Claude Haiku 5.5 Went From 0.0% to 39.2% on Terminal-Bench. That One Number Changes How You Build Agents.
Claude Opus 5.5 is still the better model. It just shouldn’t be doing all the work anymore.
Cover illustration generated with image2. This is not a product screenshot.
A team I’d describe as typical runs a document pipeline: pull filings, extract fields, reconcile conflicts, draft a summary a human will sign off on. Every step currently runs on the most capable model they can afford, because the one time they tried moving the boring steps to a cheaper tier, the cheap tier quietly mangled a table and nobody noticed for a week.
This story was written with the assistance of an AI writing program. That experience — try to save money, get burned, revert, never try again — is the reason most teams still overpay. And it’s the reason a single benchmark number from Anthropic’s October release deserves more attention than the price cut everyone wrote about.
The number
Claude Haiku 5.5 scores 39.2% on Terminal-Bench 4.0. The previous generation, Haiku 4.5, scored 0.0%.
Zero. Not “lower.” Not “acceptable for simple cases.” The prior small model could not complete that class of task at all.
This matters because of what it implies about a specific architecture. The idea of “expensive model plans, cheap model executes” has been circulating for two years. It kept failing for an unglamorous reason: the cheap model couldn’t execute. You’d decompose a task beautifully, hand the steps down, and the execution tier would fall over on step three. So teams did the rational thing and ran everything on the expensive tier.
A jump from 0.0% to 39.2% isn’t an incremental gain. It’s a capability crossing a threshold that a whole design pattern was waiting on.
What the tiering actually buys you
Cognition tested this directly. Running Devin Fusion with Haiku 5.5 as the execution sidekick and Opus 5.5 as the lead planner, they report a FrontierCode score of 66.2 — higher than either model scored on its own (Haiku 5.5 alone: 46.4%; Sonnet 5.5 alone: 52.1%) — while cutting both cost and latency.
Sit with that for a second. The combination beat both components. That isn’t a cost optimization that sacrifices a little quality at the margin. It’s a configuration where splitting the work made the work better.
The explanation is unremarkable once you say it out loud: planning and executing are different jobs. Planning means holding several candidate approaches in view and choosing between them under uncertainty. Executing means taking an already-specified step and getting it exactly right. A single model doing both is averaging across two skill profiles. Separate them and each tier operates inside its strength.
What changed in October is not that this reasoning became true. It’s that the execution tier finally got good enough to hold up its end.
Mechanically the split is almost free: same base URL, same request shape, only the model field differs between the planning call and the execution call. The thing worth checking before you refactor is coverage, because not every route carries both tiers — support is listed per channel on pages like Crazyrouter's Anthropic model page, and whichever route you use, confirm the parameters you depend on survive it.
The constraint nobody mentions in the launch coverage
There is a catch, and it is pointed directly at the architecture I just described.
Haiku 5.5 is the only Claude model on the current price list that charges by prompt length. Below 100,000 input tokens: $0.10 in, $0.50 out per million. Above it: $0.50 and $2.50 — five times the rate, applied to the entire request, including output and caching. The threshold is evaluated per request. One hundred thousand tokens stays in the cheap tier; one hundred thousand and one does not.
Now consider what a coding subagent actually does. It accumulates. Every turn carries the conversation so far, the files it has read, the results tools returned. A subagent ten turns into a task is carrying a lot. The exact workload this model is marketed for is the workload most likely to drift across the line — and to do it silently, because nothing in the response tells you which tier you just paid.
Anthropic’s own savings math leans on this: roughly 90% of Haiku 4.5 requests came in under 100,000 tokens, which is how “90% cheaper below, 50% cheaper above” averages out to the headline figure of about 75%. That average describes the old traffic pattern. It does not describe an agent loop you are about to build.
Get rhythm’s stories in your inbox
Join Medium for free to get updates from this writer.
Worth checking the tiering against whatever route you actually call through, since resellers restate it with their own snapshot dates — Crazyrouter’s Haiku 5.5 pricing page lists the same 100K rule and matches the official rate as of its 13 October snapshot. Reseller and official pricing are separate figures; don’t average them into one budget.
So the engineering task isn’t “switch the model.” It’s context discipline: trim what each subagent carries, estimate before dispatch, split tasks that approach the ceiling. Do that and the 40x gap holds. Skip it and you land in a tier that’s only 8x cheaper, having told your finance team 40.
Where Opus 5.5 still earns its price
The case for the expensive tier has shifted from “it’s smarter” to something more specific and more useful.
Anthropic’s own measurements emphasize step count rather than accuracy. A 200,000-line audit finished in under three hours against more than twenty hours for Opus 5, using 2.5x fewer tokens. A 680,000-line migration completed inside a day. On a research task, 16 of 18 reports cleared the quality bar — while Fable 5.1 and Opus 5 cleared none.
Customers report the same shape. Optiver matched Opus 5’s quality in roughly half the turns, time and output tokens, cutting that workload’s cost by 40 to 50 percent. GitHub solved more terminal tasks in VS Code than with Opus 5, in less than half the steps. Kiro used about 40% fewer calls and half the tokens.
That pattern — winning by needing fewer attempts rather than by being more correct per attempt — is also where the “40% cheaper to run” figure comes from. List price only dropped 20%. The rest is efficiency.
And efficiency is precisely what you want from a planner. A planning tier that reaches a good decomposition in four turns instead of nine doesn’t just cost less; it gives the execution tier fewer chances to be pointed in the wrong direction.
Read the benchmarks the way Anthropic asks you to
One more thing, because it cuts against the instinct these numbers provoke.
Anthropic states plainly in the Opus 5.5 material that at current capability levels, benchmark margins have become a less reliable guide to real-world differences, and that the real gap to Fable 5.1 is narrower than the scores suggest.
The conditions matter too. Unless noted, results use adaptive thinking at max effort; on Terminal-Bench, Opus 5.5 ran at xhigh and GPT-6 Astra at high — each at its own best setting. Production safeguards were active, which Anthropic says likely depressed measured performance. Terminal-Bench 4.0 carries a standard error of ±2.6 points; the leaderboard’s 51.8% for Opus 5 reproduced at 52.3%, which Anthropic calls noise. The OSWorld figures are an offline subset.
The practical rule: read magnitudes, not decimals. 39.2% against 0.0% is a magnitude. It justifies redesigning a pipeline. A two-point gap between two flagship models does not justify anything.
What this means for the team in the first paragraph
They should not move their pipeline to the cheap tier. They should split it.
Reconciling conflicting filings, deciding which source to trust, judging whether a summary is defensible — that stays on Opus 5.5, because the cost of getting it wrong dwarfs the token difference. Pulling fields, compressing documents, routing by type, running the already-specified steps — that moves down, because those outputs can be checked by a program rather than by a reviewer’s impression.
That last clause is the whole test. A step is safe to move down when a program can tell you it went wrong. Field present and well-formed. Figure traceable to the source document. Enumerated value inside the allowed set. If you cannot write that check, you have not earned the right to demote the step — because cheap models rarely fail by refusing. They fail by answering confidently, and confident wrong answers are exactly what a human spot-check misses.
Which suggests the order most teams get backwards. Run it on the expensive tier first, not to save money but to discover what “correct” looks like well enough to encode it. Then demote, one checkable step at a time.
Starting cheap and upgrading when it fails sounds thriftier. It isn’t, because on day one you have no acceptance criteria, so you have no way to detect failure. You’ll just ship the mangled table again and find out in a week.
The limit of all this
One honest caveat. Anthropic still positions Sonnet 5.5 and Opus 5.5 as the better choice for complex agentic coding, and describes Haiku 5.5 as suited to “more narrowly scoped tasks.” Terminal-Bench puts the gap at 39.2% against 70.6%. Long-horizon coding is not where the cheap tier belongs, and a tiered architecture that hands it a genuinely open-ended problem will fail — just more cheaply.
The claim here is narrower and, I think, more durable: an architecture that has been theoretically appealing and practically broken for two years now has a working execution tier. That’s worth more than a price cut, and it’s the part the launch coverage mostly skipped.
