Worker-tier routing sends planning to a strong model and execution to a cheaper tier. Quality lives in the plan; the executor runs the spec.
| Tier | $/M in/out (approx.) | Role |
|---|---|---|
| Fable 5 | $10 / $50 | Rare hardest-step escalation (refusals) |
| Opus 5 | $5 / $25 | Long-horizon orchestration |
| Sonnet 5 | $2 / $10 intro → $3 / $15 | Default worker |
| Haiku 4.5 | $1 / $5 | Parallel sub-agents |
| DeepSeek V4 Flash | $0.14 / $0.28 | Cheap executor (complete spec required) |
DeepSeek V4 Flash scores 79.0 on SWE-bench Verified against Opus 5’s 96.0, and 60/100 on a real agentic build - strong at executing a complete plan, weak when inventing architecture. One Flash output token costs ~1/89th of an Opus 5 token. Practitioners calibrate it as Sonnet-class for clear specs, not Opus-class for vague ones.
Conditions that make the split work
- The plan is genuinely complete - every ambiguity becomes an invention.
- The prompt prefix stays stable - cache hits are where the economics live.
- Still verify before done; a cheap worker is not an excuse to skip checks.
- Effort vs fast mode are separate dials on Claude tiers - do not confuse them with model choice.
Evidence: DeepSeek V4 Flash is a worker tier, not an Opus killer. Stack map: Why most agents default to the wrong Claude tier.