Model chooser
Say what you want to do and whether you do it once or again and again. You get one model, the reason, and what to use if you cannot use it. Checked against vendor docs, public leaderboards and Hugging Face.
Every recommendation
A one-off job wants the strongest model you can open in an app today. A recurring job wants the one that holds quality at a cost, speed and consistency you can live with for a year. They are rarely the same model.
| Task | One-off | Recurring |
|---|---|---|
| Build or fix software | Claude Opus 5 | Claude Sonnet 5 |
| Write copy and long-form content | Claude Opus 5 | Claude Sonnet 5 |
| Research a market, company or topic | GPT-6 Astra | Claude Opus 5 |
| Classify, tag or extract data | Claude Sonnet 5 | GPT-5.6 Luna |
| Read PDFs, contracts and scanned documents | Claude Opus 5 | Mistral OCR 4 |
| Run a customer-facing chat assistant | Claude Opus 5 | Claude Sonnet 5 |
| Automate multi-step work with an agent | Claude Fable 5.1 | Claude Opus 5 |
| Analyse spreadsheets and numbers | GPT-6 Astra | Gemini 3.8 Flash |
| Generate or edit images | GPT Image 2.5 | Nano Banana 2 (Gemini 3.1 Flash Image) |
| Generate video | Gemini Omni Flash | MiniMax H3 |
| Transcribe audio, calls and meetings | ElevenLabs Scribe v2 | ElevenLabs Scribe v2 |
| Generate voiceover and speech | ElevenLabs v3 | Cartesia Sonic 3.6 |
| Search and answer from your own content | Gemini 3.8 Flash | Gemini Embedding 2 |
| Translate and localise | Claude Opus 5 | Gemini 3.8 Flash |
Checked . Hugging Face figures fetched 17 September 2026.
Use it inside your agent
The same recommendations are served over MCP, and a skill tells Claude when to reach for them. Install both, then ask “which model should I use to transcribe our sales calls every week?” in Claude Code or Claude.
claude mcp add --transport http --scope user model-chooser https://gtmagentcy.com/api/mcp
mkdir -p ~/.claude/skills/model-chooser && curl -fsSL https://gtmagentcy.com/model-chooser/skill/SKILL.md -o ~/.claude/skills/model-chooser/SKILL.md
Add https://gtmagentcy.com/api/mcp as a custom connector in Customize → Connectors, then upload the skill zip under Skills.
Streamable HTTP at https://gtmagentcy.com/api/mcp. No auth, read-only. Tools: choose_model, get_recommendation, list_tasks. Speaks protocol 2026-07-28 and the older initialize handshake.
Read the skill before you install it. It only calls the three tools above.
How it picks
Quality first. You are one person with one job, so the price per call barely matters and the chat app is the interface. We pick the model that does the task best today.
Quality you can hold. The job runs hundreds of times, so cost per run, speed, a stable API, structured output and, sometimes, open weights you can host decide it.
Vendor model docs, public leaderboards such as LMArena, SWE-bench and Artificial Analysis, and live Hugging Face download and like counts for open models. Every page lists its links.
Wiring it into a workflow that runs every week, with evals, fallbacks and a cost you can predict, is the work. Fifteen minutes, no deck.