Register Model

Register New Model

The model registry (src/lib/model-registry-data.json) is the single source of truth for model behavior. All behavioral lookups (isStrictSamplingModel, isAnthropicAdaptiveThinkingModel, supportsXhighEffort, supportsMaxEffort, getDefaultMaxOutputTokens) read from it. Registering a model is a data-only change in the common case — do NOT touch src/lib/utils.ts.

Sync with models.dev

For a full registry refresh ("sync the registry with models.dev"), do NOT follow the single-model workflow below. Instead:

A. Fetch the full dataset

curl -sSL --max-time 60 "https://models.dev/api.json" -o /tmp/models-dev-api.json

B. Compare with the registry using OFFICIAL-lab priority

models.dev contains 180+ labs; third-party mirrors (snowflake-cortex, nano-gpt, qiniu-ai, llmgateway, helicone, vivgrid, opencode, opencode-go, greenpt, alibaba-cn, etc.) often report different context/output/temperature values than the vendor. Always compare against the vendor's own lab first:

anthropic, openai, google, deepseek, xai, mistral, moonshotai, alibaba, zhipuai, minimax, xiaomi, kimi-for-coding

Only fall back to a third-party lab when the vendor lab no longer lists the model (e.g. deprecated models like claude-opus-4, gpt-5-codex, o1-mini).

Note: labs are case-sensitive — the MiniMax vendor lab uses MiniMax-M2 while the registry uses lowercase minimax-m2. Normalize IDs when matching.

C. Apply updates

1. Spec updates: set contextWindow / maxOutputTokens from the vendor lab's limit.context / limit.output. Ignore third-party differences. When a vendor drops a model, use the best available source or leave it. Ambiguity: the registry may alias a newer model under an older id (e.g. qwen-max has alias qwen3-max) — align to the spec the alias points to.
2. strictSampling: set "strictSampling": true for EVERY entry whose main provider shows temperature: false — this includes GPT-5.x (except gpt-5.3-chat-latest, which is temperature: true), the o1/o3/o4 series, Kimi K2.5/K2.7/K3, and new-gen Claude (Sonnet 5, Opus 5, Fable 5).
3. New models: confirm scope with the user first (add all new chat/code models vs. only flagship models). Apply the single-model workflow below per model (official-lab spec, standard reasoning profile, alphabetical insert).
4. Removals: entries no longer present in models.dev at all (not even via alias) are deprecated — confirm with the user before removing, since user configs may still reference them. Removing requires cleaning up test assertions that reference those ids (e.g. claude-jupiter-v1-p).
5. Efforts: reasoning_options may be a dict OR a list of {type: toggle|effort|budget_tokens} objects. toggle means reasoning is optional → include none. Use the effort values when present (e.g. kimi-k3 → ["none", "low", "high", "max"]).

D. Update tests + verify

• Add/remove model ids in src/lib/utils.test.ts describe blocks (isStrictSamplingModel, isAnthropicAdaptiveThinkingModel, supportsMaxEffort, supportsXhighEffort, getDefaultMaxOutputTokens) — flip assertions like isStrictSamplingModel('gpt-5.2') when strictSampling changes. Watch for spec-value assertions (e.g. gemini output 64000 → 65536).
• src/lib/model-registry.test.ts: add non-standard effort coverage (e.g. kimi-k3).
• src/components/models/ModelDialog.test.tsx: the strict-sampling extraArgs test uses a strict model id — repoint it if that model is removed.
• Run npm run check:all and fix any failures.

Workflow

1. Fetch the model spec from models.dev

Fetch the full dataset (more reliable than the model page):

curl -sSL --max-time 30 "https://models.dev/api.json" -o /tmp/models-dev-api.json

Then extract the entry for the target model. The Anthropic lab uses a nested models map; other labs use a flat id -> spec map:

python3 -c "
import json
data = json.load(open('/tmp/models-dev-api.json'))
# Search all labs for the model ID
for lab, models in data.items():
    models = models.get('models', models) if isinstance(models, dict) else {}
    for mid, m in models.items():
        if '<model-id>' in mid:
            print(lab, mid)
            print(json.dumps(m, indent=2))
"

If the model is not in the API data, it may be a hypothetical/requested model — confirm with the user before proceeding.

2. Map spec fields to the registry entry

From the models.dev entry, extract:

Determine platform from the lab: