# DeepSeek V3.1 on the Cotera agent benchmark

> Cheapest model in the matrix. Also the most likely to bail at "let me search for…" without ever calling a tool.

Source: https://cotera.co/benchmarks/models/deepseek-v3.1

---

DeepSeek's V3.1 — the Chinese open-weights frontier model that defines the cost floor for serious reasoning. We tested via DeepSeek's first-party API.

**Score:** 3/5 benchmarks passed · **Cost across the matrix:** $0.113

## Strengths
- Total spend was $0.113 — cheaper than every other model in the matrix by a wide margin.
- The three tasks it actually engaged with all passed cleanly under $0.06 each.
- Reddit was $0.046 — second-cheapest Marketing pass after DeepSeek's own willingness to skip verification on Sales.

## Watch-outs
- Bailed twice. On both Sales and Coding, DeepSeek wrote a one-line intent statement ("I'll help you research..." / "Let me create a TypeScript function...") then stopped — no tool calls on Sales, exactly one tool call on Coding. Total damage: $0.001 to produce nothing.
- This is a known V3.1 failure mode when the system prompt has a verbose persona block. The model treats the agent loop like a chat opener and waits for the user to confirm.
- When it does engage, it's excellent. When it doesn't, you get a polite preamble and a closed run. You need supervisor logic to retry these in production.