Default model route
Map routine OpenClaw chat, planning, and tool narration to deepseek-v4-flash.
This keeps high-frequency agent turns on the lowest DeepSeek V4 route while preserving the same model-family positioning.
API routing
DeepSeek V4 Flash API routing playbook
Route by job type, not by brand preference. V4 Flash handles default OpenClaw and API traffic; V4 Pro and other providers enter only when a clear trigger is met.
Routing table
Default model, escalation, and metric
This page adds a crawlable operational table that turns the Flash-first story into a concrete policy.
| Task | Default | Escalation | Metric |
|---|---|---|---|
| OpenClaw routine agent turns | DeepSeek V4 Flash | V4 Pro after repeated failed checks | Task completion cost, latency, retry rate |
| Search and retrieval answers | DeepSeek V4 Flash | Claude/GPT sample audit for regulated pages | Citation quality, hallucination rate, cache hit ratio |
| Code explanation and batch fixes | DeepSeek V4 Flash | V4 Pro for complex debugging plans | Accepted patch rate, follow-up turns, token cost |
| Long-form strategy or policy review | V4 Pro or Claude/GPT audit | Human review | Review accuracy, missed risks, edit distance |
OpenClaw
Adapter-aware routing rules
OpenClaw needs dedicated model-routing language because agent traffic repeats context, tools, and planning scaffolds.
Escalation policy
Escalate to deepseek-v4-pro only for long reasoning, failed self-checks, or high-risk review.
Pro should improve difficult tasks without turning every OpenClaw request into a premium request.
Prompt cache design
Keep OpenClaw system prompts, tool schemas, and retrieval wrappers stable across runs.
Stable prompt scaffolding is the direct path to cache-hit economics on repeated agent workflows.
Comparison fallback
Keep GPT, Claude, Gemini, and Grok as explicit audit or customer-required routes.
The comparison layer should support routing decisions without weakening the Flash-first message.