A.I.

Frontier models

GPT-6 Astra

GPT-6 Astra is OpenAI's most capable model for hard end-to-end work: complex reasoning, coding, computer use (the model drives a desktop or browser for you), research, and document creation. Model id: gpt-6-astra.

As of OpenAI's announce (fetched 5 Sep 2026), Astra was rolling out to a limited set of organizations and, over the coming days, to ChatGPT Plus, Pro, Business, and Enterprise, plus the OpenAI API, Microsoft Azure, and AWS Bedrock. Pro, Business, and Enterprise plans also get GPT-6 Astra Pro. Enterprise administrators enable Astra for the workspace; access is off by default at launch. That access language will go stale. Re-check the live pages before publish. Prefer this announce wording over older “Trusted Access Program” card language.

Standard API pricing: $10 per million input tokens and $50 per million output tokens. Cached input $1.00. Cache writes $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. This page is a learning model card under the OpenAI door on rycode.dev, not a news essay.

Training

Knowledge cutoff: April 30, 2026 (API model page).

What OpenAI has not published for this SKU on the API page or announce extract used here: parameter count, training FLOPs, dataset mix, and any RLHF / RLVR / RLAIF recipe tied specifically to Astra. Mark those UNKNOWN. Do not use secondary "N GPUs trained Astra" claims without an OpenAI primary.

The GPT-6 Astra System Card was published September 3, 2026 (OpenAI Deployment Safety hub). Use it for safety and health eval numbers below, not as a substitute for a missing training recipe.

Inference

reasoning.effort (how hard the model thinks before answering) supports low, medium, high, xhigh, and max. There is no none on Astra. Reasoning tokens are supported.

Migration note from OpenAI's latest-model guide: if you currently use none or minimal on an older model, start with low on Astra and compare. Prefer reasoning.effort on the Responses API, or reasoning_effort on Chat Completions.

Chat Completions works for Astra, but tool calling requires the Responses API. Remove unsupported sampling knobs: temperature, top_p, and top_logprobs (and related Chat Completions logprobs includes). For EU data residency, OpenAI says use Standard processing; Fast mode is not for that path in the guide extract used here.

Context

Context window: 1,050,000 tokens. Max output: 128,000 tokens.

Modalities (API model page): text in and out; image input only; audio not supported; video not supported.

Features listed as supported: streaming, function calling, structured outputs. Fine-tuning: not supported.

Methods

Tools supported on Astra when using the Responses API (API model page): web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP (Model Context Protocol), and tool search.

Related product behaviors OpenAI names in the guide family: mid-turn steering, async and programmatic tool calling patterns, compaction when long sessions fill the context window, and misalignment monitoring as a safety control surface. Exact train-time alignment recipe for Astra: UNKNOWN.

OpenAI states Astra meets the Critical cybersecurity threshold under its Preparedness Framework (announce + system card). Treat that as a safety classification, not a capability leaderboard score.

Relative to GPT-5.6 Sol (flagship of the prior GPT-5.6 family at $4 / $20 per million tokens, effort includes none): Astra is the GPT-6 flagship SKU at higher price and without none effort. Terra and Luna remain lower-cost GPT-5.6 siblings and are not Astra.

Benchmarking

Numbers below are from OpenAI's GPT-6 Astra announce page (openai.com/index/gpt-6-astra/), fetched 5 Sep 2026. The Sol column is OpenAI's Sol, not Anthropic's Sol column from Anthropic pages. Do not merge lab tables. CursorBench for Astra: UNKNOWN (not on the OpenAI announce extract used here).

Eval GPT-6 Astra GPT-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% (prose also ~98% / saturates)83.0%
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 scores; higher is better): 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 vs Sol in its testing program. In an internal Codex deployment simulation across 54,218 tasks, Astra had 34 severity-3-or-higher flags (0.063%) vs more for Sol; OpenAI states about 53% fewer severity level 3 actions vs Sol on that comparison. Production safe-completion rates improve vs Sol across categories in the card tables (example category names appear on the hub; quote individual rates only with the live card open at QC time).

Hardware

Per-SKU training or serving silicon for Astra: UNKNOWN. Buyers purchase API tokens (or Azure / Bedrock meters), not an "Astra GPU" SKU. Weights are closed.

Company infrastructure context only: OpenAI's compute post names Stargate Abilene (Oracle + NVIDIA GB200) as a site that trained GPT-5.5. That is not a published "Astra trained on N GB200s" claim. Do not invent one.

Sources