Tests computational research workflows, helping you assess what analysis and execution you could delegate so you can focus on interpreting results. Native-agent and independent setups remain separate, and success does not establish a scientific discovery's validity.
| Model / creator | Version | Reported score |
|---|---|---|
 Claude Mythos 5.2Anthropic · Proprietary | 0.1 / Claude Code / offline | 75.3% |
| Claude Fable 5.2Anthropic · Proprietary | 0.1 / Claude Code / offline | 74.6% |
| Claude Fable 5.2Anthropic · Proprietary | 0.1 / independent mini-swe-agent | 73.333% |
 GPT-6 AstraOpenAI · Proprietary | 0.1 / Claude Code / offline | 68.1% |
 Gemini 4 CarbonGoogle · Proprietary | 0.1 / Claude Code / offline | 64.8% |
| Claude Opus 5.5Anthropic · Proprietary | 0.1 / Claude Code / offline | 58.7% |
| Claude Fable 5.1Anthropic · Proprietary | 0.1 / Claude Code / offline | 52.6% |
effort: max · safeguards: false · November 9 launch-table comparison; highest available effort unless explicitly noted. Preserve the benchmark track and model configuration; competitor figures are as reported by Anthropic. · Tools enabled
Anthropic ↗effort: max · 70 tasks; 10 trials/task; Claude Code --bare; offline/cached resources; standard error ±4.3 points clustered by task. · Tools enabled
Anthropic ↗effort: max · 70 tasks; 10 trials/task; Claude Code --bare; offline/cached resources; standard error ±4.3 points clustered by task. · Tools enabled
Anthropic ↗effort: highest available · November 9 launch-table comparison; highest available effort unless explicitly noted. Preserve the benchmark track and model configuration; competitor figures are as reported by Anthropic. · Tools enabled
Anthropic ↗effort: highest available · November 9 launch-table comparison; highest available effort unless explicitly noted. Preserve the benchmark track and model configuration; competitor figures are as reported by Anthropic. · Tools enabled
Anthropic ↗effort: highest available · November 9 launch-table comparison; highest available effort unless explicitly noted. Preserve the benchmark track and model configuration; competitor figures are as reported by Anthropic. · Tools enabled
Anthropic ↗effort: highest available · November 9 launch-table comparison; highest available effort unless explicitly noted. Preserve the benchmark track and model configuration; competitor figures are as reported by Anthropic. · Tools enabled
Anthropic ↗effort: max · Claude Fable 5.2 (Max, Default Fallback); exact independent record; implementation per archived benchmarking methodology. · Tools enabled
effort: high · Claude Fable 5.2 (High, Default Fallback); exact independent record; implementation per archived benchmarking methodology. · Tools enabled
effort: xhigh · Claude Fable 5.2 (Xhigh, Default Fallback); exact independent record; implementation per archived benchmarking methodology. · Tools enabled
effort: medium · Claude Fable 5.2 (Medium, Default Fallback); exact independent record; implementation per archived benchmarking methodology. · Tools enabled
effort: low · Claude Fable 5.2 (Low, Default Fallback); exact independent record; implementation per archived benchmarking methodology. · Tools enabled
Tests extracting text, tables and formulas from documents, helping you choose a model for turning client files into usable data with less manual copying.
AutomationBenchTests whether an agent can finish business workflows across apps and APIs, helping you assess what sales, support or operations work you could automate to spend more time with clients.
DeepSWETests whether an agent can understand a codebase and complete substantial engineering changes, helping you assess which model could take on more development work while you focus on requirements and review..
