Skip to content
Models / OpenAI

GPT-6 Astra

OpenAI · Model release

Released

Catalog checked 2026-09-28

Input $/1M
$10.00
Output $/1M
$50.00
Context
1.05M
Max output
128K
Reasoning
Yes
Included in
11 plans
API priceUSD / 1M tokens
Published API rates in US dollars per million tokens
RateInputOutputCache readCache write
StandardInput$10.00Output$50.00Cache read$1.00Cache write$12.50
Above 272K input tokens (priced at 2x input and cache rates, 1.5x output, for the full request)Input$20.00Output$75.00Cache read$2.00Cache write$25.00
Reasoning tokens
Billed at the output token rate

Published rate checked 2026-09-23. Sources below.

Processing tiers

The same model, processed on a different service tier. Rates for prompts up to the long-context threshold; speed claims are the provider’s and are not modelled.

API rates by processing tier, US dollars per million tokens
TierInputCache readOutputStatus
StandardInput$10.00Cache read$1.00Output$50.00StatusAvailable
BatchInput$5.00Cache read$0.50Output$25.00StatusAvailable
FlexInput$5.00Cache read$0.50Output$25.00StatusAvailable
FastInput$20.00Cache read$2.00Output$100.00StatusAvailable
UltrafastInput$60.00Cache read$6.00Output$300.00StatusAvailable

Current published base rates. Context tiers, cache-write duration and other conditions may change the rate for a request. These are not a reconstructed historical invoice.

API token estimate

Token estimate unavailable

This model has conditional rates. Aggregate token totals cannot establish which request or schedule conditions apply.

Capabilities & limitsProvider specifications
Context window
1,050,000
Maximum output
128,000
Input
text · image
Output
text
Knowledge cutoff
Apr 30, 2026
Reasoning
Supported
Tool calling
Supported
Structured output
Supported
Before you chooseChecked 2026-09-29
Usage limits
API request and token limits depend on your account usage tier. They are separate from ChatGPT and Codex subscription allowances.
Thinking
Effort: low, medium, high, xhigh, max.
Long prompts
Above 272K input tokens: 2× input and cache prices, 1.5× output prices for the full request.
Processing options
Batch and Flex: 50% of Standard token prices. Fast mode: 2× applicable token prices.

Google DeepMind · Published Sep 30, 2026

Developer reported · Gemini 4 Argon launch

Show all 17 results
Methodology & sources

Google’s launch sheet combines evaluations it computed with results it republished from leaderboards and system cards. This is one reporting set; it does not establish identical harnesses across models. Argon generally uses the Gemini API at the highest thinking setting, with a single attempt and trial averaging on smaller benchmarks.

These scores are reported by Google DeepMind. Model developers identify authorship; results may originate with another evaluator. Exact benchmark versions remain distinct. StackReplay calculates no composite score.

DeepSWE v1.1 · 74.1%

Long-horizon software engineering in real-world codebases.

Version: v1.1 · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

Google DeepMind · Published Sep 30, 2026 · Developer reported

Reporting source
Google DeepMind · Gemini 4 Argon launch
Evaluation origin
Leaderboard or system card cited by Google (republished result)
Effort
Maximum available reasoning; best available reported result when maximum-setting results are unavailable
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Argon uses Google’s mini-swe agent; Astra is reported from the leaderboard, Fable and Opus from system cards. Thinking levels follow the highest scoring level reported by Datacurve.

No confidence interval reported in this comparison.

Original methodology ↗ · Original evidence ↗

FrontierSWE v2 · 65.5%

Agentic software engineering on coding tasks.

Version: v2 · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

Google DeepMind · Published Sep 30, 2026 · Developer reported

Reporting source
Google DeepMind · Gemini 4 Argon launch
Evaluation origin
Proximal public leaderboard (republished result)
Effort
Maximum available reasoning; best available reported result when maximum-setting results are unavailable
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Results republished from Proximal’s public leaderboard.

No confidence interval reported in this comparison.

Original methodology ↗ · Original evidence ↗

Vibe Code Bench · 89.6%

Code generation for prompted software-building tasks.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

Google DeepMind · Published Sep 30, 2026 · Developer reported

Reporting source
Google DeepMind · Gemini 4 Argon launch
Evaluation origin
Vals AI public leaderboard (republished result)
Effort
Maximum available reasoning; best available reported result when maximum-setting results are unavailable
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Results republished from the Vals AI public leaderboard.

No confidence interval reported in this comparison.

Original methodology ↗ · Original evidence ↗

Terminal-Bench 4.0 · 58.2%

Agentic task completion in a terminal environment.

Version: 4.0 · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

Google DeepMind · Published Sep 30, 2026 · Developer reported

Reporting source
Google DeepMind · Gemini 4 Argon launch
Evaluation origin
Terminal-Bench public leaderboard (republished result)
Effort
Maximum available reasoning; best available reported result when maximum-setting results are unavailable
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Argon is computed by Google; competitor results are republished from the official leaderboard, using the highest scoring thinking level reported by its authors.

No confidence interval reported in this comparison.

Original methodology ↗ · Original evidence ↗

PostTrainBench v1.1 · 44.3%

Post-training of language models under a fixed compute budget.

Version: v1.1 · Benchmark-reported weighted score · percent · Higher is better · Task subset: Not separately specified

Google DeepMind · Published Sep 30, 2026 · Developer reported

Reporting source
Google DeepMind · Gemini 4 Argon launch
Evaluation origin
Google DeepMind (source-computed)
Effort
Maximum available reasoning; best available reported result when maximum-setting results are unavailable
Harness
OpenCode
Tools
Single NVIDIA H100, 10-hour budget
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Google computed every model with OpenCode and a 10-hour budget on one NVIDIA H100; the benchmark’s metric aggregates four base models and seven benchmarks.

No confidence interval reported in this comparison.

Original methodology ↗ · Original evidence ↗

Vals Index · 63.1%

Economic knowledge work across finance, coding, legal and tax tasks.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

Google DeepMind · Published Sep 30, 2026 · Developer reported

Reporting source
Google DeepMind · Gemini 4 Argon launch
Evaluation origin
Vals AI (republished result)
Effort
Maximum available reasoning; best available reported result when maximum-setting results are unavailable
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Vals AI results republished by Google.

No confidence interval reported in this comparison.

Original methodology ↗ · Original evidence ↗

AutomationBench · 41.4%

End-to-end execution of business workflows.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

Google DeepMind · Published Sep 30, 2026 · Developer reported

Reporting source
Google DeepMind · Gemini 4 Argon launch
Evaluation origin
Zapier public leaderboard (republished result)
Effort
Maximum available reasoning; best available reported result when maximum-setting results are unavailable
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Private task set; results republished from Zapier’s public leaderboard.

No confidence interval reported in this comparison.

Original methodology ↗ · Original evidence ↗

Vals Finance Agent v2 · 53.5%

Multi-step financial research by an agent.

Version: v2 · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

Google DeepMind · Published Sep 30, 2026 · Developer reported

Reporting source
Google DeepMind · Gemini 4 Argon launch
Evaluation origin
Vals AI (republished result)
Effort
Maximum available reasoning; best available reported result when maximum-setting results are unavailable
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Vals AI results republished by Google.

No confidence interval reported in this comparison.

Original methodology ↗ · Original evidence ↗

Harvey’s Legal Agent Benchmark · 5.4%

Legal research and drafting by an agent.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

Google DeepMind · Published Sep 30, 2026 · Developer reported

Reporting source
Google DeepMind · Gemini 4 Argon launch
Evaluation origin
Vals AI (republished result)
Effort
Maximum available reasoning; best available reported result when maximum-setting results are unavailable
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Vals AI results republished by Google.

No confidence interval reported in this comparison.

Original methodology ↗ · Original evidence ↗

Terminal-Bench Science 0.1 · 68.1%

Scientific and mathematical tasks in a terminal environment.

Version: 0.1 · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

Google DeepMind · Published Sep 30, 2026 · Developer reported

Reporting source
Google DeepMind · Gemini 4 Argon launch
Evaluation origin
Terminal-Bench Science public leaderboard (republished result)
Effort
Maximum available reasoning; best available reported result when maximum-setting results are unavailable
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Argon uses a sixfold verifier timeout to address verification timeouts; competitor results come from the official leaderboard.

No confidence interval reported in this comparison.

Original methodology ↗ · Original evidence ↗

LABBench 2 · 85.4%

Scientific tasks with a terminal, bioinformatics tools and internet access.

Version: 2 · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

Google DeepMind · Published Sep 30, 2026 · Developer reported

Reporting source
Google DeepMind · Gemini 4 Argon launch
Evaluation origin
Google DeepMind (source-computed)
Effort
Maximum available reasoning; best available reported result when maximum-setting results are unavailable
Harness
Not reported
Tools
Linux terminal, bioinformatics tools, Python, R and internet
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Google computed every model with a Linux terminal, bioinformatics tools, Python, R and internet access.

No confidence interval reported in this comparison.

Original methodology ↗ · Original evidence ↗

RiemannBench · 72.0%

Mathematical problem solving.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

Google DeepMind · Published Sep 30, 2026 · Developer reported

Reporting source
Google DeepMind · Gemini 4 Argon launch
Evaluation origin
Surge public leaderboard (republished result)
Effort
Maximum available reasoning; best available reported result when maximum-setting results are unavailable
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Results republished from Surge’s public leaderboard.

No confidence interval reported in this comparison.

Original methodology ↗ · Original evidence ↗

GraphWalks · up to 128k BFS F1 · 98.7%

Breadth-first graph traversal with contexts up to 128k tokens, scored by F1.

Version: Not reported · BFS F1 · percent · Higher is better · Task subset: · up to 128k BFS F1

Google DeepMind · Published Sep 30, 2026 · Developer reported

Reporting source
Google DeepMind · Gemini 4 Argon launch
Evaluation origin
Google DeepMind (source-computed)
Effort
Maximum available reasoning; best available reported result when maximum-setting results are unavailable
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Google computed every model on 650 problems with context lengths up to 128k tokens.

No confidence interval reported in this comparison.

Original methodology ↗ · Original evidence ↗

GraphWalks · 256k–1M BFS F1 · 71.8%

Breadth-first graph traversal with 256k to 1M-token contexts, scored by F1.

Version: Not reported · BFS F1 · percent · Higher is better · Task subset: · 256k–1M BFS F1

Google DeepMind · Published Sep 30, 2026 · Developer reported

Reporting source
Google DeepMind · Gemini 4 Argon launch
Evaluation origin
Google DeepMind (source-computed)
Effort
Maximum available reasoning; best available reported result when maximum-setting results are unavailable
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Google computed every model on 200 problems with context lengths from 256k to 1M tokens.

No confidence interval reported in this comparison.

Original methodology ↗ · Original evidence ↗

Chartography · 71.0%

Visual interpretation and analysis of charts.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

Google DeepMind · Published Sep 30, 2026 · Developer reported

Reporting source
Google DeepMind · Gemini 4 Argon launch
Evaluation origin
Surge public leaderboard (republished result)
Effort
Maximum available reasoning; best available reported result when maximum-setting results are unavailable
Harness
Not reported
Tools
No tools
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Without tools; results republished from Surge’s public leaderboard.

No confidence interval reported in this comparison.

Original methodology ↗ · Original evidence ↗

LVBench · 87.5%

Understanding information in long videos.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

Google DeepMind · Published Sep 30, 2026 · Developer reported

Reporting source
Google DeepMind · Gemini 4 Argon launch
Evaluation origin
Google DeepMind (source-computed)
Effort
Maximum available reasoning; best available reported result when maximum-setting results are unavailable
Harness
Not reported
Tools
No tools
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Google computed every model without tools. Video sampling differs due to API limits: Argon 1 FPS, Astra 800 frames, Fable 300 frames, Opus 600 frames.

No confidence interval reported in this comparison.

Original methodology ↗ · Original evidence ↗

CWE-bench v1 · 68.0%

Software vulnerability remediation.

Version: v1 · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

Google DeepMind · Published Sep 30, 2026 · Developer reported

Reporting source
Google DeepMind · Gemini 4 Argon launch
Evaluation origin
CWE-bench public leaderboard (republished result)
Effort
Maximum available reasoning; best available reported result when maximum-setting results are unavailable
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Pass@1 results republished from the official leaderboard. Its pass@4 tiebreak is not part of this sheet, so equal pass@1 scores remain ties.

No confidence interval reported in this comparison.

Original methodology ↗ · Original evidence ↗

OpenAI · Published Sep 29, 2026

Developer reported · GPT-6.1 Sol launch · Max effort

Methodology & sources

Launch chart results at Max reasoning effort. GPT evaluations ran in OpenAI research environments or its API. Competitor scores were taken from public reports. Exact chart values, versions, subsets and documented fallbacks are preserved.

These scores are reported by OpenAI. Model developers identify authorship; results may originate with another evaluator. Exact benchmark versions remain distinct. StackReplay calculates no composite score.

GDP.pdf · 31%

Answering professional questions about complex PDF documents.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

OpenAI · Published Sep 29, 2026 · Developer reported

Reporting source
OpenAI · GPT-6.1 Sol launch · Max effort
Evaluation origin
OpenAI research environment or API (source-computed)
Effort
Max
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
OpenAI research environment or API; deployment not individually specified
Checked
Sep 30, 2026

Exact percentage converted from the numerical fraction embedded in the official launch chart. Harness, fallback and tool details are retained only where the publication specifies them.

Original methodology ↗ · Original evidence ↗

AutomationBench 1.0.6 · 41.4%

Completing multi-step business workflows with 47 available tools.

Version: 1.0.6 · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

OpenAI · Published Sep 29, 2026 · Developer reported

Reporting source
OpenAI · GPT-6.1 Sol launch · Max effort
Evaluation origin
OpenAI research environment or API (source-computed)
Effort
Max
Harness
Not reported
Tools
47 tools across sales, marketing, operations, support, finance and HR
Fallback
Not reported
Provider
OpenAI research environment or API; deployment not individually specified
Checked
Sep 30, 2026

Exact percentage converted from the numerical fraction embedded in the official launch chart. Harness, fallback and tool details are retained only where the publication specifies them.

Original methodology ↗ · Original evidence ↗

OSWorld 2.0 Offline · v2026.08.08 · 73.49%

Computer-use workflows scored by partial reward on the August 8 offline task set.

Version: 2.0 · Partial reward · percent · Higher is better · Task subset: Offline set, v2026.08.08

OpenAI · Published Sep 29, 2026 · Developer reported

Reporting source
OpenAI · GPT-6.1 Sol launch · Max effort
Evaluation origin
OpenAI research environment or API (source-computed)
Effort
Max
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
OpenAI research environment or API; deployment not individually specified
Checked
Sep 30, 2026

Exact percentage converted from the numerical fraction embedded in the official launch chart. Harness, fallback and tool details are retained only where the publication specifies them.

Original methodology ↗ · Original evidence ↗

Factual error rate Difficult prompts · OpenAI Sep 2026 · 3.91%

Answers containing a factual error on difficult, previously flagged ChatGPT conversations.

Version: Not reported · Answers with at least one factual error · percent · Lower is better · Task subset: OpenAI de-identified conversations where users flagged an earlier factual error

OpenAI · Published Sep 29, 2026 · Developer reported

Reporting source
OpenAI · GPT-6.1 Sol launch · Max effort
Evaluation origin
OpenAI research environment or API (source-computed)
Effort
Max
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
OpenAI research environment or API; deployment not individually specified
Checked
Sep 30, 2026

Exact percentage converted from the numerical fraction embedded in the official launch chart. Harness, fallback and tool details are retained only where the publication specifies them.

Deliberately error-inducing prompts, not representative of ordinary usage. Lower is better.

Original methodology ↗ · Original evidence ↗

Where you can use it1 API route · 11 subscriptions
OpenAI APIDirect API
ChatGPT Plus
$20 / month ↗

Monthly only. Canceling keeps access until the paid period ends.

ChatGPT Business (Standard seat)
$25 / month ↗

Workspace subscription; allowance depends on seat tier.

ChatGPT Pro 100
$100 / month ↗

Monthly; upgrades restart the billing cycle. Canceling keeps access until the paid period ends.

ChatGPT Pro 200
$200 / month ↗

Monthly; upgrades restart the billing cycle. Canceling keeps access until the paid period ends.

ChatGPT Pro 500
$500 / month ↗

Monthly. Canceling keeps access until the paid period ends.

Command Code Max 10×
$100 / month ↗

Monthly cycle; purchased credits carry over without expiry.

Command Code Max 20×
$200 / month ↗

Monthly cycle; purchased credits carry over without expiry.

Copilot Business
$19 / month ↗

Credits reset on the first of the month; unused included credits do not carry over.

Copilot Enterprise
$39 / month ↗

Credits reset on the first of the month; unused included credits do not carry over.

Copilot Pro+
$39 / month ↗

Credits reset on the first of the month; unused included credits do not carry over.

Copilot Max
$100 / month ↗

Credits reset on the first of the month; unused included credits do not carry over.

Pricing, assumptions & evidence

Published base rate

Input $10.00 · Output $50.00 · Cache read $1.00 · Cache write $12.50

Above 272K input tokens (priced at 2x input and cache rates, 1.5x output, for the full request): input $20.00, output $75.00, cache read $2.00.

Missing token-category prices are not zero. Workload pricing applies exact recorded categories and admitted routes. No benchmark score or quality ranking is inferred from prices.

Aliases, routes and identity

Catalog ID gpt-6-astra

Model release (kind: release)

Developer: OpenAI