TokenLab Model Picker

Turn an outcome into a short list of live TokenLab model IDs. Work in three steps: outcome → operation → model. Read live catalog signals every time; never answer from memory.

What this skill delivers

• The operation the outcome needs, and the endpoint for it.
• A shortlist with exact public model IDs: a default, an alternative, and a fast or low-cost option.
• Why each model fits, from contract and catalog signals.
• One command the user can rerun to refresh the list.
• A caveat for anything volatile: price, supply, or a benchmark claim.

Inputs to collect

• The outcome in the user's words, and the inputs they already have (text, image, video, audio).
• Limits that matter: length, resolution, aspect ratio, voice, language, format.
• Priority: quality, speed, or cost, and any budget.
• For chat and coding: the client and the protocol it speaks.

Workflow

Step 1: outcome → operation

Outcome	Operation	Outcome skill
New image from text	text-to-image	tokenlab-photoreal-stills, tokenlab-product-photography, tokenlab-text-in-image, tokenlab-logos-and-vectors
Change an existing image	image-edit, image-to-image	tokenlab-edit-image, tokenlab-restore-and-upscale, tokenlab-character-consistency
Video from text	text-to-video	tokenlab-cinematic-video
Video from a still	image-to-video	tokenlab-animate-image, tokenlab-product-ad-video, tokenlab-talking-avatar
Video from reference images	reference-to-video	tokenlab-character-consistency, tokenlab-product-ad-video
Video between two frames	start-end-to-video	tokenlab-keyframe-transition
Change or continue a video	video-to-video, video-extension	tokenlab-edit-video, tokenlab-extend-video
Move a character like a reference clip	motion-control	tokenlab-motion-control
Speech from text / text from speech	tts / stt	tokenlab-voiceover, tokenlab-transcribe-and-subtitle
Song, instrumental, sound effect	music	tokenlab-music
3D asset / explorable world	text-to-3d, image-to-3d / world-generation	tokenlab-image-to-3d-asset, tokenlab-world-generation
Typed yes/no, choice, or score	systemone	tokenlab-decisions
Text translation	translation	tokenlab-transcribe-and-subtitle
Chat, coding, agent loop	chat formats	tokenlab-api-integration, tokenlab-native-endpoints

Step 2: operation → candidates

• Scene shortlist: GET https://api.tokenlab.sh/v1/models?recommended_for=<scene>. Scenes: image, video, music, 3d, tts, stt, embedding, rerank, translation, decision. If the API rejects a scene with invalid_recommended_for, use the category filter instead.
• Category filter: GET /v1/models?category=<category> (for example category=decision).
• Keep only models whose tokenlab.public_contract_summary.public_operations (or public_contract) includes the operation.
• A scene ranks models by their measured success rate in the last 24 hours, not by output quality. Read tokenlab.agent_preferences.<scene>: preferred_rank, success_rate_24h, sample_count_24h, and status. status: "insufficient_samples" means no ranking signal yet, not a bad model.
• Check supply per operation: tokenlab.deliveryAvailability.operations[<op>] shows whether Verified and Official supply exist.

Step 3: candidates → shortlist

1. For each candidate, read GET /v1/models/{model} and confirm:
   • lifecycle.stage is active;
   • the operation, its endpoint in request_endpoint_by_operation, and the inputs the user has (image, video, audio, reference count);
   • the size, duration, and resolution limits the user needs, from operation_constraints.
2. Read GET /v1/models/{model}/pricing and compare cost per output unit (image, second, character, token, or request).
3. Pick the default for fit and reliability, an alternative with a different strength (quality, speed, or input type), and a fast or low-cost option. Start from the outcome skill's generated Models list (refreshed daily from the live catalog), then confirm it live.
4. Chat and coding: read tokenlab.accepted_request_formats to match the client's protocol (openai_chat_completions, openai_responses, anthropic_messages, gemini_generate_content).
5. Decisions: discover category=decision and confirm the systemone operation. A decision model returns typed answers; never select it as a chat model.

MCP equivalents: list_models, get_model, get_model_pricing, compare_models (in every tool profile; the default profile is core).

Output format

• One sentence: the outcome, the operation, and your assumptions.
• A table: Role (default / alternative / fast), Model ID, Why, Watch out.
• One rerunnable command, for example curl "https://api.tokenlab.sh/v1/models?recommended_for=video".
• One line on what to confirm before a paid call (price, supply, limits).

Avoid

• Do not name a single universal best model, and do not hardcode one model as the only answer.
• Do not recommend a model that is absent from the live catalog or does not publish the operation.
• Do not infer an endpoint, field, or capability from a model name; read the contract.
• Do not invent prices, model counts, or benchmark results. A benchmark claim needs a cited source and date.
• Do not use physical route names or provider-prefixed names as model IDs.

Quality bar

• Every listed model is live, publishes the operation, and fits the user's inputs and limits.
• Each pick has a reason from catalog or contract data, not from the model name.
• The answer says when the list was checked and how to refresh it.

Edge cases

• The user asks for the cheapest: still state what the cheap model cannot do (inputs, length, resolution, audio).
• The catalog is unreachable: say so, and give the outcome skill's table as examples only, not as live truth.
• The user names a model: check that it publishes the operation; if not, say so and offer live alternatives.

Related skills

• tokenlab-run: run the chosen model.
• tokenlab-prompting: prompt craft for the chosen family.
• tokenlab-cost-routing: fallback chains with live pricing.