Opus vs Sonnet: the routing question your task already answered

Two Claude tiers can do the job in front of you, and one costs 1.67x the other. Claude Opus 4.8 meters at $5 in and $25 out per million tokens; Claude Sonnet 5 meters at $3 and $15,¹ the same 1.67x per token in and out, because every model in this family bills at one-to-five input-to-output. Paying up for Opus feels like buying insurance against a wrong answer. On a wide band of everyday work, our own routing eval says that premium buys nothing you can measure.

Today the premium is steeper than sticker. While Sonnet 5 holds its introductory $2/$10 rate through August 31, 2026,¹ Opus costs 2.5x Sonnet on the same call. Step down one more rung and Claude Haiku 4.5 runs at $1/$5, a third of Sonnet’s sticker. Four tiers, a ten-to-one spread from Haiku to Fable 5, and one live question underneath every routing decision: does the pricier model get your task right, run after run, in a way the cheaper one does not?

Pay Opus’s 1.67x premium only where a grader cannot check the answer. Deterministic, gradeable work (extraction, unit-testable code, format transforms) goes to the cheapest tier that clears your checks, Haiku 4.5 then Sonnet 5; Opus 4.8 and Fable 5 earn their multiple on long-horizon or open-ended runs no script can score. Route by how important the task feels instead and you pay the premium, or Fable’s refusal tax on top, for an answer the cheaper tier already returned.

Claude Opus pricing vs Sonnet: a fixed 1.67x premium (2.5x while the intro rate holds)

Claude Opus 4.8’s API pricing is $5 per million input tokens and $25 output, against Claude Sonnet 5’s $3 and $15 (introductory $2 and $10 through August 31, 2026).¹ That is a flat 1.67x premium at sticker, 2.5x while the intro rate holds, identical in and out. Batch processing halves both rates and cached-read input tokens reprice far lower, so the ratios below track base per-token prices.

Every model in the current family bills at a one-to-five input-to-output ratio, so a single multiplier describes the whole gap in both directions. Set Sonnet 5’s sticker at 1x and the Claude Opus vs Sonnet vs Haiku ladder reads cleanly.

Whether the premium earns out is the rest of this page.

Until August 31 every Opus call costs 2.5 Sonnet calls; after that the premium settles at its steady 1.67x. Both numbers only bite if the premium buys something, which is the next section’s question.

On checkable work, Sonnet kept pace with Opus and returned more correctness per dollar

If capability separated the two, price would be moot. On short work a script can grade, it does not. We wrote benign, deterministically scored tasks (a seating-arrangement logic puzzle and a wire-transfer tool-use approval among them) and pruned them to 14 items where Opus had stumbled in a calibration pass, a set that if anything tilts toward the cheaper challenger.² At low effort, on the non-refused trials, the two arms landed here:

Sonnet’s point estimate sits above Opus and the intervals overlap almost entirely, so this sample supports one read only, no measured separation: on tasks picked for Opus’s failures, the cheaper model was at least its equal. Why overlapping intervals cannot license a ranking is a question the significance-testing guide takes up.

One scope note the study carries: the Sonnet arm was a low-effort spot-check outside the pre-registered Opus-versus-Fable design, so treat it as the current best read and hold the ranking open.

Price then decides what capability would not. In the prior 40-task round we scored a unitless quality-per-dollar composite, pass rate over mean cost per task on sticker pricing, and Sonnet returned about 387 against Opus’s 187, roughly a twofold value lead; while Sonnet’s intro rate holds through August 31, 2026, its figure rises to about 580 and the lead widens toward threefold.² These are point composites, their uncertainty living in the component pass rates, so they rank the tiers by value while carrying no interval bounds of their own. The margin is not close.

Consistency did not favor the premium either. At low effort Sonnet’s reliability@k was 0.946 against Opus’s 0.875; at extended-high effort Opus itself reached 0.946.² The sample is too small to rank the two, and what it rules out is the premise that the premium buys run-to-run consistency. The tier-by-tier routing rule this feeds is routing Claude models by task.

On checkable short-horizon work, Sonnet 5 was statistically indistinguishable from Opus 4.8 and led on correctness per dollar. The premium bought no measurable capability here.

Claude Sonnet vs Haiku is the same rule, one rung cheaper

The logic does not stop at Sonnet. Claude Haiku 4.5 runs at $1/$5, a third of Sonnet’s sticker and half of it during the intro window, and Anthropic positions it as the fastest model with near-frontier intelligence, one rung quicker than Sonnet on the same latency ladder.¹ That speed is the reason latency-sensitive routing (tool-call dispatch, interactive extraction) defaults to the bottom of the ladder, so for the most deterministic slice of checkable work (short extraction, classification, format transforms, tool-call routing) the same question applies: does Sonnet get your task right where Haiku fails, often enough that its 3x earns out?

Our routing eval measured Sonnet, Opus, and Fable, not Haiku, so treat the Sonnet-vs-Haiku call as a hypothesis to settle on your own gradeable set; we cannot hand it to you as a result. The route-by-checkability rule points the same way it did one rung up: test the cheapest tier first, escalate only where the checks fail. Concretely, run your classification check on Haiku 4.5 first; if it clears on single-clause inputs but drops on multi-clause ones, that failed check is your signal to step up to Sonnet 5, and the escalation costs 0.33x to 1x. What you should not do is skip Haiku because the label reads “fastest” where Opus reads “most capable,” any more than you reach for Opus on the strength of that label. The label is not the measurement.

On the most deterministic tasks, price down to Haiku 4.5 and let your checks decide when a task earns Sonnet’s 3x, with the tier label kept out of the call.

Refusal behavior changes the price you actually pay

Sticker price and effective price diverge the moment a model declines. Claude Fable 5 ships a stricter safety classifier, and Anthropic is candid that it flags benign coding and debugging more often.³ In our set, Fable refused 75 to 86% of these benign, checkable tasks (21 of 28 at low effort, 24 of 28 at extended-high), and 44 of 45 refusals carried the API’s cyber category on tasks we judge benign, the seating puzzle and the wire-transfer approval among them.² Turn on server-side fallback and the refusals reroute to Opus 4.8: fallback fired on 20 of 28 calls and lifted the pass rate to 96.4%.² You pay Fable’s 3.33x and read Opus’s answer. The instrumentation for catching that silent swap is logging refusals and model fallbacks; the consumer-side version of the same math is do you need Fable 5.

For the Opus-versus-Sonnet decision this is the reassuring part. In the same sweep neither refused a single task (Opus 0 of 28 at both effort levels, Sonnet 0 of 28 at low effort),² so on this benign coding work neither adds a refusal tax, and the choice between them stays a clean trade of price against measured capability. Refusal economics is a Fable-tier problem, and it is the reason a cheaper model can be the more reliable choice per dollar once you count only the calls you actually get an answer to.

A rescued call bills at the pricier model for the cheaper model’s answer. Routing spends effective cost, and the sticker price hides it.

When to use Opus vs Sonnet: a table by task risk

When to use Opus vs Sonnet comes down to one predicate: can a script check the answer? If a grader can score the output, the cheaper tier is the value knee, so route down to Haiku 4.5 or Sonnet 5 and gate on the check. If no grader can score it (a long-horizon or open-ended run), the premium finally has room to pay, so reserve Opus 4.8, with Fable 5 above it. Which model fits each kind of work is which Claude model to use for coding, writing, and research; this table prices the tiers you route between.

Put the two halves together and the rule becomes a table. Read down the task column to the row that matches, and the tier and its relative cost follow. The multiples are versus Sonnet 5 at sticker.

The premium tiers earn their multiple on work a grader cannot check. A CSV-to-JSON transform with a unit test routes to Haiku 4.5; a forty-file state-management migration with no single gradeable key routes to Opus 4.8, because the check that would catch a cheap model’s mistake does not exist.

That boundary is also our eval’s blind spot: a deterministic exact-match harness cannot score a long-horizon run where a small early error compounds across dozens of steps, and that is exactly the workload Opus and Fable are built to win.² We did not measure it, so we do not claim it in either direction. Everything a test can grade, though, routes to the cheapest tier that clears it, which is where the Opus-versus-Sonnet answer already lives. How a model earns the “most capable” label in the first place is how benchmarks get gamed.

Our eval’s blind spot is the premium’s home turf: the long-horizon runs a grader cannot score are exactly what Opus 4.8 and Fable 5 are built for. That splits the workload cleanly. The checkable half is the settled half, and it routes to the cheapest tier that clears the check.

Let the check route the task, then let price break the tie

The rule has a break-even underneath it. Say Sonnet clears 95 of 100 checkable tasks and Opus clears all 100. Opus asks 1.67x per token for those five extra passes, and each of the five misses Sonnet made was caught by the same check that made the work checkable in the first place. A caught miss is cheap; a 1.67x premium on every token to avoid it usually is not.

• Default checkable work to the cheapest tier that clears your checks. Extraction, tool-calls, format transforms, unit-testable code: on our set we measured no separation across Sonnet and Opus, and the correctness-per-dollar knee sat at the bottom of the ladder.
• Reserve Opus 4.8 for long-horizon, compounding work, and Fable 5 for what needs the frontier. Those are the runs a grader cannot score and the only place the 1.67x and 3.33x have room to pay.
• Let price decide; do not route by the leaderboard. A top-of-the-benchmark label is a peak claim about someone else’s tasks; only your own set, run after run, tells you which tier clears yours.
• Watch the intro clock. Sonnet 5’s $2/$10 holds only through August 31, 2026; after it, Sonnet is 1x and the Opus premium reads its steady 1.67x.

The pricing is the easy input. The hard one is the reliability measurement sitting under it, and the research desk keeps that: the full method, per-tier intervals, and the results behind every number here live in the routing eval, and the metric for consistency past a single pass rate is how to measure agent reliability. Route by checkability, and “most capable” goes back to being a label, priced like one.