Most coding requests don't need the strongest model. Renaming a variable, writing a test, reading a stack trace: a model that costs a twentieth as much does these just as well. The hard part is knowing which request is which before you send it. That's what Sparrow Auto does.
How it works
- It looks at each task first. A small, fast model reads the request and rates it as simple, medium or hard. If that check is unsure, Auto rates the task higher, not lower.
- It picks the lowest-cost model that fits. Simple work goes to models like DeepSeek V4.1 Flash, medium work to models like Claude Haiku 5.5, hard work to Claude Sonnet 5.5 and models of its class.
- It keeps one model per task. Switching models mid-task would throw away the prompt cache, which is where most of the savings in long coding sessions come from.
- It steps up when things go wrong. If tests or builds keep failing, or a task drags on, Auto moves that task to a stronger model.
- It works with your plan. On a plan, Auto only uses models the plan includes, so you're never charged extra by surprise.
You pay the price of whichever model answered, and every response says which one it was.
The test
The same test we used for the low-cost coding models: ten Python tasks checked by hidden tests, and an agent job that fixes six failing tests in a small repository. We ran it through Sparrow Auto and through Claude Sonnet 5.5 on its own.
| Sparrow Auto | Claude Sonnet 5.5 | |
|---|---|---|
| Tasks passed | 10 / 10 | 10 / 10 |
| Agent job | 6 / 6 in 3 turns | 6 / 6 in 3 turns |
| Cost of the whole run | $0.0053 | $0.0693 |
| Time for the ten tasks | 92 s | 38 s |
Same results, 92% lower cost. Auto sent six of the ten tasks to Claude Haiku 5.5 and four to DeepSeek V4.1 Flash, and did the agent job on Haiku.
It was slower: 92 seconds against 38 for the ten tasks. Part of that is the check before each task, and part is that DeepSeek V4.1 Flash thinks longer than Sonnet does. For an agent working through a long task, that trade is usually worth it. If you are waiting on every answer, pick a model yourself.
Use it
In the SparrowCodes CLI, Sparrow Auto is a model like any other: pick it with /models. The sidebar shows which model it chose for the current task and how much of your plan is left.
Through the API, use the model id sparrow/auto:
curl https://api.sparrowcodes.com/v1/chat/completions \
-H "Authorization: Bearer $SPARROWCODES_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "sparrow/auto", "messages": [{"role": "user", "content": "Add input validation to this function: ..."}]}'Sparrow Auto works with credits and with every plan, including the free one.