Which Model for Which Job¶

In one line: Match the model tier to the task (small and fast for high-volume simple work, mid-tier for most products, the frontier tier for hard reasoning and long agent runs, open weights when data or cost demand it) and confirm the choice on your own eval set, not on leaderboards. Last reviewed: 2026-10-21

The short version¶

• Every major provider sells a ladder: a small fast model, a mid-tier workhorse, and a frontier model. Prices across one ladder differ by 20x to 100x as of 2026-10.
• Default to the mid tier for a new product, then move individual routes down (to save money) or up (to fix quality) based on evals.
• Small models handle classification, extraction, routing and short replies very well. Frontier models earn their price on hard reasoning, complex code and long autonomous agent runs.
• Reasoning effort is a second dial. The same model at lower effort can be much cheaper and faster; test that before switching models.
• Open-weight models make sense for privacy, offline use, fine-tuning or very high steady volume. They trail the best closed models but the gap is small for many tasks.
• Model names below were current on 2026-10-21. They change every few months. The method on this page does not.

The current ladders (as of 2026-10)¶

Prices are list prices per million tokens, input / output, standard tier, from each provider's own pages. Batch and cached input are cheaper.

Notes: Claude Haiku 5.5's price rises to $0.50 / $2.50 for prompts over 100,000 tokens. Gemini Pro-tier prices rise for prompts over 200,000 tokens. Google offers a free API tier on its Flash and Flash-Lite models. All of these accept text and images; check each provider's docs for audio, video and tool support. For the lab histories behind these names, see Claude, GPT and Gemini.

Open-weight families you can run yourself or rent from many hosts include Qwen, Llama, DeepSeek, Mistral, Gemma and OpenAI's gpt-oss. See open vs closed weights, Qwen, Llama, DeepSeek, Mistral and local LLM runners.

Job-by-job guidance¶

For why reasoning effort matters so much, see reasoning models. For the cost arithmetic, see cost and latency control.

How to choose, properly¶

flowchart TD
    A[Write 30-100 real test cases with expected outputs] --> B[Run the mid-tier model]
    B --> C{Meets quality bar?}
    C -->|yes| D[Try the small tier and lower effort]
    D --> E{Still meets bar?}
    E -->|yes| F[Ship the cheaper option]
    E -->|no| G[Ship mid tier]
    C -->|no| H[Fix prompt and retrieval first]
    H --> I{Meets bar now?}
    I -->|yes| D
    I -->|no| J[Try strong default or frontier, higher effort]

1. Build the test set first. Real inputs, expected outputs, and a scoring method. See set up an eval harness.
2. Fix the prompt and context before buying a bigger model. Many "the model is not smart enough" problems are missing context.
3. Measure cost per completed task and latency, not just accuracy.
4. Re-run when models change. New releases arrive every few months; your test set makes switching a one-day job.

Why leaderboards are not enough¶

Public benchmarks are useful for a shortlist and misleading for a final decision. Scores saturate, test data leaks into training data, and benchmark tasks rarely look like yours. Read contamination and saturation, human preference arenas, knowledge and reasoning benchmarks and agent benchmarks.

Other factors that can decide it¶

• Where you can buy it. Claude is available through Anthropic and the major clouds (Amazon Bedrock, Google Cloud, Microsoft Foundry). OpenAI models are available through OpenAI and Microsoft Azure. Gemini through Google. Your cloud contract or data residency rules may decide for you. See GenAI platforms.
• Data terms and retention. Some models or features require specific retention settings. Check before committing.
• Rate limits at your volume.
• Model lifecycle. Prefer pinned model versions and check the provider's deprecation schedule; Anthropic publishes minimum retirement dates for its current models.
• Context length if you process very long documents. Current Claude models offer 1M tokens as of 2026-10.

Common mistakes¶

• Using the frontier model for everything. You pay 20-100x for tasks a small model does just as well.
• Using the small model for everything. You save money and lose customers.
• Choosing from a leaderboard. Test on your data.
• Not pinning versions. Behaviour changes without a deploy.
• Ignoring the promotional price expiry. Some prices above are temporary; budget on the post-promotion price.

Next steps¶

• Build vs buy for the bigger decision
• LLM serving if you self-host
• Cost and latency control
• Chat assistants if you are choosing an app rather than an API

Sources¶

• Anthropic, Models overview (IDs, descriptions, default effort, 1M context, retirement commitments; as of 2026-10): https://platform.claude.com/docs/en/about-claude/models/overview
• Anthropic, Pricing (all Claude rates incl. Haiku 5.5 long-prompt pricing; as of 2026-10): https://platform.claude.com/docs/en/about-claude/pricing
• OpenAI, Models (GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna descriptions and prices; as of 2026-10): https://developers.openai.com/api/docs/models
• Google, Gemini API pricing (Gemini 4 Argon, 3.8 Flash, 3.5 Flash-Lite, 3.1 Pro Preview, promotional dates, free tier; as of 2026-10): https://ai.google.dev/gemini-api/docs/pricing