Devstral 2 123B vs Mistral Small 4
Mistral's code specialist vs its latest small model — which is better for agentic workloads?
Head-to-head: agentic-core-v1
Harness: agentic-core-v1 — 10 tasks × 3 runs = 30 total. Binary pass/fail per run. Harness version: openclaw@2026.4.22. Full methodology: what we measure and why.
| Metric | Devstral 2 123B | Mistral Small 4 |
|---|---|---|
| Runs passed / total | 27 / 30 | 29 / 30 |
| Pass rate | 90.0% | 96.7% |
| Campaign cost (30 runs) | $0.05 | $0.03 |
| Cost per run | $0.00167 | $0.00100 |
| Cost per passing run | $0.00185 | $0.00103 |
| Provider | Mistral AI | Mistral AI |
| Campaign date | 2026-05-17 | 2026-05-31 |
Devstral 2 123B
Code-specialist training shows on multi-step planning tasks; efficient tool dispatch strategy
task_09 failure consistent with non-reasoning mid-tier models; 123B footprint requires managed API
Mistral Small 4
96.7% pass rate at $0.001/run — best cost-efficiency in the dataset for frontier-tier accuracy
Single task_09 failure; less tested on non-coding task variants than Devstral 2
Verdict
Mistral Small 4 leads by 2 runs (29/30 vs 27/30). That gap represents 2 task categories where Devstral 2 123B fails consistently. It also costs less per passing run ($0.00103 vs $0.00185). Mistral Small 4 is the stronger choice on both dimensions.
About agentic-core-v1
agentic-core-v1 is modelbattles' flagship benchmark harness. Ten tasks drawn from real engineering work: fix a failing test, refactor duplicated code, investigate a production log, add a null guard, trace through a codebase. Each task runs three times per model. A run passes if and only if the checker accepts the output — no partial credit, no manual grading.
The harness is deterministic: same task, same environment, same checker across all models. Scores are comparable. task_09 is the persistent difficulty point — it requires the model to recognise a structurally impossible calculation and refuse to produce a wrong answer instead of looping. Most models fail it at least once.
Read: What we actually measure and why · How to read an agentic-core-v1 score
More comparisons
- Claude Sonnet 4.6 vs Claude Haiku 4.5 — A 1-point score gap and a 16× cost gap — is Sonnet worth it?
- DeepSeek V4 Flash vs Claude Sonnet 4.6 — Same agentic-core-v1 score, 36× cost difference — what does Sonnet actually buy you?
- Mistral Small 4 vs Claude Haiku 4.5 — The two most cost-efficient models in the top tier — one passes more, the other costs less.
- GPT-5.5 Instant vs Claude Sonnet 4.6 — Both mid-tier flagships at 90%+ — OpenAI vs Anthropic on real agentic tasks.
- Claude Opus 4.8 vs DeepSeek V4 Pro Thinking — Both 30/30 on agentic-core-v1. One costs $7.34. The other costs $0.12.