Coding · Model selection
Choose models for coding: API prices, context windows and tool calling
Last updated：
Start by defining the task: explaining a function, refactoring a project, or running an Agent with repeated tool calls. Prices help estimate cost and capability fields narrow the candidates; use your own task samples for the final choice.
Tikway’s model catalog includes protocols, context, capabilities and pricing. This guide provides a selection method without an untested “best model” ranking.
1. Identify requirements by task
| Task | Priorities | Acceptance checks |
|---|---|---|
| Explain code or make simple edits | Text capability, input/output cost | Accurate explanation and important edge cases. |
| Refactor a project | Context, output limit, instruction following | Project tests pass and references remain consistent. |
| Agent tool execution | Function/tool support and client protocol | Correct arguments, continued execution after tool results, bounded loops. |
| Find UI issues from screenshots | Image understanding and input charges | Recognizes the image rather than guessing from text. |
| Batch classification or extraction | Structured output, cost and error handling | Format acceptance rate and cost per successful task. |
Clients also impose protocol requirements. For Claude Code, check Anthropic configuration. A catalog tool-support flag does not establish compatibility with every client.
2. A large context window does not require sending every file
Input can include system instructions, history, code, tool definitions and tool results. Select relevant files and necessary context before comparing windows. For unlisted parameters, consult model-specific references rather than filling in guesses.
Output limits and context capacity are separate. Input fitting in context does not mean a complete refactor fits in one answer; split tasks and check truncation.
3. Compare budgets for the same workload
Hold input/output token counts constant and switch models in the free calculator to compare base estimates. Then record actual task usage, tool rounds and completed results.
A cheaper model may need more retries or turns. One-request pricing is not the same as cost per completed task. Check cache, search and multimodal charges separately.
4. Build a small comparison with your own tasks
Choose common samples such as a bug fix, a refactor and a tool operation. Use the same prompt, materials and acceptance criteria for every candidate. Record:
- Exact model ID, protocol, date and client version.
- Test results, manual corrections and truncation.
- Input/output usage, actual costs and rounds needed to finish.
- Time to first text and complete task, including method and sample size.
Choose models that complete the task, then compare costs. One fast request does not demonstrate sustained availability, and a small sample does not rank all tasks.
5. Integration and troubleshooting
OpenAI SDK · Claude Code · Cherry Studio · Streaming troubleshooting