A.I.
Rycode

Jeffrey 3 min read

Frontier models

GPT-6 Astra

GPT-6 Astra is OpenAI’s flagship for hard end-to-end work: reasoning, coding, computer use, research, and documents. The model id is gpt-6-astra. It ships on the OpenAI API, ChatGPT plans as they roll out, Microsoft Azure, and AWS Bedrock.

Status
Current
AA score
53 (max effort)Artificial Analysis Intelligence Index, one independent score across many tests, fetched 22 Sep 2026
Price
$10 in, $50 outUS dollars per million tokens (word pieces) in and out, unless noted

Training

Knowledge cutoff is April 30, 2026 on the API model page. Parameter count, training FLOPs, and the full post-training recipe for Astra are not published on the API page or announce. The GPT-6 Astra System Card (published 3 Sep 2026) covers safety and health evals, not a training recipe.

Inference

reasoning.effort supports low, medium, high, xhigh, and max. There is no none. Reasoning tokens are supported. Chat Completions works for Astra, but tool calling requires the Responses API. Unsupported sampling knobs on this path include temperature, top_p, and top_logprobs.

Context

The documented context window is 1.05 million tokens. Max output is 128K tokens. Inputs are text and images; output is text. Audio and video are not supported. Streaming, function calling, and structured outputs are supported. Fine-tuning is not.

Methods

Tools named for Astra on the Responses API include web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search. OpenAI also describes mid-turn steering, async tool calling, compaction for long sessions, and misalignment monitoring. OpenAI states Astra meets the Critical cybersecurity threshold under its Preparedness Framework.

Benchmarking

Numbers below are from OpenAI’s GPT-6 Astra announce. The Sol column is OpenAI’s Sol, not Anthropic’s. CursorBench for Astra is not on that announce extract.

EvalGPT-6 AstraGPT-5.6 Sol (OpenAI)Notes from OpenAI page
Terminal-Bench 4.057.9%37.3%OpenAI also lists Claude Fable 5.1 at 55.8%
Terminal-Bench Science 0.164.6%22.4%OpenAI lists Fable 5.1 52.6%, Opus 5 30.0%
OSWorld 2.0 (latency sim prose)72.6% @ ~40 min/task65.7% @ ~75 min~47% less time than Sol
AutomationBench41.4%18.1%OpenAI lists Fable 5.1 31.4%, Opus 5 26.9%
BrowseComp91.5%90.4%
BenchCAD95.9%83.3%
DeepSWE v1.174.1%72.7%
FrontierMath Tier 4 (v2)97.6%83.0%Prose also describes ~98% / saturates
GPQA Diamond96.0%94.6%
Humanity's Last Exam (w/ tools)57.2%-OpenAI lists Fable 5.1 65.0%
ARC-AGI-3saturates 99.9%; human action-efficiency baseline on 96% of levels-prose
ExploitBench100%-prose

Health (system card Table 6, length-adjusted): HealthBench Professional 63.4 (Sol 60.5); HealthBench 58.1 (Sol 57.0); HealthBench Hard 36.3 (Sol 33.1).

Safety (system card, not a capability leaderboard): OpenAI reports stronger jailbreak robustness versus Sol. In an internal Codex deployment simulation across 54,218 tasks, Astra had 34 severity-3-or-higher flags versus more for Sol (about half on that comparison). Production safe-completion rates improve versus Sol across categories in the card tables.

Hardware

OpenAI has not published training or serving silicon details for Astra on the model docs. Company infrastructure posts name Stargate Abilene (Oracle + NVIDIA GB200) as a site that trained GPT-5.5. That is company context, not a published Astra GPU map. Buyers purchase API tokens (or Azure / Bedrock meters), not an Astra GPU SKU. Weights are closed.

Access and pricing

Standard API pricing is $10 input and $50 output per million tokens. Cached input is $1.00. Cache writes are $12.50. Prompts with more than 272K input tokens are priced at 2× input and cache rates and 1.5× output for the full request. Batch and Flex are 50% of Standard. Fast mode is 2× the applicable rates. Access rolled out through limited organizations and ChatGPT Plus, Pro, Business, and Enterprise plans; check live pages for current availability. Pro, Business, and Enterprise also get GPT-6 Astra Pro. Enterprise administrators enable Astra for the workspace.

Sources