Model names change every month; the categories stay the same. Use this sheet to decide what kind of model you need before comparing brands. Browse candidates in our Models Directory.
| If you need to… | Use… | Why |
|---|---|---|
| Chat, write, summarize, general help | A general-purpose LLM | Best all-round quality |
| Solve math, logic, complex coding or planning | A reasoning model | Spends extra compute “thinking” before answering |
| Handle millions of simple requests cheaply | A small/fast model | Lower cost and latency |
| Keep data on your own servers | An open-weight model, self-hosted | Full control over data and deployment |
| Search documents by meaning | An embedding model | Turns text into vectors for similarity search |
| Understand images, charts or screenshots | A multimodal/vision model | Accepts images as input |
| Create images | A diffusion/image model | Purpose-built for visual generation |
| Transcribe or generate speech | A speech model (STT/TTS) | Optimized for audio |
| Autocomplete and refactor code | A code-specialized model or coding assistant | Trained heavily on code |
Quick rules
- Start with a strong general model to see if the task is feasible at all.
- Then go smaller until quality drops — the cheapest model that passes your tests wins.
- Use reasoning models selectively: they are slower and costlier per answer.
- Prefer retrieval over bigger models when the problem is missing knowledge.
- Re-test when models update: providers change models behind the same name less than they used to, but new versions arrive constantly.
Questions to ask any provider
- What happens to my data, and is it used for training?
- What is the context window and the price for input vs. output?
- What are the rate limits?
- When will this model version be retired?
Matching tasks to model types
| Task | Usually a good fit |
|---|---|
| Everyday writing, Q&A, summaries | A capable general chat model |
| Hard math, logic, complex coding | A reasoning model |
| High-volume classification or extraction | A small, fast, cheap model |
| Private data that can’t leave your servers | An open-weight model you host |
| Images, audio or video input | A multimodal model |
| Search over your documents | Embedding model plus a chat model (RAG) |
| Real-time voice conversation | A low-latency speech or multimodal model |
A quick decision process
- Define the task and what “good” looks like.
- Shortlist two or three models of the right type based on capability, price and data requirements.
- Test on 20–30 real examples and score the outputs.
- Compare cost and speed at your expected volume.
- Check data terms and compliance for your use case.
- Choose, monitor and re-test when providers release new versions.
Cost sanity check
Estimate monthly tokens: number of requests × (input tokens + output tokens). Multiply by each model’s price per million tokens. A model that is slightly better but ten times more expensive is rarely the right choice for high-volume tasks — but may be worth it for rare, high-value ones.
Red flags when choosing a provider
- No clear information on how your data is used or retained.
- No way to pin a model version, so behavior can change without notice.
- Unclear pricing for long inputs, images or tool calls.
- No status page, support channel or service-level commitments for business use.
- Benchmark claims without details on how they were measured.
Related reading
For deeper background, read our quick take on why benchmark scores won’t pick your model and our guide to building an evaluation set before you commit.
What to watch next
- Price changes and new model tiers from major providers.
- Smaller models closing the gap with large ones.
- Model routing tools that pick the right model per request.
- Data residency and compliance options for enterprise models.
Key takeaways
- Match the model type to the task: general, reasoning, small, open-weight or multimodal.
- Test candidates on your own examples rather than relying on leaderboards.
- Compare cost and speed at your expected volume.
- Check data handling terms before sending sensitive information.
- Re-evaluate when providers update models.
Frequently asked questions
Should I always use the most powerful model?
No. Smaller models are often good enough, faster and much cheaper for routine tasks.
When are open-weight models the best choice?
When you need data control, customization, offline use or predictable costs at scale.
How do I compare models fairly?
Use the same prompts and real examples, score outputs blind if possible, and include cost and speed.