Claude Code ships a per-model table of what each effort level costs
retained text
Hidden inside the Claude Code binary, `effort_cost_index` provides the following per-model token-usage estimate for the same task. It is normalized so `high = 1`.
|Model|low|medium|high|xhigh|max|xhigh to max|
|:-|:-|:-|:-|:-|:-|:-|
|sonnet-5|0.47|0.74|**1**|2.41|5.59|**+132%**|
|opus-4-8|0.72|0.90|**1**|1.65|1.88|\+14%|
|opus-5|0.67|0.76|**1**|**1.60**|**1.70**|**+6%**|
|fable-5|0.60|0.77|**1**|1.74|1.91|\+10%|
(fable-5-1 and mythos-5-1 were re-tuned between the two builds: 1.38/1.74 in the newer one.)
**Opus 5 flattens hard at the top.**
* `high -> xhigh` is +60%, then
* `xhigh -> max` is only **+6%**.
Sonnet 5 does the opposite and more than doubles at that same step.
So "max" means quite different things depending on model. On Opus 5 it's a small increment over xhigh. On Sonnet 5 it's genuinely the expensive tier.
# Tokens per Attempt, Not by Success
The UI string is, verbatim:
>`{from} effort uses more tokens per task than {to}.`
That's **consumption per task attempt**. It does *not* say the task gets completed, or completed correctly, or that the same amount of work happens at each level. It also doesn't price retries — a level at 0.76x per attempt isn't 0.76x cheaper if it needs a second attempt.
# How it held up when I measured it
Ran 8 agents on Opus 5 (2 real tasks x 4 effort levels), same repo, same base commit, isolated clones:
|Step|Catalog predicts|Observed (output tokens)|Observed (total incl. cache)|
|:-|:-|:-|:-|
|medium to high|x1.32|x1.17|x1.01|
|medium to xhigh|x2.11|x1.78|x2.53|
|medium to max|x2.24|x1.78|x1.93|
Direction right, magnitudes loose. But it nailed the thing that mattered: it predicted xhigh and max would be near-identical in cost, and on output tokens they were — 397,827 vs 397,761. Three digits apart.
Two caveats on my numbers: cache reads are 99%+ of the totals, so the "total" column is really measuring how many turns each agent took. And n=2 tasks, so treat it as a sanity check, not a benchmark.
I couldn't find any documentation of how the catalog numbers were derived or over what workload. Treat it as a vendor estimate.document fetched 2026-09-19 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1wjr89j/claude_code_ships_a_permodel_table_of_what_each