.png)
GPT-6 Astra Benchmarks Explained: Scores, Charts, Price
OpenAI released GPT-6 Astra on September 3, 2026, and the GPT-6 Astra benchmarks in its launch table look dominant: 97.6% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and a perfect 100% on ExploitBench. But independent testing tells a more modest story. If you are a developer, freelancer or business owner in the US deciding whether Astra is worth $10 per million input tokens, the scores only help once you know who ran the test and under what settings.
Quick Answer: How Good Are the GPT-6 Astra Benchmarks?
Short version:
• In OpenAI's own tables, Astra wins on math, cybersecurity and computer use.
• On the independent Artificial Analysis Intelligence Index, Astra scores 61.2, basically tied with GPT-5.6 Sol (60.9) and behind Claude Fable 5.1 (65.7).
• On agentic coding it is a near tie with Claude Opus 5 and Fable 5.1, and its real edge is cost per task.
• The 99.9% ARC-AGI-3 score used a custom harness, so do not expect that from a plain API call.
• In OpenAI's own tables, Astra wins on math, cybersecurity and computer use.
• On the independent Artificial Analysis Intelligence Index, Astra scores 61.2, basically tied with GPT-5.6 Sol (60.9) and behind Claude Fable 5.1 (65.7).
• On agentic coding it is a near tie with Claude Opus 5 and Fable 5.1, and its real edge is cost per task.
• The 99.9% ARC-AGI-3 score used a custom harness, so do not expect that from a plain API call.
What Is GPT-6 Astra?
GPT-6 Astra is the first model in OpenAI's GPT-6 family. The API model ID is gpt-6-astra, with a context window of about 1.05 million tokens and up to 128,000 output tokens. It accepts text and images and outputs text. It reached ChatGPT Plus, Pro, Business and Enterprise users within days of launch, and it is available through the OpenAI API, Microsoft Azure and Amazon Bedrock. Pro, Business and Enterprise plans also unlock a GPT-6 Astra Pro tier.
GPT-6 Astra Benchmarks vs Claude Fable 5.1 and Opus 5 (OpenAI-Reported)
The chart below plots four headline results from OpenAI's launch table. Every number here is vendor-reported, meaning OpenAI chose the settings and ran the comparison, so read it as OpenAI's best case.
GPT-6 Astra vs Fable 5.1 vs Opus 5: Selected Benchmarks (%)
Source: OpenAI launch table (vendor-reported). Higher is better.
Terminal-Bench 4.0 (coding in the terminal)
Astra
57.7
Fable 5.1
55.8
Opus 5
52.3
DeepSWE v1.1 (agentic software engineering)
Astra
74.1
Fable 5.1
67.4
Opus 5
73.7
Humanity's Last Exam, with tools (hard reasoning)
Astra
57.2
Fable 5.1
65.0
Opus 5
63.6
FrontierMath Tier 4 (research-level math)
Astra
97.6
Fable 5.1
87.8
Opus 5
73.2
Two things stand out. Astra's lead on FrontierMath Tier 4 is large, and that is where the "generational leap" talk comes from. But on DeepSWE v1.1 the field is tightly bunched, with Opus 5 within a fraction of a point and Google's Gemini 3.8 Flash at 73.8%. On Humanity's Last Exam with tools, Claude Fable 5.1 actually comes out ahead.
A Note on the Terminal-Bench Number
Some coverage lists Astra at 57.9% on Terminal-Bench 4.0 while OpenAI's table shows 57.7%. OpenAI also reportedly changed a few figures when it withdrew and republished its announcement. We use 57.7% from the table and treat the gap as noise, but it is a good reminder to check the footnotes.
GPT-6 Astra Computer Use Benchmarks: Where It Pulls Ahead
Computer use is the category OpenAI leaned on hardest, and it is where Astra looks strongest in practice.
| Benchmark | GPT-6 Astra | GPT-5.6 Sol | Other |
|---|---|---|---|
| OSWorld 2.0 | 72.6% | 65.7% | Opus 5: 70.2% |
| ScreenSpot-Pro (no external tools) | 92.7% | 76.9% | Fable 5: 87.3% |
| Agents' Last Exam | 59.3% | 53.6% | Opus 5: 55.5%, Fable 5.1: 48.7% |
| Time per OSWorld 2.0 task (OpenAI latency simulation) | about 40 min | about 75 min | Roughly 47% faster |
For anyone delegating real desktop work, such as filling forms, updating a CRM or building a spreadsheet, the time drop matters as much as the score. One caveat: Claude Fable 5.1 is missing from OpenAI's OSWorld row, and Anthropic reports its own number on a different release of the test, so a clean head-to-head does not exist yet.
GPT-6 Astra ARC-AGI-3 Score: Why 99.9% Needs an Asterisk
Astra's 99.9% on ARC-AGI-3 is the number most headlines repeat. ARC-AGI-3 tests whether a model can learn interactive environments it has never seen. According to OpenAI's footnote, the result came from a modified Responses API harness with two settings changed. Secondary coverage reports that ordinary stateless API calls score lower. So treat 99.9% as a ceiling under tuned conditions, not what your app will see on day one.
Independent GPT-6 Astra Benchmarks: The Artificial Analysis Index
Artificial Analysis runs every model through one neutral harness, which makes it the fairest cross-model check available. Here is how the Intelligence Index (v4.1.1) looks on a 0 to 100 scale.
Artificial Analysis Intelligence Index v4.1.1
Independent neutral harness. Higher is better.
Fable 5.1
65.7
Opus 5
63.1
GPT-6 Astra
61.2
GPT-5.6 Sol
60.9
Astra effectively ties the model it replaced and sits about five points behind Fable 5.1. That is the strongest counterweight to the "AGI era" framing. Independent testing did show real gains elsewhere:
- Coding cost: on the Artificial Analysis Coding Agent Index, Astra scored 67.0, roughly level with the Fable models, while matching Fable 5's score at under half the cost per task.
- Hallucinations: the rate on Artificial Analysis's knowledge test reportedly fell from 92% to 51% at maximum effort. Better, but still high enough that you should verify factual claims.
GPT-6 Astra Cybersecurity Benchmarks and the "Critical" Gate
Astra scored a perfect 100% on ExploitBench (Sol: 78.5%, Opus 5: 70%) and 88.0% on SRE-Bench against 12.5% for Fable 5.1, which refused parts of these tests. OpenAI says Astra is the first model to cross the "Critical" cybersecurity threshold in its own Preparedness Framework, so the most advanced cyber features are limited to vetted defenders in its Daybreak program. In ChatGPT or Codex you may be asked to confirm risky actions, and in the API a flagged task can stop outright. Expect occasional friction on security-related work.
GPT-6 Astra Price: What It Costs in US Dollars
| Item | GPT-6 Astra |
|---|---|
| Input (per 1M tokens) | $10 |
| Output (per 1M tokens) | $50 |
| Cache reads | 90% discount |
| Fast mode | Up to 2x speed at 2x price |
| Price vs GPT-5.6 Sol | 2.5x higher |
That is roughly on par with Claude Fable 5.1. Per-token price is not per-job price: Astra uses far fewer tokens on coding tasks, so it can finish a job cheaper than rivals. On general knowledge work, Artificial Analysis found Astra about 75% more expensive per task than Sol at maximum effort.
GPT-6 Astra vs Claude Fable 5.1: Which Should You Pick?
- Choose Astra for computer-use agents, long terminal sessions, math-heavy work and token-efficient coding at scale.
- Choose Fable 5.1 for hard general reasoning, tool-assisted research and tasks where the neutral index matters more than a vendor table.
- Test both on your own prompts. A 1 to 2 point benchmark gap rarely survives contact with real work.
Is GPT-6 Astra AGI?
OpenAI president Greg Brockman has called Astra a "generational leap" and suggested it may signal the AGI era. There is no agreed definition of AGI, and the saturated benchmarks behind that pitch (ARC-AGI-3 and FrontierMath) are narrow or harness-sensitive. On the neutral aggregate Astra looks incremental. It is a strong model that leads specific tests under specific conditions, which is what the evidence supports.
Final Verdict
The honest reading of the GPT-6 Astra benchmarks is that source and setup decide the story. Astra is a clear step up from GPT-5.6 Sol in computer use, speed and coding efficiency, and it dominates a few very hard tests. It is not a clean sweep over Claude Fable 5.1, and the neutral index shows an incremental gain. Pick the benchmark that matches your job, then run your own test.
Benchmarks were compiled from OpenAI's launch materials and coverage by OfficeChai, Emergent, DataCamp and Artificial Analysis in September 2026. Vendor-reported figures are not independently reproduced, and newer results may have appeared since.
