Every team eventually faces the same question: "Would the bigger model be better?" The honest answer is usually "somewhat — but is somewhat worth 2x to 5x the price for this task?" That is not a philosophical question. It is a division problem, and you can answer it in an afternoon with an evaluation set and the published price list.
The price ladder you are climbing
Claude's current lineup spans a 10x price range on input tokens, and the list prices are the same across the Claude API and the cloud marketplaces (committed-use discounts are negotiated with your cloud provider separately).
| Model | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| Claude Haiku 4.5 | $1.00 | $5.00 |
| Claude Sonnet 5 | $3.00 ($2.00 intro through Aug 31, 2026) | $15.00 ($10.00 intro) |
| Claude Opus 4.8 | $5.00 | $25.00 |
| Claude Fable 5 | $10.00 | $50.00 |
There is a second, subtler dial: the effort parameter (low, medium, high — the default — plus xhigh and max on supported models). Effort shapes how many tokens the model spends on all response tokens, including internal reasoning, and reasoning tokens are billed as output tokens. So the same model at different effort levels has genuinely different cost per task, not just different latency.
Computing cost per quality point
The method has three steps, and none of them requires ML expertise:
1. Build a graded eval for the task. Fifty to a few hundred representative inputs with a pass/fail or scored rubric. Anthropic's own prompt-engineering guidance lists clear success criteria and empirical tests as prerequisites before you tune anything — the same artifact serves here.
2. Run the ladder. Run the identical eval on each candidate configuration — for example Haiku 4.5, Sonnet 5 at default effort, Opus 4.8 at default effort, Opus 4.8 at xhigh. Record two numbers per configuration: eval score and average billed cost per task (from the usage object in each response, multiplied by list prices).
3. Compare the marginal steps. For each step up the ladder, compute (score gained) ÷ (dollars added per 1,000 tasks). Somewhere on that curve the number collapses — you pay several times more per task for a score gain your users would never notice. That is your stopping point.
Reading the curve honestly
Three things distort this analysis if you let them. First, public benchmarks are not your task — a model that tops a coding leaderboard may add nothing to your invoice-classification job. Only your own eval counts. Second, output tokens dominate the bill for many workloads: output is 5x the input price on every current model, so a configuration that produces terser answers can beat a "cheaper" model that rambles. Third, quality problems are not always model problems. Anthropic's docs note that not every failing eval is a prompting problem — and equally, not every failing eval justifies a model upgrade. Re-run the ladder after any significant prompt change, because the crossover point moves.
Also remember which price you are measuring against. Claude Sonnet 5's introductory pricing ($2/$10 per million tokens) ends August 31, 2026; from September 1 it is $3/$15. If your marginal-dollar math was done on the intro price, the conclusion may flip when the standard price takes effect — date-stamp the analysis.
When "spend more" is still right
Diminishing returns is a per-task verdict, not a company-wide one. Anthropic positions Opus 4.8 as the starting point for complex enterprise and agentic work, and Fable 5 for workloads that need the highest available capability. If your eval shows a capability cliff — the cheaper tiers simply fail the task — then the marginal dollar is buying you a working product rather than a quality point, and the calculation is easy. The discipline is to demand that evidence per task instead of defaulting the whole organization to the top of the ladder.
Where to go next
Pair this with measuring actual savings from optimization work and model tiering for routing the easy traffic downward. The feature matrix shows which models are available on each platform.