Claude Code with a model you can choose
Watch the case videoKeep the coding workflow you already use and connect a supported model through SeedRouter. Give Claude Code a concrete project task, inspect the resulting change and compare the full session before choosing a model for everyday work.
For developers and teams connecting their existing tools.
What can Claude Code do with SeedRouter?
Claude Code is a coding agent that works with project files and development tools. SeedRouter supplies the model connection through documented API settings, while the coding client handles file access and command permissions. Use that separation to compare models without redefining your entire development process.
Understand an unfamiliar part of a project
Ask Claude Code to trace one behavior through the files that implement it. Require file references and a short explanation of the data flow, giving you a useful starting point before a bug fix or a larger change.
Explore this use caseMake a change with a visible result
Give Claude Code an expected behavior and a bounded area to edit. Inspect the patch, run the relevant check and compare what actually changed with the request instead of judging the session only by how quickly it answered.
Explore this use caseCompare models on the same assignment
Run Claude Code against an equivalent project state and task brief when comparing model options. Keep the acceptance criteria fixed so you can distinguish fewer revisions from a shorter but incomplete attempt.
Explore this use case
Connect Claude Code with the right settings
The SeedRouter setup guide uses ANTHROPIC_BASE_URL and ANTHROPIC_API_KEY, then selects an exact model identifier. Apply the settings in the environment where Claude Code starts, and review existing model aliases before opening a session.
- Use https://api.seedrouter.ai as the documented base address.
- Store the Claude Code credential privately, outside shared project files.
- Keep model-specific setup notes attached to the model and client version tested.
Claude Code connection map
- API base
- https://api.seedrouter.ai
- Credential variable
- ANTHROPIC_API_KEY
- Example model ID
- claude-fable-5-1
- Client selection
- Choose the exact model when starting the session
Check existing shell variables and aliases so background work uses models you actually intend to select.
Use the recorded tests as a starting point
Our linked guide records multi-step file-reading sessions with Claude Code 2.1.280 on September 27, 2026. Those results show what was checked with Claude Opus 5.5 and Claude Fable 5.1 at that time; your project still needs its own task-level check.
- Read the historical test scope before treating a model as interchangeable.
- Ask Claude Code to report actual command output when it runs a check.
- Inspect changed files for unrelated edits and missed requirements.
Claude Code evidence record
- Recorded client
- Claude Code 2.1.280
- Recorded date
- September 27, 2026
- Observed task
- Multi-step work with project file reads
- Your next check
- One local change with a known expected result
A historical pass is not a guarantee for every newer client release or a different model combination.
Give Claude Code an acceptance brief
Download the runnable Python fixture: a catalog helper groups items correctly but crashes on an empty list. Give Claude Code the included task, then inspect its patch and run the supplied checks. The baseline was checked locally; this example has not been executed through a model.
- Write the expected behavior before the agent edits the file.
- Keep Claude Code inside the task boundaries your project needs.
- Save the patch and check output with your model comparison notes.
Claude Code brief: a boundary case
- File / function
- catalog.py · group_by_category
- Expected empty result
- An empty mapping: {}
- Baseline check
- Empty input fails; existing grouping passes
- Run locally
- python3 check.py
The download intentionally fails. Compare your Claude Code patch with the separate authored repair and identical checks linked below.
Choose a model for Claude Code
Use Claude Code model guides to narrow the options, then compare the work on your project. Current prices belong on the live model pages; the table below shows the documentation to inspect, without claiming that one model wins every coding task.
| Starting point | Reference to inspect | Task-level question |
|---|---|---|
| Claude Code with Claude Fable 5.1 | The Claude setup guide and model contract | Does it complete the change within the requested scope? |
| Claude Code with Claude Opus 5.5 | The recorded file-reading tests and current rates | How many revisions does the accepted result need? |
| Historical Kimi K3 experiment | The dated SeedRouter configuration and test notes | Not supported by Anthropic; review the tested scope and unverified client features. |
Check a complete coding session
A useful Claude Code trial ends with an inspected artifact. Start in a project where you know the expected behavior and can run a relevant check, then record both the result and the cost of getting there.
- 1
Choose a small, checkable task
Pick a bug or change with a clear completion condition. Give Claude Code the relevant files, project instructions and existing test command, keeping unrelated cleanup outside the assignment.
- 2
Confirm the active model and execute
Check the connection settings, then run the bounded assignment. Inspect Claude Code tool results and file changes during the session so a successful response does not hide an unexecuted command or missing edit.
- 3
Review the outcome and usage
Compare the diff with the request and verify the recorded check output. Include all Claude Code calls and follow-up revisions in your cost comparison, then decide whether the model suits similar work.
Claude Code connection questions
Does Claude Code use my Claude subscription here?
With the configured SeedRouter key active, these model calls use your SeedRouter balance. Claude Code subscription access and API credentials are different arrangements; verify the credential selected by the client rather than assuming that a saved subscription login pays for the session.
Can every text model run in Claude Code?
Anthropic supports Claude Code with Claude models and does not support non-Claude models through a gateway. SeedRouter’s linked Kimi guide describes a dated, limited experiment; it does not establish official support or parity for features such as deferred tools. Start with the documented Claude setup, then check your installed client and assignment.
Why can a short Claude Code task use many tokens?
The session includes instructions, tool descriptions and project context in addition to your prompt. Claude Code may make several model calls while reading, editing and checking files, so evaluate total session usage rather than estimating cost from the length of the final answer.
Does prompt caching work with Claude Code?
The September 27 test record observed cached prompt reads after the first request in multi-step sessions. Check current model pricing and your actual Claude Code usage records for the relevant billing lines; that historical observation is not a fixed saving promised for every run.
How do I verify a Claude Code change?
Inspect the actual file diff and the result of a check that covers the requested behavior. Ask Claude Code to distinguish commands it ran from checks it only recommends, and keep any unresolved failure visible when deciding whether to accept the patch.
What does the Claude Code download contain?
The Claude Code case ZIP contains a small Python helper, its failing empty-input check, a passing grouping check and a bounded assignment. Run it locally with Python 3 before using your configured client. The separate JSON brief records setup and evidence fields; neither file configures credentials or starts a paid session.
Setup guides and the recorded evidence
The September 27 guide records historical Claude Code tests; the gateway documentation explains connection settings. For this original catalog task, a separate authored repair passes both local checks while the unchanged baseline fails on empty input. Download the diff and test receipt to inspect that comparison. It is not output from an evaluated Claude Code session.
- Claude Code runnable source case
- Claude Code original and reference repair package
- Claude Code reference repair and scope
- Claude Code reference patch
- Claude Code before-and-after test receipt
- Claude Code setup and recorded tests
- Dated Kimi coding experiment
- Claude Code gateway documentation
- Current model pricing
Bring the next small change in your project
Choose a supported model and a task with a known check. Use Claude Code to produce the patch, review what it actually did and make your model choice from the complete session.
