Claude Sonnet 4
Balanced Claude model for coding, analysis, agent workflows, and cost control
Shogo Agent Index
Supported#78 of 9226.2range 19.5–33 · 28% of weight measured
Agentic coding
26
11% measured
Long-horizon autonomy
36
50% measured
Knowledge-work agents
23
20% measured
Computer use
19
40% measured
Security & context learning
25
50% measured
Price
- Input
- $2.60/M
- Output
- $13.00/M
- Cache read
- $0.26/M
- 3:1 blend
- $5.20/M
- Agent-effective
- $1.20/M
Agent-effective = 20 input tokens per output token, 85% cache hits. Why.
Agentic benchmarks
| Benchmark | Score | Normalised | Index weight | Harness / source |
|---|---|---|---|---|
| SWE-bench Verified | est. | 8% | predicted 61 | |
| Terminal-Bench 2.0 | est. | 8% | predicted 23 | |
| DeepSWE | est. | 6% | predicted 18 | |
| FrontierSWE | est. | 5% | predicted 2 | |
| CursorBench | est. | 4% | predicted 23 | |
| GSO-Bench | 4.9% | 4% | OpenHands | |
| METR Time Horizons | 62.0% | 10% | source | |
| Vending-Bench 2 | est. | 6% | predicted 7 | |
| PostTrainBench | est. | 4% | predicted 15 | |
| APEX-Agents | est. | 8% | predicted 19 | |
| GDPval | est. | 7% | predicted 25 | |
| Remote Labor Index | est. | 5% | predicted 1 | |
| DeepResearch Bench | 46.6% | 5% | source | |
| OSWorld 2.0 | est. | 6% | predicted 2 | |
| OSWorld | 43.9% | 4% | claude-4-sonnet-20250514 (50 steps) | |
| Cybench | 35.0% | 5% | source | |
| CL-bench | est. | 5% | predicted 15 |
Other benchmarks
| Benchmark | Category | Score | Source |
|---|---|---|---|
| Aider Polyglot | Coding | 61.3% | source |
| ARC-AGI | Reasoning | 40.0% | source |
| ARC-AGI-2 | Reasoning | 5.9% | source |
| CritPt | Reasoning | 0.3% | source |
| DTBench | Knowledge | 77.1% | source |
| Fiction.LiveBench | Long context | 46.9% | source |
| GeoBench | Multimodal | 37.0% | source |
| GPQA Diamond | Reasoning | 79.2% | source |
| Humanity's Last Exam | Reasoning | 7.8% | source |
| LMCA | Knowledge | 29.0% | source |
| MATH Level 5 | Math | 84.4% | source |
| OTIS Mock AIME | Math | 71.1% | source |
| SciCode | Coding | 40.0% | source |
| SimpleBench | Reasoning | 45.5% | source |
| TheAgentCompany | Knowledge work | 33.1% | source |
| VPCT | Multimodal | 34.0% | source |