Model API pricing
/api/model-pricingList prices and published task-bench scores for major model APIs. Each bench is one board, and an empty cell means that board published no score.
Loading prices.
24 models
By bench
One chart per board, grouped by the kind of work. Click a name to keep that model, and Command-click to add another. On a point chart, hovering a point draws a line at that score and marks every point above it.
Computer use
OSWorld 2.0 strict
The vertical axis is the score and the horizontal axis is the thinking level.
A desktop task scores only when it finishes completely. Each cell is that model's latest run on XLANG's full set at 500 steps, and partial credit is named in the note.
OSWorld 2.0 offline
- GPT-6 Astra72.6%
- GPT-5.6 Sol65.7%
Desktop tasks with no internet, scored with partial credit. This is OpenAI's offline set, so it is a different task list from the strict row.
OSWorld 2.1 partial
- Claude Opus 5.5Partial.81.8%
- Claude Sonnet 5.5Partial.80.1%
- Claude Sonnet 5Partial.57.0%
Desktop tasks scored with partial credit on OSWorld 2.1. This is Anthropic's published partial score, not the binary strict board above.
Terminal agent
Terminal-Bench 4.0
The vertical axis is the score and the horizontal axis is the cost of the full run.
Terminal tasks in software, configuration and data analysis. Each point is one model, and the dollar amount is the cost of the full 66-task run.
Terminal-Bench 4.0, Anthropic
- Claude Sonnet 5.570.6%
- Claude Opus 5.5xhigh.66.4%
- Claude Sonnet 510.3%
Anthropic's own Terminal-Bench 4.0 run, published with Sonnet 5.5. It is not the Snorkel board in the row above.
Coding agent
DeepSWE v1.1
The vertical axis is the score and the horizontal axis is the mean cost per task, on a log scale.
Long coding tasks in real repositories. A line joins one model's thinking levels, and each point is that level's mean cost per task.
CursorBench 3.2
- Claude Fable 5.173.4%
- Claude Fable 570.5%
- Claude Opus 570.0%
- GPT-5.6 Sol67.2%
Agentic coding tasks run in Cursor. The bar is the score published for that model.
FrontierCode 1.1
- Claude Opus 5.554.4%
- Claude Sonnet 5.5Xhigh. Max scored 46.2%.52.1%
- GPT-6 Sol49.3%
- Claude Sonnet 542.4%
Whether a code change would be merged without extra edits. Sonnet 5.5's marked score is xhigh, and max is named in the note.
CursorBench 4.0
- Claude Opus 5.557.8%
- Claude Sonnet 5.555.5%
- Claude Sonnet 534.1%
Coding tasks from real Cursor sessions. This is CursorBench 4.0, not the 3.2 comparison above.
Research agent
Terminal-Bench Science 0.1
- GPT-6 AstraOpenAI's run.64.6%
- Claude Fable 5.152.6%
- Claude Opus 529.0%
- Claude Fable 524.7%
- GPT-5.6 Sol22.4%
Scientific workflows that analyse data, run simulations and fit models. A note names a score from another lab's run.
Workflow agent
AutomationBench
- GPT-6 AstraOpenAI's run.41.4%
- Claude Fable 5.131.4%
- Claude Opus 526.9%
- GPT-5.6 Sol19.6%
- Claude Fable 517.1%
Multi-step business workflows. A note names a score from another lab's run.
Knowledge work
GDPval-AA v2
Bars show the spread between these scores.
- Claude Fable 5.11,853
- Claude Opus 51,824
- Claude Fable 51,723
- GPT-5.6 Sol1,711
Professional knowledge-work tasks, scored as Elo. A longer bar is a higher score, and the axis runs from the lowest score here to the highest.
GDPval-AA v2.1
Bars show the spread between these scores.
- Claude Opus 5.51,846
- Claude Sonnet 5.5Pre-release run. A structured-output bug may understate this score.1,844
- GPT-6 SolMay predate an image-understanding fix.1,487
- Claude Sonnet 51,449
Professional knowledge-work tasks, scored as Elo on GDPval-AA v2.1. A note names a pre-release run or a score that may predate a fix.
AA-Briefcase v1.1
Bars show the spread between these scores.
- Claude Opus 5.51,822
- Claude Sonnet 5.5Pre-release run. A structured-output bug may understate this score.1,811
- GPT-6 SolMay predate an image-understanding fix.1,483
- Claude Sonnet 51,359
Long-horizon knowledge work, scored as Elo. A note names a pre-release run or a score that may predate a fix.
Reasoning
Humanity's Last Exam
- Claude Opus 5.5With tools.67.7%
- Claude Sonnet 5.5With tools.64.5%
- Claude Sonnet 5With tools.54.9%
Multidisciplinary questions, answered with tools. Every cell on this board is a with-tools score.
Vision
Chartography
- Claude Opus 5.5No tools.64.4%
- Claude Sonnet 5.5No tools.61.6%
- GPT-6 SolNo tools. Image-understanding fix. Anthropic's check says this score did not change.53.6%
- Claude Sonnet 5No tools.15.6%
Reading charts, without tools. A note names a score checked after an image-understanding fix.
Scores
Each row is one board, and the shaded cell is the highest score on that row. A cell shows the best score with its thinking level in brackets, and hovering it lists the other figures for that score.
| Bench | Claude Fable 5.1Anthropic | Claude Fable 5Anthropic | Claude Opus 5.5Anthropic | Claude Opus 5Anthropic | Claude Opus 4.8Anthropic | Claude Opus 4.7Anthropic | Claude Sonnet 5.5Anthropic | Claude Sonnet 5Anthropic | Claude Sonnet 4.6Anthropic | GPT-6 AstraOpenAI | GPT-6 SolOpenAI | GPT-5.6 SolOpenAI | GPT-5.6 TerraOpenAI | GPT-5.6 LunaOpenAI | GPT-5.5OpenAI | Gemini 3.8 FlashGoogle | Grok 4.7xAI | Grok 4.6xAI | Grok 4.5xAI | Kimi K2.6Moonshot | Kimi K2.7 CodeMoonshot | DeepSeek V4 ProDeepSeek | Qwen 3.7-PlusAlibaba | MiniMax M3MiniMax |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Computer use OSWorld 2.0 strict | — | — | — | max 44.3%. xhigh 33.3%. high 36.9%. medium 33.0%. low 18.8%. v2.1. Partial 77.7% at max. | Partial 54.8%. Batched tools. | Partial 48.9%. Batched tools. | — | — | max 8.3%. medium 9.3%. Partial 33.9% at medium. | — | — | v2026.08.08. Partial 62.7%. | — | — | Partial 49.5%. | — | — | — | — | Partial 22.1%. | — | — | Partial 21.5%. | Partial 22.3%. |
Computer use OSWorld 2.0 offline | — | — | — | — | — | — | — | — | — | 72.6% | — | 65.7% | — | — | — | — | — | — | — | — | — | — | — | — |
Computer use OSWorld 2.1 partial | — | — | 81.8%Partial. | — | — | — | 80.1%Partial. | 57.0%Partial. | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — |
Terminal agent Terminal-Bench 4.0 | max, Claude Code. Run $6.2k. | max, Claude Code. Run $7.3k. | — | xhigh, Claude Code. Run $6.1k. | max, Claude Code. Run $6.5k. | — | — | max, Claude Code. Run $9.6k. | — | max, Codex. Run $3.3k. | — | max, Codex. Run $2.5k. | max, Codex. Run $1.7k. | max, Codex. Run $0.3k. | — | high, mini-SWE-agent. Run $1.8k. | xhigh, Grok Build. Run $3.7k. | high, Grok Build. Run $3.6k. | high, Grok Build. Run $2.1k. | — | — | — | — | — |
Terminal agent Terminal-Bench 4.0, Anthropic | — | — | 66.4% (xhigh) | — | — | — | 70.6% | 10.3% | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — |
Coding agent DeepSWE v1.1 | — | max 69.7% · $21.63 a task. xhigh 69.9% · $13.41 a task. high 68.6% · $9.18 a task. medium 65.4% · $6.09 a task. low 59.6% · $3.76 a task | — | max 73.6% · $11.84 a task. xhigh 73.2% · $9.07 a task. high 72.8% · $6.08 a task. medium 68.9% · $3.29 a task. low 58.1% · $1.66 a task | max 59.0% · $13.22 a task. xhigh 54.4% · $8.01 a task. high 51.8% · $4.28 a task. medium 48.7% · $3.44 a task. low 40.8% · $2.29 a task | — | — | max 53.8% · $26.40 a task. xhigh 49.7% · $11.89 a task. high 48.2% · $7.43 a task. medium 39.8% · $4.08 a task. low 30.5% · $2.19 a task | — | max 73.2% · $7.50 a task. xhigh 74.1% · $4.43 a task. high 73.2% · $3.92 a task. medium 72.8% · $3.08 a task. low 67.0% · $1.60 a task | — | max 72.7% · $6.46 a task. xhigh 70.7% · $3.60 a task. high 69.4% · $2.66 a task. medium 61.1% · $1.42 a task. low 45.4% · $0.82 a task | max 69.6% · $3.96 a task. xhigh 60.2% · $1.70 a task. high 53.8% · $0.91 a task. medium 35.1% · $0.47 a task. low 24.1% · $0.34 a task | max 67.2% · $0.61 a task. xhigh 56.9% · $0.31 a task. high 44.2% · $0.16 a task. medium 11.3% · $0.04 a task. low 1.5% · $0.01 a task | — | high 73.8% · $2.36 a task. medium 71.0% · $1.97 a task | — | xhigh 66.7% · $5.50 a task. high 65.2% · $4.38 a task. medium 67.5% · $3.45 a task. low 41.6% · $1.04 a task | $2.42 a task | — | No effort setting.. $2.82 a task | $1.67 a task | — | — |
Coding agent CursorBench 3.2 | 73.4% | 70.5% | — | 70.0% | — | — | — | — | — | — | — | 67.2% | — | — | — | — | — | — | — | — | — | — | — | — |
Coding agent FrontierCode 1.1 | — | — | 54.4% | — | — | — | Xhigh. Max scored 46.2%. | 42.4% | — | — | 49.3% | — | — | — | — | — | — | — | — | — | — | — | — | — |
Coding agent CursorBench 4.0 | — | — | 57.8% | — | — | — | 55.5% | 34.1% | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — |
Research agent Terminal-Bench Science 0.1 | 52.6% | 24.7% | — | 29.0% | — | — | — | — | — | 64.6%OpenAI's run. | — | 22.4% | — | — | — | — | — | — | — | — | — | — | — | — |
Workflow agent AutomationBench | 31.4% | 17.1% | — | 26.9% | — | — | — | — | — | 41.4%OpenAI's run. | — | 19.6% | — | — | — | — | — | — | — | — | — | — | — | — |
Knowledge work GDPval-AA v2 | 1,853 | 1,723 | — | 1,824 | — | — | — | — | — | — | — | 1,711 | — | — | — | — | — | — | — | — | — | — | — | — |
Knowledge work GDPval-AA v2.1 | — | — | 1,846 | — | — | — | 1,844Pre-release run. A structured-output bug may understate this score. | 1,449 | — | — | 1,487May predate an image-understanding fix. | — | — | — | — | — | — | — | — | — | — | — | — | — |
Knowledge work AA-Briefcase v1.1 | — | — | 1,822 | — | — | — | 1,811Pre-release run. A structured-output bug may understate this score. | 1,359 | — | — | 1,483May predate an image-understanding fix. | — | — | — | — | — | — | — | — | — | — | — | — | — |
Reasoning Humanity's Last Exam | — | — | 67.7%With tools. | — | — | — | 64.5%With tools. | 54.9%With tools. | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — |
Vision Chartography | — | — | 64.4%No tools. | — | — | — | 61.6%No tools. | 15.6%No tools. | — | — | 53.6%No tools. Image-understanding fix. Anthropic's check says this score did not change. | — | — | — | — | — | — | — | — | — | — | — | — | — |
References
- OSWorld 2.0 strict · Computer use
- A desktop task counts only when the whole task finishes. XLANG's official board, full set, 500-step budget, binary accuracy. Read on 28 September 2026 from the board dated 17 September 2026. Where a model has several thinking levels, each one is listed and the marked number is the best binary score. The board prints one or two decimals. These figures keep one. A note names the partial score and, where the row is not from the original runs, the release. v2.1, the August 2026 release and the original runs are not one task list. Not the same set as the offline row.
- OSWorld 2.0 offline · Computer use
- A desktop task without internet, with partial credit. OpenAI's v2026.08.08 offline set, partial score.
- OSWorld 2.1 partial · Computer use
- A desktop task scored with partial credit. Anthropic's Sonnet 5.5 comparison, partial score, read on 28 September 2026. Not XLANG's binary strict board.
- Terminal-Bench 4.0 · Terminal agent
- Terminal tasks covering software, configuration and data analysis. Snorkel's Harbor run of terminal-bench/terminal-bench@4.0.0, 66 tasks. The note names the effort, the agent and the cost of that full run. The board does not publish a cost per task.
- Terminal-Bench 4.0, Anthropic · Terminal agent
- Anthropic's Terminal-Bench 4.0 run. Anthropic's comparison, read on 28 September 2026. Opus 5.5 is at xhigh, its highest score. Sonnet 5.5's cell names no effort. No full-run cost is published. Not Snorkel's Harbor board.
- DeepSWE v1.1 · Coding agent
- Long-horizon changes in real repositories. Datacurve's board, 113 tasks, mini-swe-agent, four runs, dated 22 September 2026. Where that board published several thinking levels, each one is listed with its mean cost per task. The marked number is the best score. The board prints a whole percent. These figures keep one decimal from the same pass rate.
- CursorBench 3.2 · Coding agent
- Agentic coding tasks in Cursor. Anthropic's comparison.
- FrontierCode 1.1 · Coding agent
- Whether an agent's code change would be merged. Anthropic's FrontierCode 1.1 main set, read on 28 September 2026. Sonnet 5.5's best published score is xhigh. Max scored lower because the bench penalises out-of-scope edits.
- CursorBench 4.0 · Coding agent
- Coding tasks taken from real Cursor sessions. Anthropic's CursorBench 4.0 comparison, read on 28 September 2026. Not CursorBench 3.2.
- Terminal-Bench Science 0.1 · Research agent
- Scientific workflows that analyse data, run simulations and fit models. Anthropic's setup, with a standard error of about 4 points, unless the cell names another run.
- AutomationBench · Workflow agent
- Multi-step business workflows. Anthropic's comparison, unless the cell names another run.
- GDPval-AA v2 · Knowledge work
- Professional knowledge-work tasks, scored as Elo. Anthropic's comparison.
- GDPval-AA v2.1 · Knowledge work
- Professional knowledge-work tasks, scored as Elo. Anthropic's comparison of Artificial Analysis GDPval-AA v2.1, read on 28 September 2026. Not GDPval-AA v2. Sonnet 5.5 was a pre-release run that may understate the score. GPT-6 Sol may predate an image-understanding fix.
- AA-Briefcase v1.1 · Knowledge work
- Long-horizon knowledge work, scored as Elo. Anthropic's comparison of Artificial Analysis AA-Briefcase v1.1, read on 28 September 2026. Sonnet 5.5 was a pre-release run that may understate the score. GPT-6 Sol may predate an image-understanding fix.
- Humanity's Last Exam · Reasoning
- Multidisciplinary questions, with tools. Anthropic's comparison, with tools, read on 28 September 2026.
- Chartography · Vision
- Reading charts, without tools. Anthropic's comparison of Surge Chartography, no tools, read on 28 September 2026. GPT-6 Sol's official score may predate an image-understanding fix. Anthropic's own check says that score did not change.
Prices checked 27 September 2026. Bench scores were read from the linked boards on 28 September 2026. A dash means that board published no score. The OSWorld rows use different task sets.