OpenAI
GPT-6 Astra
S-tier for end-to-end agent work, backed by a complete set of 16 published Superbash visual benchmark runs.
API pricing
USD per million tokens · OpenAI API Standard (≤272K input tokens)- Input
- $10
- Output
- $50
- Cached input
- $1
- Context window
- 1.05M
Try the visual runs
Open the generated scene to test it, or compare that prompt with another model.
City Scroll Journey
Cinematic scroll journey testing AI-generated scene continuity, scroll-scrubbed camera motion, and art-directed landing craft.
Ember Glider
Sunset gliding journey testing flight energy management, checkpoint flow, and atmospheric scene craft.
Helm's Deep
Fortress siege scene testing scale, lighting, architecture, and cinematic atmosphere.
Hogwarts Broom Flight Simulator
Broom-flight scene testing depth, motion cues, castle scale, and fantasy mood.
Jabberwock
Dark fantasy encounter testing creature design, forest mood, and narrative staging.
Low-Poly Tower Defense
Diorama tower defense testing economy balance, wave design, placement rules, and combat readability.
Low Poly World
Stylized island build testing composition, color, and low-poly worldbuilding.
Mechanical Watch Simulator
Interactive watch movement testing mechanical legibility, accurate relative motion, and real-time 3D controls.
Neon Drift
Synthwave time-trial racing testing drift physics, lap timing, ghost replay, and unlock progression.
Office Life
Workplace vignette testing everyday scene logic, objects, and believable office detail.
Petri Dish
Microscopic ecosystem testing organic forms, scientific clarity, and cellular detail.
Starfall Arena
Neon arena survival testing wave escalation, upgrade builds, particle feedback, and boss design.
Stormwind Trebuchet Simulator
Counterweight siege simulation testing coupled mechanics, trajectory prediction, projectile cameras, and interactive tuning.
Universe Simulator
Cosmic system testing orbital structure, glowing bodies, scale, and simulation readability.
Vice City
Neon coastal city testing vehicles, architecture, atmosphere, and dense urban layout.
Yingzao Fashi Assembly
Timber assembly scene testing structure, joinery, construction order, and material clarity.
How GPT-6 Astra compares
Three relevant peers, side by side.
- GPT-6 Astra This model
- Claude Fable 5.1
- GPT-5.6 Sol
- Claude Opus 5
6 benchmarks · 4 models
Official benchmark profile
Astra’s published benchmark results
Full official benchmark table40 scores and peer comparisons
| Benchmark | Area | Score | Comparison |
|---|---|---|---|
| Terminal-Bench 4.0 | Coding | 57.9% | |
| Terminal-Bench Science 0.1 | Science | 64.6% | |
| AutomationBench | Automation | 41.4% | |
| Humanity’s Last Exam (with tools) | Reasoning | 57.2% | |
| Artificial Analysis Intelligence Index v4.1.1 | Reasoning | 61.2 | |
| FrontierMath Tier 4 (v2) | Reasoning | 97.6% | |
| DeepSWE v1.1 | Coding | 74.1% | |
| FrontierCode 1.1 Extended (score) | Coding | 64.5% | |
| FrontierCode 1.1 Main (score) | Coding | 53.3% | |
| GPQA Diamond | Science | 96.0% | |
| ARC-AGI-2 | Reasoning | 95.0% | |
| ARC-AGI-1 | Reasoning | 98.5% | |
| Agents’ Last Exam | Computer use | 59.3% | |
| OSWorld 2.0 · offline, v2026.08.08 | Computer use | 72.6% | |
| ScreenSpot-Pro (no tools) | Computer use | 92.7% | |
| BenchCAD (with tools) | Professional work | 95.9% | |
| BrowseComp | Research | 91.5% | |
| OpenScore String Quartets (1 − OMR-NED) | Professional work | 0.84 | |
| Design tasks (OpenAI internal) | Professional work | 50.0% | |
| Data science tasks (OpenAI internal) | Science | 40.9% | |
| Database migration (OpenAI internal) | Coding | 63.9% | |
| Artificial Analysis Coding Agent Index v1.4 | Coding | 67.0 | |
| GeneBench Pro | Science | 37.1% | |
| MedChemBench (OpenAI internal) | Science | 49.3% | |
| LifeSciBench | Science | 60.3% | |
| HealthBench Professional (length-adjusted) | Science | 63.4% | |
| ExploitBench | Cybersecurity | 100.0% | |
| ExploitGym | Cybersecurity | 42.4% | |
| ExploitBench (June–August 2026) | Cybersecurity | 39.0% | |
| SRE-Bench (single attempt) | Cybersecurity | 88.0% | |
| SEC-Bench Pro | Cybersecurity | 85.4% | |
| Computer-use safety (OpenAI internal)Lower is better | Alignment | 2.4% | |
| Computer-use safety with AutoReviewLower is better | Alignment | 1.8% | |
| Circumvention (OpenAI internal)Lower is better | Alignment | 0.0% | |
| ExploitGym honeypotLower is better | Alignment | 0.0% | |
| Impossible ExploitGym | Alignment | 100.0% | |
| Hallucination (OpenAI internal)Lower is better | Alignment | 4.2% | |
| MRCR v2 · 8 needles · 256K–512K | Long context | 100.0% | |
| MRCR v2 · 8 needles · 512K–1M | Long context | 96.3% | |
| ARC-AGI-3 | Reasoning | 99.9% |
Sources, pricing details and methodology
Verified model facts
Model specifications
- Status
- Current
- API model ID
gpt-6-astra- Context
- 1.05M
- Max output
- 128K
- License
- Not publicly verified
- Access route
- OpenAI API; rolling out through ChatGPT and Codex, Microsoft Azure, and AWS Bedrock. Access depends on account and rollout. The September 7 visual benchmarks used Codex subagents; API prices are not the recorded cost of those runs.
- Modalities
- text input, image input, text output
- Hardware
- Not publicly verified
- OpenAI API Standard (≤272K input tokens) · USD / 1M tokens
- $10 input · $1 cached input · $12.5 cache write · $50 output
Above 272K input tokens, the full request costs $20 input, $2 cached input, $25 cache write, and $75 output per 1M tokens. Batch, Flex, Fast mode, and regional pricing differ.
Benchmark sources
Maximum reported score at any effort; research/API environments. A shared table does not guarantee identical harnesses, budgets, or safeguards.
OpenAI Astra launch comparison tableLocal visual runs and their validation records are separate from these provider-published scores.
Evaluation settings
- Maximum reported score at any effort; research/API environments. A shared table does not guarantee identical harnesses, budgets, or safeguards.
- Lower is better. Maximum reported score at any effort; research/API environments. A shared table does not guarantee identical harnesses, budgets, or safeguards.
- Terminal-Bench Science 0.1
- Anthropic reports standard error of about 3.5–4.5 points per model; small gaps may be noise.
- Artificial Analysis Intelligence Index v4.1.1
- Independent index as reproduced in OpenAI’s launch table; separate from the older v4.1 observations in the cost explorer.
- FrontierCode 1.1 Extended (score)
- Astra uses additional developer guidance about keeping code and tests scoped (source footnote 8).
- FrontierCode 1.1 Main (score)
- Astra uses additional developer guidance about keeping code and tests scoped (source footnote 8).
- OSWorld 2.0 · offline, v2026.08.08
- Partial score on the offline subset. Anthropic’s modified tasks and grading are incompatible; no Fable 5.1 value is supplied here.
- ScreenSpot-Pro (no tools)
- The source labels Mythos’s 87.3 as Fable 5 (footnote 17). It is excluded from the Fable column here because the model’s safeguards differ.
- BenchCAD (with tools)
- Claude results use three evaluation modifications. Scores shown for context; no direct gap is calculated.
- HealthBench Professional (length-adjusted)
- OpenAI re-evaluation with GPT-5.4 grading. Fable 5.1 uses Opus 5 fallback for refusals; not a standalone Fable result.
- ExploitBench
- Astra and Sol evaluated without production safeguards.
- ExploitGym
- The source’s Fable columns are Mythos 5.1 (30.4) and Mythos 5 (28.4), excluded here. Astra/Sol have no six-hour limit and no production safeguards. Not a production-model comparison.
- ExploitBench (June–August 2026)
- OpenAI internal set, without production safeguards. Sol’s 5.5 is affected by a 300-turn cap; source also reports 11.5 with fewer limits.
- SRE-Bench (single attempt)
- Cyber capability evaluation without production safeguards.
- Computer-use safety (OpenAI internal)
- Lower is better. Research harness without production confirmations; provider safeguards and tools still differ.
- ARC-AGI-3
- OpenAI Responses harness with two modified settings; see source footnote 1.
Comparison methodology
Maximum reported score at any effort; research/API environments. A shared table does not guarantee identical harnesses, budgets, or safeguards. Gaps describe reported scores, not statistical significance. Percentage bars run from 0 to 100; index bars use a 0–100 display scale.
- Terminal-Bench Science 0.1
- Anthropic reports standard error of about 3.5–4.5 points per model; small gaps may be noise.
- Artificial Analysis Intelligence Index v4.1.1
- Independent index as reproduced in OpenAI’s launch table; separate from the older v4.1 observations in the cost explorer.
- FrontierCode 1.1 Extended (score)
- Astra uses additional developer guidance about keeping code and tests scoped (source footnote 8).
- FrontierCode 1.1 Main (score)
- Astra uses additional developer guidance about keeping code and tests scoped (source footnote 8).
- OSWorld 2.0 · offline, v2026.08.08
- Partial score on the offline subset. Anthropic’s modified tasks and grading are incompatible; no Fable 5.1 value is supplied here.
- ScreenSpot-Pro (no tools)
- The source labels Mythos’s 87.3 as Fable 5 (footnote 17). It is excluded from the Fable column here because the model’s safeguards differ.
- BenchCAD (with tools)
- Claude results use three evaluation modifications. Scores shown for context; no direct gap is calculated.
- HealthBench Professional (length-adjusted)
- OpenAI re-evaluation with GPT-5.4 grading. Fable 5.1 uses Opus 5 fallback for refusals; not a standalone Fable result.
- ExploitBench
- Astra and Sol evaluated without production safeguards.
- ExploitGym
- The source’s Fable columns are Mythos 5.1 (30.4) and Mythos 5 (28.4), excluded here. Astra/Sol have no six-hour limit and no production safeguards. Not a production-model comparison.
- ExploitBench (June–August 2026)
- OpenAI internal set, without production safeguards. Sol’s 5.5 is affected by a 300-turn cap; source also reports 11.5 with fewer limits.
- SRE-Bench (single attempt)
- Cyber capability evaluation without production safeguards.
- Computer-use safety (OpenAI internal)
- Lower is better. Research harness without production confirmations; provider safeguards and tools still differ.
- ARC-AGI-3
- OpenAI Responses harness with two modified settings; see source footnote 1.
Editorial verdict
Where this model fits
GPT-6 Astra completed all 16 Superbash visual benchmark briefs in independent prompt-only runs. Every published artifact is linked from the benchmark page with its manifest and validation record. S tier reflects the completed, inspectable run set and the editorial assessment of Astra as the strongest current choice for difficult multi-step product work; it is not a numeric leaderboard claim.
Best for
Watch out
Why it is ranked here
Evidence
Superbash editorial model ranking
Takeaway: GPT-6 Astra is currently placed in Tier S.
The September 2026 editorial roster places GPT-6 Astra at rank 1.
Open source →Superbash visual benchmark suite: GPT-6 Astra
Takeaway: All 16 prompt-only Astra runs are published with manifests, artifact hashes, and validation records.
The methodology record describes how the local runs were isolated and validated; it is evidence of this site’s visual suite, not a provider benchmark or independent universal score.
Open source →Related guides