Skip to content
AIMODELRACE
OpenAILanguage model

GPT-6 Astra

OpenAI's top tier: operates software like a person and leads the hardest math.

Compare head-to-headRace it in 3D

Overall Intelligence

#3

of 56
0.4 pts behind #1

7

races led

94

weeks at #1

88.4

score

Released
Sep 3, 2026
2 mo ago
Context
1.05M
128K max output
Price
$10 / $50
per 1M in/out · $20.00 blended
Speed
60 tok/s
11.5s to first token
Modalities
text · image · audio · video
outputs text
Weights
Closed
API and apps

The 30-second version

OpenAI's top tier. GPT-6 Astra operates computers like a skilled person and leads the hardest math benchmarks, making it the pick for autonomous software operation and research-grade problem solving.

Key takeaways
  • 1Leads our computer-use and research-math races in the current snapshot.
  • 2Top-three overall, within a photo finish of the leaders.
  • 3Premium pricing ($10 / $50) and modest speed: use it where its peaks matter.
  • 41M+ token context and strong tool use across OpenAI's agent stack.

What you need to know

Purpose

A frontier model for agents that act in real software environments and for problems that demand the deepest mathematical and scientific reasoning.

What OpenAI optimised for

Peak capability on reasoning and agentic computer use, tightly integrated with OpenAI's agent products and tools.

Unique strengths

  • Operates GUIs and browsers with human-like reliability.
  • Leading performance on research-level mathematics.
  • Deep integration with OpenAI's agent tooling.
Insight
How it differs from rivals. Against Claude Opus 5.5 it wins on computer use and math but is pricier and a little less steady on very long coding sessions. Against Fable 5.1 it is better at operating software, while Fable leads on extended investigations.

Where it ranks

Current snapshot · illustrative scores · click any race to watch it

CodingHard ReasoningMathScienceWriting & LanguageMultimodalTool Use & AgentsLong ContextTrust & FactualitySpeed & Cost

GPT-6 AstraBest on each axis

Overall Intelligence over time

#1 frontierGPT-6 AstraFable 5.1Gemini 4 Argon
55.060.065.070.075.080.085.090.02025Apr ’25Jul ’25Oct ’252026Apr ’26Jul ’26Oct ’26

Best use cases

01

Computer-use agents

Automating workflows across web apps and desktop software without APIs.

02

Advanced math and science

Proof sketches, research problems and quantitative modelling.

03

Hard reasoning

Problems where getting the logic exactly right is the whole job.

Strengths

  • Computer use

    Leads our computer-use race by a clear margin.

  • Research math

    Top of the research-grade math race.

  • Reasoning

    Consistently near the top on frontier reasoning.

Weaknesses

  • Cost

    Top-tier pricing; expensive for routine tasks.

  • Long-horizon coding

    Slightly behind the Claude tops on very long autonomous coding runs.

Prompts that play to its strengths

Software operation

In the browser, go to our vendor portal, download every invoice from September, rename each file to `vendor_YYYY-MM-DD_amount.pdf`, and upload them to the Finance/Invoices/2026-09 folder in Drive. Stop and ask if any invoice is missing an amount.

Why it works: A goal-oriented computer-use task with a clear end state and a sensible escalation rule.

Proof exploration

Here is a conjecture and our partial progress. Identify the weakest step in our argument, try two different approaches to repair it, and clearly separate what you have proven from what you believe.

Why it works: Uses its math strength while forcing it to be explicit about rigour.

When to choose it, and when not to

GPT-6 AstraoverOpus 5.5

The task is driving software through a GUI or research-level math.

GPT-6 AstraoverFable 5.1

You need computer use or contest math rather than a long investigation.

Pick something else when…

  • Cost-sensitive, high-volume workloads: GPT-6.1 Sol or Luna are far cheaper.
  • Video or audio understanding: Gemini 4 Argon leads.

Update history

The distilled version first; open any release for the deeper breakdown.