API parameter · output_config.effort
Claude Fable 5 effort parameter, explained
On Fable 5, effort replaced temperature as the one dial that matters. Five levels control how much reasoning the model spends — here is what each one does to your tokens, your latency and your bill.
TL;DR — 30 seconds
Effort is an API parameter (output_config.effort) that controls Fable 5's reasoning budget. Five levels: low, medium, high, xhigh, max. high is the default and the sweet spot for most work. Output tokens scale roughly 0.25×–1.9× vs high across the range — on one published measurement, the same prompt produced ~1,900 output tokens at low vs ~14,400 at max (about 7.6×). Input is billed identically at every level, so the cost gap is widest on output-heavy jobs. When a run comes back weak, retry at the same level with a sharper prompt before paying for a higher one.
The five levels, compared
Anthropic publishes the level names and the default (high) but no cost curve — the multipliers below are estimates anchored on one measured datapoint (see the note under the table). The token and cost rows use the representative workload defined in the cost section.
| Effort | low | medium | high | xhigh | max |
|---|---|---|---|---|---|
| Best for | Fastest and cheapest — routine transforms, classification, light chat | Balanced — day-to-day coding and drafting | The default — serious engineering, most production agents | Max depth — hard debugging, final reviews, Claude Code's default | The ceiling — hardest tasks, correctness over cost |
| Output tokens vs high | 0.25× | 0.5× | 1× | 1.45× | 1.9× |
| Output tokens / request | 5,000 | 10,000 | 20,000 | 29,000 | 38,000 |
| Monthly cost | $4,934 | $5,177 | $5,665 | $6,104 | $6,542 |
Multipliers are relative to high. Only the low→max ratio is measured (Simon Willison's pelican-SVG run, June 2026); the intermediate levels are interpolated. Token and cost rows assume the workload in the cost section below.
Rates and effort estimates verified 2026-08-04
What effort does to the bill
The same workload run at five settings — this is the case behind the token and cost rows above.
| Cost item | high (default) | max |
|---|---|---|
| Fresh input ($10/M) | $3,705 | $3,705 |
| Cache reads ($1/M) | $864.50 | $864.50 |
| Output ($50/M) | $950.00 | $1,805 |
| Safeguard fallback | $145.25 | $167.75 |
| Total | $5,665 | $6,542 |
Workload: 1,000 requests/month, each with 1M input tokens (×1.3 tokenizer inflation, 70% served from prompt cache) and 20K output tokens at high effort. 5% of requests are answered by the Opus 4.8 safeguard fallback. Prices are Fable 5's published $10/$50 per million tokens.
Effort moves the output column only — the input bill is identical at every level. At this mix, input and caching dominate, so the total moves only ~15% between low and max even though output tokens move ~7.6×. On output-heavy workloads (code generation, long reports, agent loops) the gap is much wider; on chat-heavy traffic it nearly disappears. That's why the expensive model at low effort is often the cheap option.
All figures above come from the same estimator used on the cost-calculator page, rounded for display. Anthropic publishes no effort cost curve — the 0.25×/0.5×/1×/1.45×/1.9× multipliers are interpolated from one measured anchor (Simon Willison's pelican-SVG run: ~1,900 output tokens at low vs ~14,400 at max on an identical prompt, June 9, 2026).
Retry first, upgrade second
A weak result is usually a prompt problem, not an effort problem. The order that costs the least:
Retry at the same level with a better prompt
Most failed or shallow runs come from an ambiguous spec, not insufficient reasoning. State the goal, the constraints, and what "done" looks like, then run the same level again — it is far cheaper than escalating.
If you must upgrade, move one level at a time
low → medium → high → xhigh → max. Each step buys more reasoning depth at the cost of latency and tokens; jumping straight to max skips the point where the marginal gain stops. xhigh is the practical ceiling for coding and agentic work — reserve max for rare, correctness-critical tasks.
Latency is part of the cost
At max, a single Fable 5 request can run for many minutes. Two retries at high can finish faster than one run at max — wall-clock time is a real budget for interactive work.
Check the bill before escalating
Output tokens scale with effort, and output is the expensive direction ($50/M on Fable 5). On input-heavy workloads the difference between levels is small; on output-heavy ones it is not. Two medium runs can beat one max run on both cost and coverage.
See it on your own workload
The cost calculator on this site runs the same estimator with your numbers — input mix, cache hit rate, effort level and fallback share.