Author(s): Anna Jey
Originally published on Towards AI.
LLM Reasoning Budget
A reasoning model can feel brilliant on one task and painfully slow on the next. The model did not suddenly get worse. You probably gave the same thinking budget to a simple lookup, a tricky code fix, and a risky production decision.
That mistake is getting easier to make because reasoning controls are moving into normal developer workflows. Anthropic’s Claude Opus 5 release notes describe thinking on by default, a full effort ladder, and support for long-context agent work. OpenAI’s reasoning model docs explain that invisible reasoning tokens are billed as output tokens and consume context space. Google’s Gemini API changelog points in the same direction from the other side: newer Flash models are being tuned for token efficiency, lower latency, and agentic planning. GitHub is also making more models available inside Copilot, including Claude Opus 5 and Gemini 3.6 Flash.
