Which Claude should you pay for? The cheapest tier that clears your task's bar — not the best one. Anthropic now ships four: Haiku 5.5 ($0.10/$0.50 per million input/output tokens) for high-volume mechanical work, Sonnet 5.5 ($2/$10) as the workhorse default, Opus 5.5 ($4/$20) for hard reasoning and everyday coding agents, and the new Fable 5.2 ($8/$40) as a scalpel for frontier, long-horizon work. Fable is genuinely the most capable — but it costs double Opus for roughly a seven-point coding edge, so you promote specific hard jobs to it, never your whole bill.

With Fable 5.2's launch on November 9, the Claude lineup spans an 80x price range on the same call. That spread is the whole decision. Get the routing right and your model bill is a rounding error; get it wrong — by reflexively sending everything to the newest, smartest model — and you pay frontier prices for work a workhorse would have nailed.

The number that actually decides it#

Here's the trap. Fable 5.2 is the best model: on Anthropic's own FrontierSWE v2 agentic-coding eval it scores 69.9% against Opus 5.5's 62.3%, and Anthropic frames it as "a step above Opus 5.5 in almost all the capabilities we evaluated." So the lazy read is: use the best one. But look at what you're buying. That ~7.6-point edge costs you 2x the token price, and — this is the part the benchmark hides — the edge only exists on hard tasks. On classification, extraction, or routine chat, Opus, Sonnet, and Fable all one-shot the job. There, the gap is zero and you've paid double for identical output.

The token price is only half your bill. The other half is what a miss costs you — retries, and the engineer who cleans them up.

That second half is the real math. A weaker model that needs three attempts plus a human fix on a hard task is more expensive than a stronger one that lands it first try, even at 2x the sticker price. But flip the task to something easy and the logic inverts completely: both succeed once, so the cheaper tier simply wins. The rule that falls out of this is simple: use the cheapest tier that reliably clears your task's bar. Not the best model — the cheapest sufficient one.

Routing the four tiers#

Haiku 5.5 — the mechanical layer. At $0.10/$0.50, it's roughly one-eightieth of Fable's price. Send it the high-volume, low-judgment work: classifying support tickets, routing requests, extracting fields from documents, tagging, first-pass filtering. If a task is "apply an obvious rule a million times," it belongs here, and spending more buys you nothing.

Sonnet 5.5 — the default workhorse. At $2/$10, this is where most of a product's LLM calls should live: user-facing chat, summarization, moderate reasoning, the bulk of agent tool-calls. Start here for anything with real language in it, and only move up when you can see it stumble — looping, missing steps, producing output you have to redo.

Opus 5.5 — the reliable heavy. At $4/$20, Opus is the answer when Sonnet isn't dependable enough: harder reasoning, planning, and everyday coding agents that need to get it right without hand-holding. For most founders, Opus is the ceiling they actually need day to day — and it's half the price of the new flagship.

Fable 5.2 — the scalpel. At $8/$40, reserve it for the jobs where the capability margin converts into money: long-horizon autonomous agents, big multi-file refactors, frontier research or science loops, anything where a failed or stalled run burns an afternoon you can't get back. On those, one-shotting the task is worth the premium. On anything else, it's a luxury tax.

Three levers before you upgrade a tier#

Picking the right tier is the big win; these three compound it. Prompt caching is the largest: a cache read is $0.20 per million tokens on both Fable and Opus, versus $8 and $4 for fresh input — so any agent that re-sends the same system prompt or context should cache it. Batch processing takes 50% off non-urgent work. And fast mode (2x price for up to 2.5x speed on Opus and Fable) is a latency purchase, not a quality one — buy it only when speed is the product the user is paying for. Right-tier routing plus caching plus batching can turn the same workload into a fraction of a flat "always use the best model" bill.

The one-line version#

Don't ask which Claude is best — Fable 5.2 is, and that's not the useful question. Ask which is the cheapest tier that clears each task's bar, and route accordingly: Haiku for the mechanical, Sonnet for the everyday, Opus for the hard, Fable for the genuinely hard and expensive-to-get-wrong. The founders who overpay are the ones who picked a single default model; the ones who get frontier capability at a workhorse bill are the ones who route. If you're also weighing the subscription side of this, we cover that in which AI coding subscription a solo founder should buy.