Models optimized to think longer at inference defined 2025, with leaps in math contests and coding agents.
OpenAI reported o3 gains on hard benchmarks; DeepSeek-R1 and other open-weight models showed competitive, distillable reasoning traces.
Inference cost became a lever: more thinking tokens, more latency/cost, higher success on verifiable tasks.
That lever is the same philosophy labs would apply in 2026 to open math problems with formal verification.
DeepSeek released competitive reasoning models with reportedly low training costs, altering market expectations.
The o1 series introduced inference with internal “chain of thought” optimized by RL, excelling at math and code.
Frontier systems combined diagrams, PDF papers, and code in one tool-verified reasoning thread.