Coding
Signals for code generation, debugging, ML engineering, and terminal-based software agent work.
See model strengths as a shape, not a slogan.
A source-backed radar view for comparing coding, reasoning, research, multimodal, and agentic benchmark signals. Missing dimensions are treated as missing evidence, not as hidden claims.
Radar is useful only when the evidence is visible.
Each axis is a source-backed category rollup with visible coverage and confidence. Missing dimensions mean this registry does not yet have sourced evidence for that model and category.
OpenAI GPT-5.4 6 sourced benchmark rows
OpenAI GPT-5.2 Thinking 6 sourced benchmark rows
Claude Opus 4.6 2 sourced benchmark rows
Gemini 3.1 Pro 8 sourced benchmark rows
DeepSeek-R1 5 sourced benchmark rows
Task Lenses
Use-case lenses change which evidence matters most. These links go to category pages built from the same registry.
Reasoning
Signals for abstract reasoning, hard question answering, math, and multi-step problem solving.
Research
Signals for synthesis, factual work, long answers, RAG, and source-heavy analysis.
Multimodal
Signals for models that need to understand images, charts, video, and text together.
Agentic
Signals for tool use, browsing, GUI control, terminal operation, and multi-step autonomous workflows.
Capability Matrix
| Model | Coding | Reasoning | Research | Multimodal | Agentic | Evidence |
|---|---|---|---|---|---|---|
| OpenAI GPT-5.4 | 57.7 1 of 5 sourced rows limited | 59 3 of 6 sourced rows sufficient | 87.8 2 of 9 sourced rows limited | 58.8 2 of 7 sourced rows limited | 70.2 2 of 10 sourced rows limited | Verified source rows 5 of 5 axes with evidence |
| OpenAI GPT-5.2 Thinking | 67.8 2 of 5 sourced rows limited | 66.3 2 of 6 sourced rows limited | 79.1 2 of 9 sourced rows limited | 79.5 1 of 7 sourced rows limited | 67.1 3 of 10 sourced rows sufficient | Verified source rows 5 of 5 axes with evidence |
| Claude Opus 4.6 | 81.4 1 of 5 sourced rows limited | 36.2 1 of 6 sourced rows limited | No sourced row | 36.2 1 of 7 sourced rows limited | 81.4 1 of 10 sourced rows limited | Verified source rows 4 of 5 axes with evidence |
| Gemini 3.1 Pro | 67.8 3 of 5 sourced rows sufficient | 72.9 3 of 6 sourced rows sufficient | 90.1 2 of 9 sourced rows limited | 63.9 2 of 7 sourced rows limited | 72.3 4 of 10 sourced rows sufficient | Verified source rows 5 of 5 axes with evidence |
| DeepSeek-R1 | 49.2 1 of 5 sourced rows limited | 54.7 3 of 6 sourced rows sufficient | 79.3 3 of 9 sourced rows sufficient | 8.5 1 of 7 sourced rows limited | 49.2 1 of 10 sourced rows limited | Verified source rows 5 of 5 axes with evidence |
Evidence Rules
Capability rows use only `src/data/results.ts`. Demo names from the prototype, including Atlas, Forge, and Nova, are not used. Every visible score links back to a source record and carries a note about harness or methodology when relevant.