deepseek-cost-router
Skill0 starsChanged 2 months agoMITView on GitHub
Route coding tasks between DeepSeek chat and reasoning models (deepseek-chat vs deepseek-reasoner) to minimize API cost while keeping quality. Use when configuring an agent to use DeepSeek, choosing a model for a task batch, or optimizing token spend.
- Reads credentials
In this repository8 other files it gives its agents.
- AGENTS.mdAGENTS.md
- improve-claude-mdSkill
- show-meSkill
- claude-plugin-devSkill
- deepseek-migrationSkill
- dsh-plugin-devSkill
- skill-authorSkill
- skill-portability-auditSkill
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 24
Default to chat for: file edits driven by clear instructions, renames, boilerplate, doc updates, formatted output, summarization, simple …
- Distinctive: line 35
If the result fails review/tests, re-run just that unit on reasoner.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Routing work to the right DeepSeek model
DeepSeek exposes two behaviors through its OpenAI-compatible API:
deepseek-chat— the fast, non-thinking path. Cheap, low latency.deepseek-reasoner— the thinking path (chain-of-thought before the final answer). Slower and pricier per call, dramatically better on hard reasoning.
The routing skill: run routine volume on chat, escalate only reasoning-heavy work.
Escalation heuristics
Default to chat for: file edits driven by clear instructions, renames, boilerplate, doc updates, formatted output, summarization, simple lookups, bulk repetitive transformations.
Escalate to reasoner for: non-obvious bugs (state/caching/async/race), algorithm design, architecture tradeoffs, tricky merges, math, contract or spec violations, "the tests fail and the cause is unclear".
A cheap escalation ladder for agent harnesses that support it:
- Try the task on chat with a tight instruction.
- If the result fails review/tests, re-run just that unit on reasoner.
- Feed the reasoner's answer back as the accepted fix; continue on chat.
Escalating units of failure instead of whole sessions keeps the premium model's token share small.
Cost mechanics worth knowing
- Context caching: repeated identical prefixes (system prompt, long file context) hit DeepSeek's automatic prefix cache and are billed at the much cheaper cache-hit rate. Keep your system prompt and repo map stable and first in the prompt; don't shuffle dynamic content in front of them.
- Reasoner output is token-hungry: thinking tokens bill as output. A reasoner call can cost many multiples of a chat call — reserve it.
- Batch offline work (code review of a branch, doc generation) into dedicated sessions where the shared context prefix is reused.
Wiring DeepSeek into coding agents
DeepSeek's API is OpenAI-compatible, and Anthropic-compatible endpoints are documented for harnesses like Claude Code — see the official integration hub: https://api-docs.deepseek.com/guides/coding_agents (per-tool guides for Claude Code, OpenCode, OpenClaw, Codex and more). You need:
DEEPSEEK_API_KEYfrom https://platform.deepseek.com- The base URL (
https://api.deepseek.com) and model id set in your tool's provider config — follow the per-tool guide for exact env vars.
DeepSeek Harness (dsh) reads DEEPSEEK_API_KEY directly; if aipx doctor
warns it's missing, export it before launching.
Verify
- Spot-check 10 recent tasks: would reasoner have changed the outcome? If no, your routing is right.
- Watch spend per model in the platform dashboard; a healthy setup keeps reasoner tokens a small fraction of total while review pass-rates hold.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

