Claude Opus 5 now costs three times as much as Opus 4.8 for the same workload, even though both models list the same price per token. The culprit is the default “adaptive thinking” feature, which silently consumes output tokens as hidden reasoning. Turning that feature off restores cost parity without hurting accuracy.

Why the price gap appears

Both versions charge $5 per million input tokens and $25 per million output tokens. In our benchmarks Opus 5’s bill was 3.1 × higher than Opus 4.8 when running identical prompts. The extra charge does not come from a higher list price; it is baked into the way Opus 5 accounts for its internal thought process.

What adaptive thinking does

When adaptive thinking is enabled, the model performs extra internal reasoning before emitting the final answer. That reasoning is counted as output tokens, even though it never reaches the user. On straightforward queries, anywhere from 42 % to 95 % of the billed output tokens are this hidden reasoning. A simple arithmetic problem that should produce a single-digit answer can therefore generate dozens of unseen tokens, inflating the cost dramatically.

The feature is not a bug—Claude’s designers intend it to improve answer quality on complex tasks. In practice, the hidden reasoning often yields modest accuracy gains on hard prompts, especially when the model must chain multiple steps or interact with external tools.

How to control costs

The model accepts a single parameter that disables adaptive thinking:

{
  "thinking": {"type": "disabled"}
}

Applying this setting to Opus 5 brings its per-task expense down to exactly what Opus 4.8 charges, while the correctness of the results stays the same on the simple workloads we tested.

When to disable

  • Extraction of data from text
  • Formatting operations (e.g., JSON conversion)
  • Single-step calls where the answer is a direct lookup or a trivial calculation

Disabling thinking in these cases eliminates the hidden token tax and keeps bills predictable.

When to keep it on

  • Tasks that require deep logical chains or nuanced reasoning
  • Agent workflows that call external tools repeatedly

In tool-heavy scenarios the cost gap narrows to roughly 33 % because the model spends less on internal reasoning while looping through tool calls.

Other notable differences

  • Context window: Opus 5 can hold up to one million tokens, which we verified by retrieving a target string placed at token 969,950.
  • Cache floor: The minimum cache size dropped to 512 tokens, half of Opus 4.8’s floor, affecting how much prior conversation the model can reuse without re-tokenising.

What to watch next

Developers should monitor usage reports for “reasoning tokens” or similar line items that indicate adaptive thinking is active. As more applications adopt Claude for agent-style operations, the trade-off between cost and quality will become a key optimisation decision. Future updates may let users tune the depth of reasoning rather than an all-or-nothing switch, which could smooth out the cost curve.

Takeaway: Opus 5’s default adaptive thinking inflates token usage on easy tasks. Disable it with the simple setting above to match Opus 4.8’s cost, and only enable it when the extra reasoning justifies the extra spend.