Anthropic, OpenAI and Google each offer a range of models, from small, fast and cheap to large and powerful. New versions arrive every few months, so rather than chasing the latest name, it helps to know how to choose.
Model tiers
Each provider's lineup follows the same pattern:
- Small models (such as Claude Haiku, GPT-6 Luna and Gemini Flash-Lite) are fast and very cheap. They're good at classification, extraction, short summaries and high-volume jobs.
- Mid-size models (such as Claude Sonnet and Gemini Flash) balance quality and cost. They're good everyday choices for chatbots, writing help and most coding.
- Large models (such as Claude Opus and Fable, the top GPT-6.2 models and Gemini Pro) handle the hardest reasoning, complex code and long multi-step tasks, at a higher price.
Prices compared
Prices are per million tokens, checked on 19 November 2026. They change often, so check the providers' pricing pages before you commit.
| Model | Input | Output |
|---|---|---|
| Claude Haiku 5.5 (prompts up to 100k tokens) | $0.10 | $0.50 |
| GPT-6 Luna | $0.10 | $0.50 |
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 |
| GPT-6.1 Terra | $1 | $5 |
| Gemini 3.5 Flash | $1.50 | $9 |
| Claude Sonnet 5.5 | $2 | $10 |
| GPT-6.2 Sol | $2 | $10 |
| Gemini 3.1 Pro Preview | $2 | $12 |
| Claude Opus 5.5 | $4 | $20 |
| Claude Fable 5.2 | $8 | $40 |
| GPT-6.2 Astra | $8 | $40 |
To see what these mean for your app, enter your own numbers in the AI API cost calculator.
What to weigh up
- Quality on your task. Public benchmarks give a rough idea, but the only test that really counts is your own examples.
- Cost. Remember that output tokens cost more than input tokens, and that caching and batch discounts can change the picture. See how to cut AI API costs.
- Speed. Smaller models reply faster, which matters for chat and voice.
- Context window. Make sure the model can take the amount of text you need. See context windows explained.
- Features. Check for what you need, such as image input, tool use for agents, or structured JSON outputs.
- Privacy and location. Read each provider's data policy, especially for personal or business data. For full control, consider an open model you run yourself, even on a Raspberry Pi.
A simple way to decide
- Collect 20 to 50 real examples of what your app will do, including tricky ones.
- Run them through a small model and a mid-size model, with the same prompt.
- Score the answers, ideally without knowing which model wrote which.
- Pick the cheapest model that's good enough, and keep your examples so you can re-test when new models come out.
Keep the model name in one place in your code, so switching later is a one-line change.
Going deeper
Build a small evaluation
A spreadsheet is enough to start: one row per test input, a column per model, and a score for each answer. For tasks with a clear right answer, score automatically. For writing quality, use a fixed rubric and, ideally, hide which model wrote each answer. Another model can grade against the rubric, but spot-check its grades by hand, as automated graders have their own biases.
Latency, not just quality
Measure two things: time to first token (how long before text starts appearing) and output speed (tokens per second). For chat and voice, time to first token matters most. Reasoning models may "think" for seconds before replying, so turn the effort down, or choose a non-reasoning setting, for latency-sensitive features.
Total cost, not price per token
A model that's cheaper per token can cost more overall if it needs longer prompts, produces longer answers, uses many thinking tokens, or fails more often and needs retries. Compare the real cost per successful task from your eval runs.
Plan for change
- Keep the provider behind one function in your code, so switching models or providers is a small change.
- Set up a fallback to another model for outages or rate limits, if your app can't afford downtime.
- Watch for deprecations. Providers retire older models on a published schedule, so pin specific model versions in production and re-run your eval before upgrading.
Open-weight models
Open-weight models can be downloaded and run on your own hardware or on cloud services. They give full control over data and no per-token fees on your own hardware, at the cost of running the infrastructure yourself. They're attractive for privacy-sensitive data, offline use, or very high volumes of simple tasks. See running AI on a Raspberry Pi for a small-scale start.
Compliance
For business use, check where data is processed and stored, how long it's retained, whether it's used for training, and what certifications the provider holds. These terms differ between consumer apps and API plans, and sometimes between regions.
Sources
The main references behind this guide. Numbers we worked out ourselves, such as the code outputs, come from running the code.