Index9

Compare 2-10 models side-by-side. Returns each model's full metadata plus a diff matrix highlighting which fields are equal and which differ. Use this when the user asks "which is cheaper / has more context / supports X" across multiple specific models. Faster than calling get_models and diffing yourself. Response: { models: ModelResponse[], diff: { contextLength, maxOutputTokens, promptPricePerMillion, completionPricePerMillion, tokenizer, inputModalities, outputModalities, capabilities, supportedParameters }, cheapestForPromptPerMillion, largestContext, missingIds, resolvedAliases?, ambiguousAliases?, suggestions?: Record<unknownId, Array<{id, name, created}>> (newest-first), missingDiagnostics? }. Each numeric/string diff field has { allEqual: boolean, values: Record<id, value|null> }. Capability/parameter diffs have { commonAll: string[], uniquePerModel: Record<id, string[]> }. cheapestForPromptPerMillion / largestContext are convenience picks across the supplied models — null when the field is missing on every model. Optional: pass `expectedPromptTokens` AND `expectedCompletionTokens` to also receive `workloadCosts` and `cheapestForRealisticWorkload` — the actual cheapest given the user's expected token mix. Each `workloadCosts[i]` carries `tokenCostUsd` (token-only), `requestCostUsd` (per-request fee), `totalCostUsd` (sum, includes request fees), and `pricingBasis` ("exact_per_token" | "rounded_per_million" | "unavailable"). This matters when prompt:completion price ratios diverge across models, or when a model has a per-request fee. Accepts the same alias formats as get_models. Unknown ids are returned in missingIds (with `suggestions[id]` as newest-first `{id, name, created}` entries when partial matches exist, plus `missingDiagnostics` carrying a machine-readable reason per id). When fewer than 2 ids resolve, this returns 400 with the diagnostics so you can retry with `suggestions[id][0].id` for each missing id.

Filter 300+ AI models by structured constraints (capabilities, price, context, modality, provider). The semantic `q` is a soft tiebreaker on top of filters — trust filters first; weak `q`-only queries surface as `meta.confidence: "low"` with a hint to drop `q` and use `sortBy=created` or `sortBy=price` instead. Call this after list_facets (or directly when you already know the filters). Always include one call with sortBy=created (no q) when recommending models to the user — your training-data list of "good" models is likely stale, and sorting purely by price surfaces old tiny models and misses this month's cheap frontier. Extract filters from user queries. Numeric and categorical constraints MUST go in structured filters; `q` is for semantic flavor. Shorthand: 1K=1000, 1M=1000000. Prices are USD per million input tokens. Examples: - "1M context under $1" → q="model", minContext=1000000, maxPrice=1 - "cheap vision model from openai" → q="cheap vision model", capabilitiesAll=["vision"], provider="openai" - "function calling under $0.50 with 128K" → q="function calling", capabilitiesAll=["function_calling"], maxPrice=0.5, minContext=128000 - "cheapest paid model" → sortBy="price", sortOrder="asc", excludeFree=true - "best coding model" → q="best coding model" - "what's new" → sortBy="created" (no q needed) Valid capabilities: function_calling, structured_output, reasoning, web_search, vision, audio_input, file_input, image_generation, audio_output. Each result: id, name, description, created (unix seconds), createdAt (ISO 8601), contextLength, maxOutputTokens, pricing.{promptPerMillion, completionPerMillion} (rounded display $/M), pricing.{promptPerToken, completionPerToken, requestUsd} (exact, use for cost math), inputModalities[] / outputModalities[], capabilities[], score. With sortBy=price, results also expose pricing.effectivePromptPerMillion and pageInfo.priceSortBasis — sort order may diverge from displayed promptPerMillion for models with per-request fee…

Get full specs for one or more models by id or alias. Accepts up to 100 ids per call — use this for batch comparison. Call after find_models to inspect candidates, or directly when the user names a model (format: 'provider/model-name'). Response: { results: (Model | null)[], missingIds: string[], resolvedAliases?: Record<alias, canonicalId>, ambiguousAliases?: Record<alias, candidateIds[]>, suggestions?: Record<unknownId, Array<{id, name, created}>> }. Each non-null result has: - id, canonicalSlug, name, description - created (unix seconds), createdAt (ISO 8601), knowledgeCutoff (ISO date or null) - contextLength (tokens), maxOutputTokens, isModerated - pricing: { promptPerMillion, completionPerMillion, promptPerToken, completionPerToken, requestUsd, imageUsd } — *PerMillion is rounded display, *PerToken is exact (use for cost math). request/image are flat per-unit fees. - architecture: { inputModalities[], outputModalities[], tokenizer, instructType } - capabilities[]: normalized capability flags (same values as find_models and capabilitiesAll/Any) - supportedParameters[]: OpenRouter parameters the model accepts (e.g., "temperature", "tools", "response_format") Entries in results are null when the id is unknown; those ids appear in missingIds. Ambiguous aliases appear in ambiguousAliases with candidate canonical ids — pass a canonical id to disambiguate. Unknown ids that partially match (e.g. "sonnet" → all Claude Sonnet variants) appear in `suggestions` as up to 5 `{id, name, created}` entries **sorted newest-first** — pick `suggestions[id][0].id` for the most current replacement without a second lookup. When token-overlap finds nothing but the id is shaped like `provider/<unknown>` and the provider exists, suggestions falls back to the 5 newest models from that provider (real created timestamps, no hardcoded "popular" list). `missingDiagnostics` (when present) gives a machine-readable reason per missing id: `unknown_provider` (the prefix before / isn't in th…

Return the live vocabulary derived from the current model cache: providers (id-prefixes with model counts), capability flags (with counts), input/output modalities, and tokenizers. Use this once at the start of a session to learn what filter values find_models will accept, instead of trial-and-error against find_models. Response: { providers: { id, modelCount }[], capabilities: { id, modelCount }[], modalities: { input: string[], output: string[] }, tokenizers: string[], totalModels: number, updatedAt: string (ISO 8601) }. No inputs.

Run model tests on 1-10 models. Use dryRun=true to estimate token usage/cost, or dryRun=false (default) to run live OpenRouter inference. When dryRun=true: - No OpenRouter API key required - No inference call is made - Either `prompt` OR `expectedPromptTokens` is required (use the latter for capacity planning at sizes you don't want to paste in full) - expectedCompletionTokens defaults to 256 when omitted Parameters: - models: 1-10 model IDs to test (all receive identical prompts) - prompt: Prompt text (required for dryRun unless expectedPromptTokens is set; required for live unless userContent provided) - dryRun: If true, return cost estimates only - expectedPromptTokens: Estimated prompt-token count for dryRun cost estimation; overrides the prompt-string heuristic. Use to model "what would N-token requests cost?" without pasting N tokens. - expectedCompletionTokens: Optional completion token estimate used by dryRun - maxTokens, systemPrompt, temperature, topP, seed, responseFormat, enforceJson, retries: Live-testing controls (ignored when dryRun=true) - stream: Use OpenRouter's SSE streaming so capacity/refusal errors surface in ~1s instead of waiting the full per-model timeout for an empty 200. Defaults to false. - firstTokenTimeoutMs: Streaming-only deadline for the first delta. Defaults to 10s. If the upstream sends no token within this window, the request aborts and returns failureReason="timeout". Ignored when stream=false. - providerSort: "throughput" | "price" | "latency" — opt-in OpenRouter provider routing. Defaults to OpenRouter's load-balanced choice. - providerOrder: ordered list of provider slugs (up to 8). Try these providers first before falling back. Useful for steering around an overloaded provider for a single model. - fallbackModels: ordered list of model ids (up to 5). OpenRouter automatically retries the request against the next id when the primary is unavailable. Use sparingly — a benchmark should usually test the model you asked for, not…