AI Model Capability Comparison
Knowledge work
Only models released within the past year are shown. Only benchmarks covering at least 3 of these models are shown. Unknown release dates are excluded; historical data is retained.
Vals IndexVals composite index
— means no published score for this candidate in this source, not zero.7/11 recent-year models have scores
| No. | Model | Cost / TestCost per test | Input costInput price per million tokens | Output costOutput price per million tokens | DurationReported test duration | Input / Output TokensReported input / output token usage | |
|---|---|---|---|---|---|---|---|
| 1 |  Claude Sonnet 5.5Claude Sonnet 5.5 | 70.85 | $18.7 | $2 | $10 | 86 min | 45.09B / 654.94M |
| 2 | Claude Opus 5.5Claude Opus 5.5 | 70.18 | $28.12 | $4 | $20 | 85 min | 31.96B / 499.75M |
| 3 |  Gemini 4 ArgonGemini 4 Argon | 69.24 | $13.37 | $4 | $20 | 44.97 min | 13.36B / 146.17M |
| 4 |  GPT-6 AstraGPT-6 Astra | 65.02 | $21.28 | $10 | $50 | 33.65 min | 4.92B / 96.77M |
| 5 | GPT-6.1 SolGPT-6.1 Sol | 63.94 | $3.61 | $2 | $10 | 44.4 min | 5.49B / 103.20M |
| 6 | Claude Haiku 5.5Claude Haiku 5.5 | 58.71 | $3.17 | $0.1 | $0.5 | 49.8 min | 23.07B / 765.85M |
| 7 | GPT-6 LunaGPT-6 Luna | 52.32 | $0.45 | $0.1 | $0.5 | 32.38 min | 10.05B / 202.17M |
PandaNpc Updated Oct 21, 2026, 06:01 UTC
Source: Epoch AI, CC BY 4.0 · Vals.ai · Artificial Analysis (artificialanalysis.ai)
Column guide
- No.
- Display order only. Sorting changes this number; it is not a published capability rank.
- Model
- The evaluated model name and its provider icon. A dash in a score column means no published score is available here, not a zero.
- Cost / Test
- Cost per test · USD
- Input cost
- Input price per million tokens · USD / 1M tokens
- Output cost
- Output price per million tokens · USD / 1M tokens
- Duration
- Reported test duration · s / min
- Input / Output Tokens
- Reported input / output token usage · tokens
- Vals.ai
- Costs and times are reported by Vals for this benchmark, not universal model prices or latency. CUA-bench estimated API cost covers evaluation usage and is a lower-bound estimate, not cost per task; time per action is an average, not a median. An asterisk marks launch-day prompt-cache behavior that increased cache-write costs. Missing values are unpublished, not zero.
- Vals Finance Agent v2Vals finance agent v2
- Tests financial analysis with search, financial reports and calculation tools, including cross-document data checks, company performance comparisons and valuation. The main metric is Partial Credit, averaged over three runs: answer checks receive severity-based weights, and a critical dealbreaker error makes that question score zero. Higher is better. This is distinct from the separate All-Pass metric and is not whole-question pass rate. Evaluation methodology ↗
- Harvey's Legal Agent BenchmarkHarvey legal agent benchmark
- Tests agents reading and editing files to produce multi-step legal work products. A task passes only when every grading requirement is met. The score is the average task pass rate from two judges; higher is better. It is not the percentage of individual criteria passed, and it measures legal agent workflows rather than LegalBench's legal reasoning questions. Evaluation methodology ↗
- Vals IndexVals composite index
- A composite index of professional tasks in areas such as finance, coding and law. Its percentage score combines benchmark results using the weighting method for that published version; higher means stronger overall performance, not the accuracy of a single task. Included domains and weights change across versions, so newer methodology must not be applied retrospectively to older score snapshots. Evaluation methodology ↗
- LegalBenchLegal reasoning benchmark
- Tests understanding and reasoning over legal text, including identifying legal issues, recalling and applying rules, and judging legal conclusions. The score is accuracy on legal reasoning questions; higher is better. Unlike Harvey's Legal Agent Benchmark, it does not measure an agent producing multi-step legal work products by editing files. Evaluation methodology ↗