modelbench日本語

AI model × effort benchmark comparison

Cost and score, by effort.Compare models and effort by cost and score

Explore published benchmark results by model and effort to help choose models and effort in Claude Code or Codex. Every value has a source.

Find models in the table, where each model name links to its page.

Models to compare
57total
Benchmarks
104total
Data retrieved(latest)
2026-11-14

Cost and score chart

Points for the same model are connected in order of increasing effort. On the cost axis, points nearer the upper left are cheaper and score higher. See “How to read the chart” below.

X-axis

Terminal-Bench 4.0

    How to read the chart
    • The y-axis shows score; the x-axis shows cost reported by the source (log scale) or effort order. On the cost axis, points nearer the upper left are cheaper and score higher.
    • Costs are amounts reported by sources; metrics differ by source (per task, per trial, estimates, and so on). The metric appears on the axis, in tooltips, and in the data points table. Costs are not calculated from API prices. Points without source costs cannot be plotted on a logarithmic axis, so only their count is shown.
    • Hollow points are approximate values read from source charts.
    • Below the chart, counts of displayed points are grouped by the party identified in the source as measuring or reporting them. Even in official announcements, the publisher may have measured values for other companies’ models.
    • Official announcements, third-party measurements, and papers use different conditions, so they are shown separately by source type. Even within a source type, results with different measurement conditions (such as harness or pass@1 / best-of-N) are placed in separate comparison groups.
    Data points and sources for this chart
    ModeleffortScoreSource cost (metric)measured/reported byExtraction methodsourceRetrieved
    Overall (performance first)

    Rankings

    More rankings

    This aggregation compares models and entries with values for every benchmark in the set, using only third-party results. Models: 8; entries: 8; benchmarks: 4 (GPQA Diamond, FrontierMath Tiers 1-3(v2), FrontierMath Tier 4(v2), SimpleQA Verified). It is not an overall score for the entire site.

    RankModel (effort)providerScore (percentile)CoveragePrice (input / output)
    1Claude Fable 5.2MaxAnthropic · Coverage 6/10 · Price $8.00 / $40.00Anthropic846/10$8.00 / $40.00
    1GPT-6 AstraMaxOpenAI · Coverage 9/10 · Price $10.00 / $50.00OpenAI849/10$10.00 / $50.00
    3GPT-6.1 SolMaxOpenAI · Coverage 9/10 · Price $2.00 / $10.00OpenAI799/10$2.00 / $10.00
    4Claude Opus 5.5MaxAnthropic · Coverage 7/10 · Price $4.00 / $20.00Anthropic507/10$4.00 / $20.00
    5GPT-6.1 TerraMaxOpenAI · Coverage 6/10 · Price $1.00 / $5.00OpenAI466/10$1.00 / $5.00
    6Claude Sonnet 5.5MaxAnthropic · Coverage 7/10 · Price $2.00 / $10.00Anthropic437/10$2.00 / $10.00
    7GPT-6 LunaMaxOpenAI · Coverage 9/10 · Price $0.10 / $0.50OpenAI149/10$0.10 / $0.50
    8Claude Haiku 5.5MaxAnthropic · Coverage 6/10 · Price $0.10 / $0.50Anthropic06/10$0.10 / $0.50
    • Scores are rank percentiles within each comparison group (first = 100, last = 0), averaged within each benchmark and then across benchmarks. They are comparable only within this aggregation.
    • Coverage is the number of benchmarks with a value for the model × effort / the number of benchmarks with third-party results.
    • measured/reported by and counts for values in this aggregation: Epoch AI (32)
    • Prices are USD per 1 million tokens (from each provider’s pricing table; they do not vary by effort). Sources are on the model pages.
    • Rankings include only models currently available in Claude Code / Codex. Rankings including legacy and unavailable models are in All model rankings (records).
    By use case

    Compare by role

    We choose benchmark categories to emphasize for each role and compare models available in Claude Code / Codex. Select a role for the Claude Code comparison, or use the links below for Codex.