Anthropic's Haiku 5.5 announcement reports most of its capability claims twice: once as a flat number in a comparison table, and once as a curve of accuracy against cost per attempt across five effort settings. The second view is the more informative artifact, because it changes what the first one can be said to prove. A model that ships with a test-time compute dial does not have a score; it has a frontier, and any single percentage quoted from it is a position on that frontier rather than a property of the weights.
What the dial actually controls. The setting exposes five levels — Low, Med, High, Xhigh and Max — matching the effort control Anthropic already offers on its larger models. Mechanically this is a per-request budget for intermediate computation: how much work the model may do before emitting the answer the caller sees. The published charts plot accuracy against cost per attempt in dollars on a log scale running from roughly $0.05 to $5, which is the right pair of axes for the decision a buyer makes, and also a reminder to read the axis. On a log scale a tenfold increase and a twofold increase occupy comparable horizontal distance, so two curves that converge near the cheap end can look like a tie while differing by a factor of five in spend.
One table, several settings. The static table compares Haiku 5.5 against Haiku 4.5, GPT-6 Luna and Sonnet 5.5 on seven evaluations. Six of the Haiku cells carry no effort annotation. One reference cell does: Sonnet 5.5's 52.1% on FrontierCode 1.1 is labelled Xhigh. That single label is enough to make the comparison indeterminate. Haiku's 46.4% on the same evaluation is 5.7 points lower, but a reader cannot tell whether the gap reflects the model or the effort the two were run at. The same ambiguity attaches to the largest claimed jumps. Haiku 4.5 has no effort setting at all, so the move from 15.7% to 72.4% on OSWorld 2.1 and from 10.2% to 45.9% on Humanity's Last Exam combines two changes at once: new weights and a new test-time budget. Part of that improvement is bought with tokens. That is not a defect of the model; it means the comparison sets a fixed configuration against a configurable one, and the two are not the same kind of object.
The computer-use numbers are qualified twice. Every OSWorld 2.1 cell in the table is marked "offline subset", and the y-axis of the accompanying chart reads partial-credit score. Both qualifiers narrow the claim. The offline subset removes tasks that need live network access, external accounts or services that may have drifted, which is exactly the class of variability that dominates browser-agent reliability in production; a score measured without it describes decoupled tasks. Partial credit rewards progress toward completion, so 72.4% is not a completion rate: a task that gets eleven of twelve steps right and then fails still scores well, while in a real workflow it returns nothing. For a browser-use deployment the operationally relevant figure is end-to-end completion on environments you do not control, and that figure is not in the table.
Where the dial does not help. Anthropic states the ceiling itself. On Terminal-Bench 4.0, Haiku 5.5 reaches 39.2% against 70.6% for Sonnet 5.5, and the post says plainly that the larger models remain better choices for complex agentic coding while Haiku is suited to compaction, summarisation and subagent work. The technical reason is compounding. A long task is a chain, and per-step reliability multiplies: at 0.98 per step a hundred-step task succeeds roughly 13% of the time. An effort setting raises per-step quality until the weights run out of headroom, but it does not shorten the chain or add the error recovery that long-horizon work needs. That is why the dial buys more on a twelve-step computer-use task than on a professional coding task that runs for an hour, and why "the small model is now nearly as good" only holds at a stated horizon length.
Per token, and per task. The pricing lines are what engineers budget against. Per million tokens at prompts up to 100k: input $0.10, output $0.50, cache read $0.01, cache write $0.125 — and every line is five times higher above 100k. Footnote 2 explains the headline "75% less to run on average": that is per task, not per token, and it is computed after accounting for the tokenizer change, which makes Haiku 5.5 spend slightly more tokens for the same piece of work. Two mechanics follow. Output costs five times input, and a higher effort setting spends its budget on precisely the expensive side, so the saving per task depends on where the dial sits. Second, the 100k boundary is a cliff rather than a gradient: agent loops grow context monotonically, so compaction cadence decides whether a request is billed at $0.10 or $0.50 per million input tokens. With cache reads at a tenth of fresh input, a byte-stable prefix is the largest single lever in a loop that resends the same multi-thousand-token instruction block on every turn — worth more than a step down in model size for that traffic.
Tools are a hidden variable. Humanity's Last Exam appears twice, at 45.9% without tools and 57.4% with them. That 11.5-point spread is wider than several of the cross-model gaps in the same table, and it belongs to the harness rather than the weights. Any result "with tools" is a joint measurement of model, retrieval corpus, tool definitions and retry policy; reproducing it means fixing all four. The same caution applies to the quoted customer evaluations. HubSpot's 92.8% averaged over three runs on a simulated CRM suite, and AlphaSense's 0.84 against 0.76 across 400 queries, are useful directional signals from teams that know their own distribution, but the harnesses are not published and cannot be re-run.
What to measure before moving volume. The numbers that decide a migration are mostly absent from launch pages. Pin the effort setting per call and log it, because a mixed-effort workload has no single cost figure. Measure cost per successful task rather than per attempt, retries included, since a cheap model that fails and re-runs is not cheap. Record cache read and write tokens separately. Check task-level refusal rates on your own inputs: safeguard tiers differ by model class — Haiku 5.5's cyber safeguards are described as more restrictive than Haiku 4.5's and less restrictive than Sonnet 5.5's — so behaviour on a classification workload can shift even while benchmark accuracy rises. And evaluate at the horizon you actually run, not at the benchmark's.
The useful part of this launch is the curve. Publishing accuracy against cost per attempt lets a reader place the model at their own budget instead of accepting one operating point, and it makes the trade explicit. What is still missing is matching discipline in the table above it: an effort label, a subset label, a tool-set label and a tokenizer version in every cell. A benchmark report carrying those is reproducible. One carrying a bare percentage is a claim about a configuration nobody wrote down.
Sources
Anthropic, "Introducing Claude Haiku 5.5", 27 October 2026 — https://www.anthropic.com/claude-haiku-5-5
Claude Haiku 5.5 System Card (evaluation methodology), linked from the launch page — https://www.anthropic.com/claude-haiku-5-5
OSWorld benchmark (Xie et al.) — https://os-world.github.io/
Terminal-Bench — https://www.tbench.ai/
Humanity's Last Exam — https://lastexam.ai/