Model landscape and selection
Model rankings change monthly. Select models with your own evals against quality, latency, cost, context, tool use and deployment constraints.
Changes fast: check the latest updates too. Reviewed 17 Oct 2026.
Key ideas
- Frontier closed models: Anthropic Claude (Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5 as of Oct 2026), OpenAI GPT, Google Gemini.
- Strong open-weight families come from Meta, Mistral, Alibaba (Qwen), DeepSeek and others; useful for self-hosting, data residency and fine-tuning.
- Use a tiered setup: a small fast model for routing, classification and extraction, a frontier model for hard reasoning.
- Selection checklist: task eval score, p50/p95 latency, $ per 1K requests, context length, tool-calling reliability, structured output support, region availability.
- Keep a model registry with a pinned model ID per use case so upgrades are deliberate and evaluated.
Common mistakes
- Choosing from public leaderboards alone. Benchmarks are saturated and contaminated; your eval set decides.
- Using floating aliases (latest) in production.