Benchgen
Models/openai/

GPT-6 Astra

DraftPublic

Model Details

GPT-6 Astra

Organization Pricing License Modality Released

Quick answer: GPT-6 Astra is OpenAI's new flagship model generation, released September 3, 2026, succeeding GPT-5.6 Sol. OpenAI positions it as its most capable and most aligned model to date — state-of-the-art on computer use, browsing, software engineering, cybersecurity, and science. It scores 97.6% on FrontierMath Tier 4, 96.0% on GPQA Diamond, 99.9% on ARC-AGI-3, and 100.0% on ExploitBench. Priced at $10 input / $50 output per 1M tokens (Fast mode: 2x price, 2x speed).

At a Glance

Where GPT-6 Astra leads

  • Near-saturates FrontierMath Tier 4 (97.6%) and ARC-AGI-3 (99.9%)
  • SOTA computer-use: OSWorld 2.0 72.6% in ~47% less time per task than GPT-5.6 Sol
  • ExploitBench 100.0%, SRE-Bench 88.0% — largest cybersecurity capability jump OpenAI has shipped, formally meeting the "Critical" threshold under its Preparedness Framework
  • Substantially improved alignment: 0% unauthorized-scope-expansion rate on an impossible-task eval where GPT-5.6 Sol was 48%

Where it's close, not clearly ahead

  • Artificial Analysis Intelligence Index (61.2) trails Claude Fable 5.1 (65.7) and Claude Opus 5 (63.1)
  • Humanity's Last Exam with tools (57.2%) trails Claude Fable 5.1 (65.0%), Fable 5 (63.8%), Opus 5 (63.6%)

Best for: production agentic workflows (computer use, coding, defensive cybersecurity) where speed-per-task and instruction-following judgment matter as much as raw benchmark ceiling.

What GPT-6 Astra Is

GPT-6 Astra is OpenAI's next-generation flagship model, combining pre-training, reinforcement learning, and alignment research advances. OpenAI frames it as a step change in computer use, professional knowledge work, coding, science, and cybersecurity, alongside what the company describes as its largest alignment improvement to date — including a new evaluation (informed by the Hugging Face incident) testing whether a model facing an impossible task expands beyond its authorized scope. Astra also contributed to two new mathematical proofs on the distribution of prime gaps.

Astra rolled out September 3, 2026 to a limited set of organizations, with staged availability to ChatGPT Plus/Pro/Business/Enterprise and via the OpenAI API, Microsoft Azure, and AWS Bedrock. Enterprise access is off by default at launch. A "Fast mode" is available in the API at 2x Standard price for roughly 2x the speed.

Specifications

FieldValue
OrganizationOpenAI
API model namegpt-6-astra
LicenseProprietary (API only)
ModalityMultimodal (text + vision)
Release date2026-09-03
AvailabilityOpenAI API, ChatGPT, Microsoft Azure, AWS Bedrock
Zero Data RetentionSupported for eligible API customers

Pricing

Input (per 1M tokens)Output (per 1M tokens)
Standard$10.00$50.00
Fast mode$20.00$100.00

Public Benchmark Scores

Scores above are reported by OpenAI's official GPT-6 Astra announcement and shown for context; not independent Benchgen measurements. Several evals reported alongside these (Internal Design/Data Science/Database Migration Tasks, internal computer-use safety/circumvention/hallucination evals, ExploitGym honeypot & "Impossible ExploitGym" variants, the dated "ExploitBench (June–Aug 2026)" subset, GeneBench Pro, MedChemBench, LifeSciBench, and OpenAI's own MRCR v2 8-needle long-context variant) are omitted here — either OpenAI-internal/single-model evals with no cross-model comparability, or variants that don't cleanly match Benchgen's existing benchmark methodology.

Frequently Asked Questions

What is GPT-6 Astra? GPT-6 Astra is OpenAI's flagship model generation released September 3, 2026, succeeding GPT-5.6 Sol, with major gains in computer use, coding, science, and cybersecurity.
How much does GPT-6 Astra cost? $10 per million input tokens and $50 per million output tokens (Standard); Fast mode is 2x price for roughly 2x speed.
Is GPT-6 Astra available via API? Yes — as gpt-6-astra via the OpenAI API, as well as Microsoft Azure and AWS Bedrock.
How does GPT-6 Astra compare to Claude Fable 5.1? Astra leads on FrontierMath, ARC-AGI, ExploitBench, and computer-use speed; Claude Fable 5.1 leads on the AA Intelligence Index (65.7 vs 61.2) and Humanity's Last Exam with tools (65.0% vs 57.2%).

Specs and scores sourced from OpenAI's official GPT-6 Astra announcement. Last updated 2026-09-07.