Claude Code ships a per-model table of what each effort level costs

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.