Here's the honest shortlist, ranked by how often we actually reach for them in our own projects.
1. Claude Opus 4.x — the reasoning workhorse. Best for long-horizon tasks, refactors, anything where the model has to hold a plan across a 40-step task.
2. GPT-4.1 / GPT-4.1 mini — the generalist. Great default for products shipped to non-technical users. Fast, cheap, predictable.
3. Gemini 2.5 Pro — the multimodal one. Vision, audio, huge context, and Google's native tooling advantage if you're already in that ecosystem.
4. DeepSeek R1 and V3 — the value pick. Frontier-ish performance at a fraction of the price. The model that made the rest of the industry nervous.
5. Grok 4 — the tool-user. Very strong at function calling and agentic loops.
6. Llama 3.x (open weights) — the self-hoster. Bring your own GPU, keep your data, pay only for electricity.
7. Mistral Large 3 — the European open frontier. Respects the 90-day retention promise, strong on French and German.
8. Qwen 2.5 and Qwen3 — the iron workhorse. Multilingual, cheap, and surprisingly strong on code.
9. Kimi K2 — the long-context specialist. Handles whole codebases in one pass.
10. gpt-oss-120b — OpenAI's open-weights play. Fine-tuneable, permissive license, legitimately good.
When in doubt: pick the model with the tool your team already uses, and switch only when measurements say so — not because Twitter said so.