AI Radar

GPT-6 Astra

Available · Language, Multimodal, Reasoning · Milestone

GPT-6 Astra is the first model of OpenAI's GPT-6 generation, a reasoning language model for text and image input released on 3 September 2026 in ChatGPT Work, Codex, the API and Amazon Bedrock, after a first rollout to a limited set of organizations. OpenAI says it is its first model to reach the Critical level of cybersecurity capability under its Preparedness Framework. [1] [2] [3] [4] [5] Primary source

Timeline of GPT-6 Astra →

Claims and evidence

  • Announced OpenAI first named the model as Astra, one of its upcoming models, in a post on preliminary cybersecurity evaluations; the GPT-6 prefix appeared at launch. A further pre-release update, Path to Astra, followed on 2026-09-01. Primary source[6]second paragraph
  • Released Rolled out first to a limited set of organizations, with wider availability in ChatGPT, the API and AWS over the following days (launch post); the API changelog lists the release on Sep 3. Primary source[2]introduction; Availability [7]September, 2026: Sep 3 [5]Published September 3, 2026; 1. Safety Overview
  • Status AvailableListed in the API model catalog and not on the deprecations page as read on 2026-10-01. Primary source[1]Snapshots [8]Upcoming deprecations
  • Successor of GPT-5.6 Sol EditorialEditorial link: Astra opens a new tier above Sol, and no source names a direct predecessor. The launch post calls GPT-5.6 Sol the previous frontier cyber-capable model, the system card compares Astra throughout with GPT-5.6 Sol, and Path to Astra measures its cyber capabilities against GPT-5.6 Sol. Primary source[2]Cybersecurity section (GPT-5.6 Sol called the previous frontier cyber-capable model) [5]1. Safety Overview [9]Assessing Astra’s cybersecurity capabilities
  • Change · Safety Rated Critical in cybersecurity (GPT-5.6 Sol: High); OpenAI added misalignment monitoring to tool-using inference in external deployment.Compared with GPT-5.6 Sol Primary source[5] [6] [10]
  • Change · Tool use Adds async tool calling, mid-turn steering and changes of reasoning effort mid-conversation in the Responses API, which is also required for tool calling.Compared with GPT-5.6 Sol Primary source[7]September, 2026: Sep 3
  • Change · Reasoning Drops the none reasoning effort level; custom temperature and top_p values and log probabilities are not supported either.Compared with GPT-5.6 Sol Primary source[7]September, 2026: Sep 3
  • Change · Training data Knowledge cutoff of 30 April 2026, against 16 February 2026 for GPT-5.6 Sol.Compared with GPT-5.6 Sol Primary source[1]Model details [11]Model details
  • Input text, image Primary source[1]Model details
  • Output text Primary source[1]Model details
  • Feature Reasoning mode Primary source[1]description (reasoning.effort); Model details (reasoning token support) [5]2. Model Data and Training
  • Feature Function callingTool calling requires the Responses API (API changelog, Sep 3). Primary source[1]Supported features
  • Feature Computer use Primary source[1]description; Supported tools [7]September, 2026: Sep 3
  • Context window 1,050,000 tokensThe docs give 922,000 maximum input tokens and 128,000 maximum output tokens. Primary source[1]Model details
  • Parameters Not disclosedNot disclosed by OpenAI. No claim madeNo source
  • Access API, consumer app, cloud partner Primary source[2]Availability [7]September, 2026: Sep 3 [3]August 31–September 4, 2026: Take on demanding work with GPT-6 Astra [4]How Bedrock availability works

Lineage

Predecessors

Successors

No known successor.

Based on

Not derived from another model.

Variants and derived

None recorded.

Siblings

None recorded.

All ancestors

All descendants

None.

Variants

No variants recorded in this record.

Related AI Radar coverage

All model releases from OpenAI on AI Radar →