OpenAI shipped GPT-6 Astra on September 3, 2026, into the most crowded AI release window of the year. Two days earlier, Google had pushed Gemini 3.8 Flash. Three days before that, Anthropic launched Claude Fable 5.1 and Mythos 5.1. And one day before Astra arrived, Meta released Muse Spark 1.3, a model that matched Opus-level intelligence at a fraction of the price.
Astra is OpenAI’s answer to all of them. The company calls it the most intelligent and aligned model it has ever built, trained on more than 100,000 GPUs at the Stargate site in Texas. The benchmarks OpenAI selected for its launch page tell a story of dominance: massive gains on ARC-AGI-3, FrontierMath, ExploitBench, and computer-use tasks. Independent testing tells a more nuanced story.
What Is GPT-6 Astra?
Astra is OpenAI’s next-generation reasoning model, succeeding the GPT-5.6 family (Sol, Luna, and Terra). The galaxy-themed branding and the name itself suggest a larger model than Sol, likely a mixture-of-experts architecture, though OpenAI has not confirmed this.
The API model name is gpt-6-astra, available in reasoning effort settings from low through max, with a Fast latency mode that doubles throughput at twice the price. Artificial Analysis lists the max reasoning variant as released September 3, 2026, with a knowledge cutoff of April 2026.
Key specifications:
- 1.05 million token context window (up to 922,000 input, 128,000 output per request)
- Text and image input, text output
- Extended chain-of-thought reasoning with configurable effort levels
- Staged rollout: Daybreak cyber-defense partners first, then ChatGPT Plus, Pro, Business, Enterprise, API, and AWS
- No free tier announced; Enterprise access stays off until an administrator enables it
OpenAI says this was its largest training run to date, and the first where earlier OpenAI models played a significant role in supervising the training process. That training footprint also explains why Samsung is courting OpenAI so aggressively, alongside parallel AI chip partnerships with Mistral and Anthropic.
Every Benchmark From OpenAI’s Announcement
OpenAI’s launch materials present Astra as state-of-the-art across computer use, browsing, software engineering, cybersecurity, science, and professional work. The official GPT-6 Astra announcement publishes around two dozen scores.
The table below lists them all, with OpenAI’s own comparison figures for GPT-5.6 Sol and Claude Fable 5.1 where disclosed. Treat vendor numbers as vendor numbers: harnesses differ between labs, and first-party benchmarks run inside highly tuned native scaffolding. The independent table further down is the fairer cross-model comparison.
| Benchmark | GPT-6 Astra | GPT-5.6 Sol | Claude Fable 5.1 |
|---|---|---|---|
| ARC-AGI-1 | 98.5% (xhigh/high), 97.5% (max) | 97.5% (xhigh), 97.0% (high), 96.5% (max) | 97.5% (max/high), 96.5% (xhigh) |
| ARC-AGI-2 | 95.0% (max), 93.3% (xhigh) | 92.5% (max), 90.0% (xhigh) | 90.0% (max/xhigh), 88.8% (high) |
| ARC-AGI-3 | 99.9% (high) Provider Adapter; 62.7% (max) standardized | 7.8% (max) standardized | Not listed |
| FrontierMath Tier 4 v2 | 97.6% | 83.0% | 87.8% |
| GPQA Diamond | 96.0% | 94.6% | 93.7% |
| Humanity’s Last Exam with tools | 57.2% | Not disclosed | 65.0% |
| Terminal-Bench 4.0 | 57.7% | 37.3% | 55.8% |
| Terminal-Bench Science 0.1 | 64.6% | 22.4% | 52.6% |
| DeepSWE v1.1 | 74.1% | 70.8% | 67.4% |
| FrontierCode v1.1 Main | 53.3% | Not disclosed | Not disclosed |
| FrontierCode v1.1 Extended | 64.5% | 60.6% | Not disclosed |
| OSWorld 2.0 offline | 72.6% in 40 min/task | 65.7% in 75 min/task | Not disclosed |
| BrowseComp | 91.5% | 92.2% | Not disclosed |
| Agents’ Last Exam | 59.3% | 52.7% | Not disclosed |
| AutomationBench | 41.4% | 46.7% | 31.4% |
| ScreenSpot-Pro, no tools | 92.7% | Not disclosed | Not disclosed |
| BenchCAD Vision2Code | 95.9% | 83.3% | 84.3% |
| SRE-Bench public | 88% pass@1 | Not disclosed | Not disclosed |
| ExploitBench | 100% | 73.5% | Not disclosed |
| ExploitGym | 42.4% | 30.3% | Not disclosed |
| SEC-Bench Pro | 85.4% | Not disclosed | Not disclosed |
| GeneBench Pro | 37.8% | Not disclosed | Not disclosed |
| LifeSciBench | 60.3% | Not disclosed | Not disclosed |
| Database Migration Tasks (OpenAI internal) | 63.9% | Not disclosed | Not disclosed |
Two patterns stand out from OpenAI’s own numbers. First, the gains over Sol are concentrated in tool-heavy, agentic work: computer use, terminal tasks, cybersecurity, and software engineering. The biggest single jump is OSWorld 2.0, where Astra not only scores higher but also finishes tasks in roughly half the time.
Second, Astra does not lead on everything, even on its own page. BrowseComp goes to Sol. AutomationBench goes to Sol. And HLE with tools goes to Fable 5.1 at 65% versus Astra’s 57%. These are the gaps OpenAI chose not to highlight.
The harness caveat
OpenAI’s most headline-friendly number is 99.9% on ARC-AGI-3. That result required OpenAI’s internal Provider Adapter harness, which preserves opaque intermediate reasoning states between steps. Under the standardized ARC harness, Astra scores 62.7%. On the Coding Agent Index, Astra reaches 67 points in the Codex harness, tying Fable 5 and Claude Opus 5, but trailing Fable 5.1 at 70 points in Claude Code.
Third-party testing evaluates models across vendor-neutral agent loops, whereas first-party announcements benchmark the model inside its native, highly tuned scaffolding. Both numbers are real, but they answer different questions: one measures the model alone, the other measures the model plus its infrastructure.
What Independent Testing Shows
The Artificial Analysis Intelligence Index scores GPT-6 Astra (max) at 61 points, which combines GDPval-AA v2, Tau-3 Banking, Terminal-Bench v2.1, SciCode, Humanity’s Last Exam, GPQA Diamond, CritPt, AA-Omniscience, and AA-LCR. That ranks eighth out of 202 models.
That 61 ties several familiar names:
- GPT-5.6 Sol (max) at 61
- Grok 4.6 (high) at 61
- Muse Spark 1.3 (xhigh) at 61, with Spark pulling ahead to 62 only in unreleased max mode
- Behind Claude Fable 5.1 (max) at 66 and Claude Opus 5 (max/xhigh) at 63
- Slightly above GLM-5.3 and GLM-5.3 Flash, the former Ox Alpha stealth model turned ultra-cheap coding alternative
- Ahead of Kimi K3, Gemini 3.8 Flash, Qwen 3.8 Max, DeepSeek V4 Pro, and MiniMax M3
For a deeper look at Anthropic’s side of that race, see our coverage of Claude Fable 5.1 and Mythos 5.1.
Independent Scores: Artificial Analysis Results Compared
Vendor harnesses differ, so this second table uses Artificial Analysis standardized runs with reasoning levels shown. This is the fairest like-for-like comparison available. Astra’s max, xhigh, and high variants score 61, 61, and 60 respectively.
| Benchmark (Artificial Analysis) | Astra | Sol (max) | Opus 5 (max) | Fable 5.1 (max) | Spark 1.3 | Grok 4.6 |
|---|---|---|---|---|---|---|
| Intelligence Index | 61 (max) | 61 | 63 | 66 | 62 (max), 61 (xhigh) | 61 (high) |
| Humanity’s Last Exam | 54.7% (max) | 49.5% | 54.9% | 59.1% | 49.1% (max), 47.5% (xhigh) | 42.9% (high) |
| GPQA Diamond | 96.3% (xhigh), 96.1% (max) | 94.1% | 93.2% | 93.7% | 94.1% (xhigh), 93.8% (max) | 94.9% (high) |
| Terminal-Bench v2.1 | 89.9% (high), 88.4% (max) | 88.8% | 86.7% | 91.4% | 85.8% (max), 85.4% (xhigh) | 88.4% (high) |
| SciCode | 54.1% (max) | 56.1% | 55.7% | 62.0% | 58.6% (xhigh), 57.3% (max) | 53.6% (high) |
| GDPval-AA v2 | 1629 (max) | 1710 | 1824 | 1853 | 1709 (xhigh) | 1730 (high) |
| DeepSWE v1.1 (DataCurve board) | 74.1% OpenAI-reported | 70.8% OpenAI-reported | 68.8% OpenAI-reported | 67.4% OpenAI-reported | 75.4% Meta-reported (max) | 65.9% xAI-reported |
| MMMU-Pro | 87% (max), 86% (high/xhigh) | 83% | 85% | 90.6% (Vals.ai) | 82% (xhigh) | No score |
| ARC-AGI-1 (ARC Prize) | 98.5% (xhigh/high), 97.5% (max) | 97.5% (xhigh), 97.0% (high), 96.5% (max) | 97.5% (max/high) | 97.5% (max), 96.5% (xhigh) | Not listed | 87.0% (xhigh/high) |
| ARC-AGI-2 (ARC Prize) | 95.0% (max), 93.3% (xhigh) | 92.5% (max), 90.0% (xhigh) | 90.4% (max), 88.3% (high) | 90.0% (max/xhigh) | Not listed | 67.1% (xhigh), 65.1% (high) |
| ARC-AGI-3 standardized harness | 62.7% (max), 59.3% (xhigh), 54.8% (high) vs 99.9% (high) Provider Adapter | 7.8% (max) | 30.2% (high) | Not listed | Not listed | 2.1% (xhigh) |
Where Astra Truly Leads: MMMU-Pro, Computer Use, and Cyber
Astra does lead outright on a handful of standardized benchmarks. On Artificial Analysis runs, it posts the highest MMMU-Pro score of any model tested (87% at max, 86% at high and xhigh), the highest GPQA Diamond score (96.3% at xhigh), and a Terminal-Bench v2.1 figure that trails only Fable 5.1.
But the biggest practical advantage is not a benchmark score. It is the combination of higher accuracy and much lower task time on OSWorld 2.0, the computer-use benchmark that measures how well a model navigates real desktop environments. OpenAI reports Astra completes tasks in roughly 40 minutes, versus 75 minutes for Sol. That is not just a capability jump. It is a cost reduction, because fewer tokens and fewer minutes mean a lower bill per task.
OpenAI also claims Astra is the first model to reach the Critical threshold for cybersecurity under its own Preparedness Framework, which is why Daybreak access starts with trusted organizations first.
New Capability: Multitasking and Delegated Computer Use
The most notable new functionality is not a benchmark but a behavior. In demo videos, Astra can keep working on a main task while assigning subtasks to other models, then reconcile the results. OpenAI also says Astra can keep notes across context windows and search earlier messages and tool output instead of compressing everything into one summary.
Demo performance looks strong, but independent evaluations of multi-agent reliability, error propagation, and cost when spawning subagents have not yet been published.
Price and Token Efficiency: Same as Fable, Pricier Than Opus
API pricing is $10 per million input tokens, $1 per million cached input, and $50 per million output, with a $12.50 cache-write rate. Fast mode doubles all rates to $20 and $100. Requests above roughly 272,000 input tokens are billed at higher rates for the full request, not just the excess.
Here is the correction to an easy misunderstanding: Astra is not cheaper than Claude. It matches Claude Fable 5.1 exactly at $10 and $50, but with more expensive cache reads ($1.00 versus Fable’s $0.25). Claude Opus 5 undercuts Astra at $5 and $25 while scoring higher on the Intelligence Index. GPT-5.6 Sol’s standard rate is also cheaper at $5 and $30.
| Model | Input per 1M | Cached input per 1M | Output per 1M | Intelligence Index |
|---|---|---|---|---|
| GPT-6 Astra standard | $10.00 | $1.00 | $50.00 | 61 |
| GPT-6 Astra Fast | $20.00 | $2.00 | $100.00 | 61 |
| Claude Fable 5.1 | $10.00 | $0.25 | $50.00 | 66 |
| Claude Opus 5 | $5.00 | $0.50 | $25.00 | 63 |
| GPT-5.6 Sol standard | $5.00 | $0.50 | $30.00 | 61 |
| GPT-5.6 Sol promo | $4.00 | $0.40 | $20.00 | 61 |
| Muse Spark 1.3 standard | $1.25 | $0.15 | $4.25 | 61 |
| Muse Spark 1.3 contributor | $0.10 | $0.002 | $0.20 | 61 |
| Grok 4.6 | $2.00 | $0.50 | $6.00 | 61 |
| GLM-5.3 | $1.40 | $0.26 | $4.40 | 60 |
| Gemini 3.8 Flash intro | $0.75 | $0.075 | $3.75 | 58 |
| Gemini 3.8 Flash standard | $1.50 | $0.15 | $7.50 | 58 |
| Kimi K3 | $3.00 | $0.30 | $15.00 | 57 |
For current list prices across providers, see our LLM API pricing comparison.
Speed
Artificial Analysis lists Astra output speed as not yet available. OpenAI reports substantially faster task completion than Sol, and early throughput figures suggest Astra runs at roughly twice the tokens per second of Opus, though still slower than Muse Spark 1.3. For latency-sensitive coding work, our breakdown of Meta Muse Code as a cheaper Claude Code alternative covers the fastest current option.
Should You Choose Astra Over Claude, Grok, or Muse Spark?
Choose Astra if you need near-Opus intelligence with strong computer use, visual reasoning, and cyber capabilities, plus faster task completion than Sol. Its advantage over Claude is speed and token efficiency, not price, since Opus undercuts it and Fable matches it. For how it now stacks up against Anthropic’s newest mid-tier model on a shared benchmark set, see our Claude Sonnet 5.5 comparison.
Choose an alternative if unit price dominates. Muse Spark 1.3 standard at $1.25 and $4.25, or $0.10 and $0.20 on the contributor tier, beats Astra decisively at the same Intelligence Index level. Grok 4.6 at $2 and $6 also matches Astra’s 61 points for far less. Chinese models such as Qwen 3.8 and GLM-5.3 at $1.40 and $4.40 also undercut Astra, a pattern we explain in why Chinese AI models are so much cheaper than OpenAI and Anthropic.
Frequently Asked Questions
When was GPT-6 Astra released?
September 3, 2026, as the max reasoning variant. It arrived two days after Gemini 3.8 Flash, three days after Claude Fable 5.1 and Mythos 5.1, and one day after Muse Spark 1.3.
How intelligent is GPT-6 Astra?
It scores 61 on the Artificial Analysis Intelligence Index, eighth of 202 models, tied with GPT-5.6 Sol (max), Grok 4.6 (high), and Muse Spark 1.3 (xhigh). It sits below Opus 5 (63) and Fable 5.1 (66).
How much does GPT-6 Astra cost?
$10 per million input tokens, $1 cached, and $50 output in standard mode, or $20 and $100 in Fast mode. That matches Fable 5.1 and costs twice as much as Opus 5.
Is GPT-6 Astra the best model in the world?
By OpenAI’s selected benchmarks, it leads on MMMU-Pro, ARC-AGI-3, FrontierMath, ExploitBench, and OSWorld 2.0. By independent average intelligence, it ties for eighth. Its 99.9% ARC-AGI-3 uses OpenAI’s Provider Adapter harness; under the standardized ARC harness, the score is 62.7%.
What is new in GPT-6 Astra besides benchmarks?
Multitasking computer use with delegation to other models, persistent notes across context windows, searchable tool history, a 1.05 million token context window, and Critical-level cybersecurity capability under OpenAI’s framework.
Bottom Line
GPT-6 Astra is a strong flagship for buyers who value top-tier agentic performance, computer use, and token-efficient task completion at roughly twice the speed of Opus. It does not dethrone Claude on independent average intelligence, it matches Fable on price while Opus stays cheaper with a higher score, and it is far from the cheapest 61-point model. Muse Spark 1.3, Grok 4.6, GLM-5.3, and Gemini 3.8 Flash all undercut it by wide margins.
Astra’s real pitch is capability per task, not price per token. For the cheaper tiers, see our GPT-6 Sol and Luna benchmarks and pricing breakdown, where OpenAI halved prices with mostly flat scores.
