Max (With Tools) · 14 ranked
GPT-6.2 Astra
OpenAI GPT-6.2 Astra
OpenAI released GPT-6.2 Astra on November 17, 2026 for complex professional work, scientific reasoning, and demanding software projects.
At a glance
- Context window
- 1.05M tokens
- Max output
- 128K tokens
- Text input / output · per 1M tokens
- $8.00 / $40.00
- Knowledge cutoff
- 2026-05-18
- GPQA Diamond
- #1 of 200 models on the GPQA Diamond leaderboard
Benchmark snapshot
Best result in 6 of 10 categories
AcademicGPQA Diamond
97.0Thinking (No Tools) · #1 of 200
CodingFrontierCode 2.0
49.0Extra-High (With Tools) · #2 of 36
General EvaluationECI
174Reported best (effort unspecified) · #1 of 168
Source: Epoch AI (epoch.ai)MultimodalMMMU-Pro
89.5Max (No Tools) · #3 of 136
Source: Artificial Analysis (artificialanalysis.ai)ProfessionalAA-AnalystAgent
62.5Max (With Tools + Internet) · #1 of 39
Source: Artificial Analysis (artificialanalysis.ai)ReasoningARC-AGI-1
98.5Thinking Low (No Tools) · #1 of 87
Benchmark details
34 results · Best published score per benchmark, grouped by category
| Benchmark · setting | Score | Rank |
|---|---|---|
| Academic | ||
| FrontierMath Erdős Max (With Tools) · Mathematics Source: Epoch AI (epoch.ai) | 10.3 | 1 / 14 |
| FrontierMath Tier 4 v2 Max (With Tools) · Mathematics Source: Epoch AI (epoch.ai) | 100.0 | 1 / 73 |
| FrontierMath v2 Max (With Tools) · Mathematics Source: Epoch AI (epoch.ai) | 92.6 | 4 / 85 |
| GPQA Diamond Thinking (No Tools) · Scientific Reasoning | 97.0 | 1 / 200 |
| HLE-Diamond Max (No Tools) · Knowledge Exams | 72.2 | 2 / 16 |
| MysteryMechanism Max (With Tools) · Scientific Reasoning Source: Vals AI. © 2026 Vals AI. All rights reserved. No affiliation or endorsement. | 53.2 | 1 / 32 |
| Terminal-Bench-Science 0.1 Extra-High (With Tools) · Scientific Computing Source: Artificial Analysis (artificialanalysis.ai) | 78.6 | 1 / 32 |
| Coding | ||
| FrontierCode 2.0 Extra-High (With Tools) · Repository Engineering | 49.0 | 2 / 36 |
| Terminal-Bench 4.0 Extra-High (With Tools) · Agentic Development Source: Artificial Analysis (artificialanalysis.ai) | 64.1 | 5 / 90 |
| Vibe Code Bench v1.1 Max (With Tools) · Code Generation & Editing Source: Vals AI — Vibe Code Bench v1.1. © 2026 Vals AI. All rights reserved. No affiliation or endorsement. | 89.9 | 9 / 121 |
| Computer Use | ||
| AutomationBench-AA Max (With Tools) · limited comparability · Tool Orchestration Source: Artificial Analysis (artificialanalysis.ai) | 77.2 | 4 / 18 |
| General Evaluation | ||
| AA Intelligence Index v5.0.1 Max (task-specific tools) · Capability Indices | 62.1 | 2 / 62 |
| Agent Arena Index Max (With Tools + Internet) · Preference Arenas Source: Arena Agent Arena Index | 1291 | 6 / 36 |
| ECI Reported best (effort unspecified) · Capability Indices Source: Epoch AI (epoch.ai) | 174 | 1 / 168 |
| Vals Index v2.2 Max (With Tools) · Capability Indices Source: Vals AI — Vals Index v2.2. © 2026 Vals AI. All rights reserved. No affiliation or endorsement. | 67.7 | 7 / 36 |
| Vals RSI Index Max (task-specific tools) · Capability Indices Source: Vals AI | 29 | 3 / 3 |
| Long Context | ||
| EBR-bench Max (No Tools) · Memory & Persistence Source: Epoch AI (epoch.ai) | 92.4 | 1 / 30 |
| Multimodal | ||
| GDP.pdf Max (No Tools) · Documents & Charts Source: Artificial Analysis (artificialanalysis.ai) | 29.8 | 9 / 96 |
| MMMU-Pro Max (No Tools) · Visual Understanding Source: Artificial Analysis (artificialanalysis.ai) | 89.5 | 3 / 136 |
| Professional | ||
| AA-AnalystAgent Max (With Tools + Internet) · Data Analysis Source: Artificial Analysis (artificialanalysis.ai) | 62.5 | 1 / 39 |
| BenchCAD 2.1 Core Max (With Tools) · Design & Engineering | 82.5 | 2 / 7 |
| EMB (Excel Modeling) Max (With Tools) · Finance Source: Vals AI — EMB (Excel Modeling). © 2026 Vals AI. All rights reserved. No affiliation or endorsement. | 76.0 | 3 / 78 |
| Harvey LAB-AA v1.1 (All-pass) Max (With Tools) · Legal Source: Artificial Analysis (artificialanalysis.ai) | 15.0 | 1 / 12 |
| Harvey's Legal Agent Benchmark Max (With Tools) · Legal Source: Vals AI — Harvey's Legal Agent Benchmark. © 2026 Vals AI. All rights reserved. No affiliation or endorsement. | 16.3 | 5 / 52 |
| Legal Research Bench Max (With Tools + Internet) · Legal Source: Vals AI — Legal Research Bench. © 2026 Vals AI. All rights reserved. No affiliation or endorsement. | 42.8 | 29 / 91 |
| Tax Agent Bench Max (With Tools + Internet) · Finance | 21.8 | 52 / 74 |
| Reasoning | ||
| ARC-AGI-1 Thinking Low (No Tools) · Abstract Generalization | 98.5 | 1 / 87 |
| ARC-AGI-2 Extra-High (No Tools) · Abstract Generalization | 95.8 | 1 / 80 |
| ARC-AGI-3 Extra-High (No Tools) · Provider Adapter · limited comparability · Abstract Generalization | 100.0 | 1 / 15 |
| ARC-AGI-3 Max (No Tools) · Standard · Abstract Generalization | 70.3 | 2 / 26 |
| SimpleBench Reported best (effort unspecified) · Commonsense Source: SimpleBench | 77.3 | 16 / 106 |
| Safety & Alignment | ||
| SimpleQA Verified Max (No Tools) · Honesty & Factuality Source: Epoch AI (epoch.ai) | 78.5 | 2 / 5 |
| Science & Health | ||
| BioMysteryBench Max (With Tools) · Biology & Genomics Source: Vals AI — BioMysteryBench. © 2026 Vals AI. All rights reserved. No affiliation or endorsement. | 80.7 | 5 / 28 |
| HealthBench Professional Max (No Tools) · length-adjusted · limited comparability · Clinical Workflows | 66.3 | 3 / 19 |
Rank is this model's position among every model with a published result on that leaderboard; the bar shows how far up the field that is. Orange marks first place on leaderboards with at least 5 models. Unranked leaderboards publish a composite or direction-less score.
Back to benchmark snapshot ↑API price vs benchmark score
Methodology & notes
Uses listed uncached text API rates at a 4:1 input/output mix against one published protocol and setting (arc3-semi-private-standard). API rates are separate from evaluation token costs and do not measure cost per solved task. Selected point basis: Standard pricing. Up to 272,000 input tokens Official OpenAI per-million-token pricing. Full request uses the applicable input-token band.
API pricing
Listed text rates, USD per million tokens · effective 2026-11-17
- Input
- $8.00
- Output
- $40.00
- Cached input
- $0.40
Blended price4:1 input:output · 90% of input tokens cached · same basis as the price ranking
$8.928/ 1M tokens
Specifications
- Provider
- OpenAI
- Model type
- Reasoning model
- Released
- November 17, 2026
- Architecture
- Undisclosed
- Context window
- 1.05M tokens
- Max output
- 128K tokens
- Knowledge cutoff
- 2026-05-18
- Input
- Text, Image
- Output
- Text
- Weights
- Closed / API only
- Commercial use
- Proprietary
Capabilities
Model analysis
OpenAI released GPT-6.2 Astra on November 17, 2026 for complex professional work, scientific reasoning, and demanding software projects. OpenAI launch announcement
Task difficulty and cost shape model selection
Vals measured Astra at 68.44% on its Science Index, up from 65.09% for GPT-6 Astra, while average test cost fell from USD 8.03 to USD 4.76. Evaluation source
In the independent Artificial Analysis Intelligence Index v5.0.1, max effort scored about 62, at 10.7 minutes and USD 4.00 per task. For repeatable workloads, Sol offers lower rates. Astra is an option where better deliverables or fewer revisions justify higher cost. Independent model results
Coding and engineering benchmarks measure different outcomes
On Cognition’s FrontierCode 2.0, both models ran at max effort in Codex. Astra scored about 44.2% and Sol about 43.8%. On BenchCAD 2.1 Core with a customized Codex harness, Astra scored 82.5 at USD 5.43 per case and Sol scored 73.1 at USD 1.60 per case. FrontierCode · BenchCAD 2.1
In the independent Artificial Analysis Coding Agent Index with Codex at max effort, Astra is level with GPT-6 Astra and Sol trails GPT-6.1 Sol by three points, driven mainly by repository question answering. On the Vals Science Index, Astra scored 68.44% and Sol 57.76%, at USD 4.76 and USD 1.12 per test, respectively. Artificial Analysis release tests · Vals Science Index
The API supports a 1.05 million token context
gpt-6.2-astra accepts text and images and returns text. Maximum input and output are 922,000 and 128,000 tokens. The knowledge cutoff is 2026-05-18. Responses, Chat Completions, and Batch are supported, with function calling, structured outputs, file and web search, and computer use. The Responses API offers multi-agent and discounted background completion features in beta. OpenAI has not disclosed parameter counts or a language inventory. Official model documentation
Long requests and faster tiers change the bill
Standard short-context rates per million tokens are USD 8 input, 0.4 cached input, and 40 output. Above 272,000 input tokens, input and cache rates double and output costs 1.5 times the short-context rate for the full request. Batch is half the Standard price and Fast is twice the Standard price. Ultrafast costs three times Standard through February 28, 2027 and six times Standard from March 1, Pacific time. ChatGPT Work and Codex serve Plus, Pro, Enterprise, and Edu users, with Ultrafast access depending on plan and workspace settings. Official API pricing
Pro 500 supports Ultrafast on both models. Pro 200 supports it on Sol at launch. Enterprise and Edu access depends on workspace settings. ChatGPT availability details
Tool calling uses the Responses API. Eligible background requests can choose a two-hour completion window at 30% below Standard rates or an eight-hour window at 40% below Standard. OpenAI API changelog
Amazon Bedrock and Microsoft Foundry also announced general availability for both base models on launch day. Region and deployment options vary by platform. AWS announcement · Microsoft Foundry model documentation
OpenAI tightened safeguards for higher-risk requests
The system card describes deployment safeguards and remaining limitations in how the models report blockers. GPT-6.2 system card
Last reviewed 2026-11-18
GPT-6.2 Astra vs competitors
Comparable results only · record is GPT-6.2 Astra's wins–losses
| Benchmark | GPT-6.2 AstraThis model | vs Claude Fable 5.2Record 9–17–2 |
|---|---|---|
| GPQA Diamond | 97.0 | 93.9 |
| ECI | 174 | 174 |
| Vibe Code Bench v1.1 | 89.9 | 93.8 |
| SimpleBench | 77.3 | 94.9 |
| GDP.pdf | 29.8 | 26.8 |
| Legal Research Bench | 42.8 | 46.2 |
| Terminal-Bench 4.0 | 64.1 | 67.6 |
| ARC-AGI-1 | 98.5 | 98.5 |
| FrontierMath v2 | 92.6 | 95.1 |
| EMB (Excel Modeling) | 76.0 | 77.4 |
Lineage
Predecessors and successors · oldest to newest
- GPT-6 AstraSep 3, 2026
Generational jump vs GPT-6 Astra
Showing the 2 largest gains and 2 smallest gains or drops out of 16 like-for-like benchmarks.
Max (With Tools) · 49 ranked
Max (With Tools) · 31 ranked
Max (With Tools) · 82 ranked
Price & benchmark trends
Across curated predecessors and successors, ordered by release date. Points use listed rates and published results, not a history of price changes within one model. Related branches may appear; adjacent points do not imply direct succession. Undated models appear last.
API price across generations
Input and output share one linear USD axis · uncached text rates · lower is cheaper. Retired models may retain their last recorded price. Rate tiers and context limits can differ; inspect each model's pricing conditions.
Benchmark performance across generations
ECI score (capability index, higher is better) · Higher is better. Same recorded metric, protocol and setting; review exclusions are respected. Score gaps do not establish statistical significance.
Inspect ECI →Orange marks this model. Hover, tap or focus the chart to see exact values. Missing values leave gaps. Horizontal spacing shows release order, not elapsed time.
| Model · released | Input / 1M | Output / 1M | Price effective | ECI |
|---|---|---|---|---|
| 1GPT-6 AstraSep 3, 2026 | $10.00 | $50.00 | 2026-09-03 | 167 ↗ |
| 2GPT-6.2 AstraNov 17, 2026 · This model | $8.00 | $40.00 | 2026-11-17 | 174 ↗ |
Primary sources
- GPT-6.2 Astra technical reportdeploymentsafety.openai.com
Community discussion
Share the API, setup, and tradeoffs you actually tested. Rate the model once; comment as often as you like.
Used this model? Sign in with Google to rate it and share what worked, what did not, and which API or setup you tested.
No community comments yet.
