra, and when did it launch?

GPT-6 Astra is OpenAI’s model for reasoning, coding, computer use, research, and document creation, launched September 3, 2026. Its API model ID is gpt-6-astra, and availability was rolling out in stages as of September 5.

What does GPT-6 Astra cost?

Confirmed Standard API pricing is $10 per million input tokens and $50 per million output tokens. Launch-announced Fast mode offers up to twice Standard speed at twice Standard price; exact caching, long-prompt, Batch, and Flex terms were not historically confirmed for the September 5 cutoff.

How do I enable experimental context management in Codex?

Set experimental_mode = true under [features.context_management], as verified in the September 4 Codex 0.153.4 schema. The schema confirms the setting but does not establish account eligibility or guarantee successful runtime activation.

How does GPT-6 Astra compare with Claude Opus 5?

Astra’s Standard input and output token prices are twice Opus 5’s, and Astra scored higher on three selected comparisons in OpenAI’s launch table. Those scores are maxima across effort settings, and production behavior can differ, so they do not establish a universal winner for coding or computer use.