Stop guessing which model to use
There are dozens of capable models and the “best” one depends entirely on your constraints. This picker turns that into five questions and gives you a concrete recommendation plus the reasoning behind it — so you can move on and build.
How it works
The tool evaluates your answers as a decision tree, applying the most binding constraints first:
- Privacy is checked before anything else. If data must stay on-premise, hosted APIs are off the table and the tool recommends self-hosted open-weight models.
- Task type comes next — hard reasoning points to dedicated reasoning models, code to Claude 3.5 Sonnet, and very long documents to Gemini’s huge context window.
- Output type routes image generation and vision-input tasks to multimodal models.
- Budget and volume then break ties between similarly capable options, favouring cheaper small-flagship models when cost or scale dominate.
Each recommendation comes with a one-paragraph rationale so you understand the trade-off rather than blindly trusting an answer.
Tips for using the recommendation
- Treat the result as a strong starting point, not gospel — the runner-up named in the rationale is often nearly as good and may suit your existing stack.
- If you are cost-sensitive and high-volume, the small-flagship models (GPT-4o mini, Gemini 1.5 Flash) almost always win; only escalate to premium models for the requests that genuinely need them.
- Validate the cost implication with the LLM Pricing Calculator before committing to a model at production volume.
How to interpret “task type”
The single most important question is what kind of cognitive work you are asking the model to do. A brief guide to the categories:
Reasoning / problem-solving. Multi-step logic, mathematics, code debugging, or any task where the model needs to think through a chain of steps before answering. These tasks benefit from dedicated reasoning models that spend tokens thinking before producing a response, at the cost of higher latency and token usage.
Long-document analysis. Summarising contracts, analysing reports, searching across a large knowledge base. The key constraint here is context window size — you need a model whose context is large enough to hold the document. Gemini’s very large context window is the current leader for extremely long inputs, while Claude’s context is competitive for most documents under a few hundred pages.
Code generation and editing. Claude 3.5 Sonnet and GPT-4o perform strongly on code benchmarks, particularly for realistic coding tasks and agentic coding workflows. For pure autocomplete in an IDE, a smaller, faster model often wins on latency.
Creative writing. Models vary considerably in their creative output style. Claude tends toward more varied and literary prose; GPT-4o has a particular style that works well for commercial content. For genuine creative work, trying your actual prompt on two or three models and comparing is more useful than any benchmark.
Classification or structured extraction. These are highly suitable for smaller, faster, and cheaper models — GPT-4o mini or Gemini Flash — because the task does not require deep reasoning, just reliable pattern matching and output structure.
The privacy dimension
If data must not leave your infrastructure — either for legal reasons (GDPR data-minimisation obligations, attorney-client privilege, healthcare data, export-controlled information) or contractual reasons (NDA-covered materials, financial data under confidentiality agreements) — then hosted cloud APIs are unavailable regardless of capability.
For on-premise or private-cloud deployments, the realistic options are open-weight models like Llama 3 (Meta), Mistral, or Falcon that can be run on your own hardware. The capability gap versus frontier API models varies by task — for some classification and extraction tasks, an open-weight model running locally is nearly as good; for complex reasoning and coding, the gap is currently meaningful.