Claude Opus 4.6: When Adaptive Thinking Became the Default

Anthropic released Claude Opus 4.6 on February 5, 2026 — the release that made adaptive thinking the recommended way to control reasoning. Instead of setting a fixed token budget for "thinking," you let the model decide how much to reason and steer it with an effort level. It's a small API change with a big practical effect.

What changed

• Adaptive thinking (recommended). No more budget_tokens to tune — the

model dynamically decides when and how much to think, and interleaves reasoning between tool calls automatically.

• The effort parameter (GA). low | medium | high | max controls thinking

depth and overall token spend — the single most useful cost-quality dial.

• 128K max output and a 1M context window, at standard pricing.

Why it mattered

This was the moment "how much should it think?" stopped being a number you guessed and became a dial you tune per workload. For production teams, effort is one of the most direct levers on cost — a point we explore in LLM Cost Optimization. Opus 4.6 also followed instructions far more closely than earlier models, which meant prompts written to overcome old reluctance ("CRITICAL: you MUST…") often started to over-trigger and needed softening.

How Internative uses it

We build on the latest models through our AI Studio, tuning effort per workload to balance quality and cost rather than over-spending by default. Talk to our team to get the model and settings matched to your use case.