Opus 5.5 effort levels are the real price
Updated September 22, 2026 · first published September 22, 2026
Claude Opus 5.5 lists at $4 per million input tokens and $20 per million output tokens. That is 20% below Opus 5. The rate card is the smaller half of the cost story. The larger half is the effort setting, which decides how many tokens the model spends thinking before it answers.
One model, five prices
Opus 5.5 exposes effort from low through medium, high, xhigh and max. Each level is the same model at the same per-token rate, but it produces a very different number of tokens. Independent benchmarking reported by The Decoder found that four of the five levels sit on the cost–performance Pareto frontier. It also found that at max effort Opus 5.5 used about 119,000 output tokens per task, far more than competing models at their top settings.
At $20 per million, 119,200 output tokens is $2.38 per task before any input. At medium, Artificial Analysis measured 25,700 output tokens, or $0.51. With input counted, max costs $5.98 per task and medium $1.34: the same model is about 4.5× more expensive per task depending on one request parameter.
Measured on the Artificial Analysis Intelligence Index v4.3.2 (ten evaluations), read 22 September 2026. Every row uses the same benchmark suite, so the rows compare directly.
| Model and effort | Intelligence Index | Cost per task | Output tokens per task | On the cost–performance frontier |
|---|---|---|---|---|
| Opus 5.5 · low | 42.3 | $0.55 | 10,200 | No |
| Opus 5.5 · medium | 51.2 | $1.34 | 25,700 | Yes |
| Opus 5.5 · high | 53.6 | $1.82 | 35,600 | Yes |
| Opus 5.5 · xhigh | 56.0 | $3.46 | 65,700 | Yes |
| Opus 5.5 · max | 57.6 | $5.98 | 119,200 | Yes |
| Opus 5 · high (reference) | 48.1 | $3.61 | 46,200 | No |
Source: Artificial Analysis. On the frontier means no setting of GPT-6 Sol, GPT-6 Astra, Opus 5, Opus 5.5 or Fable 5.1 scores higher for less. Opus 5.5 low is off it because GPT-6 Sol at xhigh scores 44.1 for $0.53. Going from xhigh to max adds 1.6 points for 73% more cost.
Medium is the setting Anthropic benchmarks against
Anthropic's headline comparison is made at default effort, which is medium. It says Opus 5.5 at medium beats GPT-6 Astra at max effort on knowledge work for about a fifth of the cost per task. So Anthropic's own case for the model is built on medium, and max is the exception. A team that sets max everywhere "to be safe" pays for tokens the launch benchmarks did not need.
How to set effort without guessing
- Default to medium. Make it the explicit value in code, not an SDK default that can change under you.
- Raise effort per route, not globally. Hard debugging, long-horizon planning and final review steps can justify high or max. Classification, extraction and routine edits rarely do.
- Log effort with every request. Cost per request is not interpretable without it. A spend spike with a flat request count is usually an effort change.
- Compare effort levels on the same replayed tasks. Record success rate and output tokens per completed task at each level, and stop raising effort when success stops improving.
- Cap output where the product allows it. A max-token limit bounds the worst case. At max effort that bound is the only thing between one runaway task and a several-dollar request.
The cut from $25 to $20 per million output tokens saves 20%. Moving a workload from max to medium effort can save far more than that. Treat effort as a budget line, owned and reviewed like any other.
Related
Want this applied to your own LLM spend? FinOps LLM runs a free audit of your AI costs and shows where the savings are. Book free audit →