Claude Opus 5.5 is cheaper per token. Check the cost per task first

Anthropic released Claude Opus 5.5 on September 22, 2026, at $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. Cache reads dropped from $0.50 to $0.20 per million. The model has a 1M-token context window and up to 128K tokens of output, and the pricing docs list batch at half price. Opus 5 stays available as a legacy model.

Anthropic says Opus 5.5 matches Fable 5.1 on most work and writes output more than 30 percent faster than Opus 5. The benchmark scores on the launch page, such as 66.4 percent on Terminal-Bench 4.0, are runs Anthropic did and reported itself. OpenAI launched GPT-6 Sol about 90 minutes later, which we covered separately.

Four request patterns that now fail

The migration guide lists changes that return a 400 error instead of a weaker answer:

• Thinking is always on. A request that disables thinking or sets budget_tokens is rejected.
• Any temperature, top_p or top_k value other than the default is rejected.
• Assistant prefill, where you write the first part of the model's reply for it, is rejected.
• Edited or reordered thinking blocks from earlier turns are rejected. Anthropic calls this preserved thinking.

Computer use on the Claude API and Google Cloud also needs the new computer_toolset_20260801 tool version. A wrapper that sets temperature to 0 for repeatable output, or prefills an opening brace to force JSON, fails on its first call. Finding those patterns takes one search. It also matches ordinary assistant turns in a conversation history, so expect some harmless hits:

grep -rnE "temperature|top_p|top_k|budget_tokens|\"role\": ?\"assistant\"" src/

The number Artificial Analysis added

Artificial Analysis, which runs its own evaluations, scores Opus 5.5 at 58 on its Intelligence Index at max effort, the highest it has measured. GPT-6 Astra and Fable 5.1 score 53. At medium effort, which is the default, Opus 5.5 scores 51. The same report counts output tokens: about 119,000 per task at max effort, against about 27,000 for GPT-6 Astra.

Output tokens are what you pay for. At $20 per million, 119,000 tokens is about $2.38 of output per task. Astra's 27,000 tokens at $50 per million is about $1.35. A lower price per token does not make a task cheaper on its own. Artificial Analysis puts Opus 5.5's cost per task level with Opus 5, so at max effort the price cut goes into more thinking, not a smaller bill.

Why a build studio cares

Switching to Opus 5.5 is a code change, not a new model string. Run the search above, fix what it finds, and run the test suite against the new model ID before any client traffic moves. Then measure cost per completed task on your own prompts at the effort level you plan to ship. The headline 58 comes from max effort, and max effort is where the token count climbs. Medium is the default, and for most production calls it is the setting to measure first.

Next step: read the migration guide before you change the model string. If you want a codebase checked for these patterns before a switch, write to us at hello@gattyworks.com.