A cheaper small model changes the cost of a bounded task. It does not determine where that task should begin and end. Anthropic’s October 7 Haiku 5.5 release makes partitioning work across models a more consequential engineering decision: the application must decide what context to transfer, what answer to accept, and when delegation costs more than direct execution.
Our Opus and Sonnet 5.5 assessment examined model efficiency and fallback attribution. This follow-up concentrates on a different mechanism: whether a small model can take a well-defined piece of work out of a larger model’s critical path without adding error or coordination overhead.
What is new in the release
Anthropic positions Haiku 5.5 for frequent, narrowly scoped work and introduces adjustable effort to the Haiku class. Its launch reports roughly 75% lower average running cost than Haiku 4.5, rather than a universal saving on every request. The same announcement halves Sonnet 5.5 cache-read pricing to $0.10 per million tokens and describes computer/browser support in SDK beta. Haiku 5.5 announcement
The model documentation specifies a one-million-token context window, up to 128K output, and adaptive thinking with medium effort by default. Pricing changes at 100,000 input tokens: input/output rates are $0.10/$0.50 per million for the shorter band and $0.50/$2.50 above it. The published maximum context and the economical operating region are therefore different quantities. Model overview
These facts motivate a workload study. They do not establish that adding Haiku to an existing agent improves its delivered work. No benchmark in this article was independently reproduced.
The unit of delegation is an evidence-bearing answer
Consider an illustrative agent preparing a business presentation. It needs a revenue figure from a supplied report, an explanation of a change, and a recommendation. The figure lookup is a plausible bounded subtask; the recommendation depends on wider context and conflicting objectives.
A useful handoff for the lookup would specify the entity, reporting period, unit, requested line item, permitted source, and the evidence location to return. Its result should distinguish an exact value, an ambiguous match, and an absent value. A short answer without a source location forces the parent to redo the search.
This is my proposed design, not a special API contract supplied by Anthropic. It exposes the economic boundary: the child must remove work that the parent would otherwise perform. If the parent rereads the whole report and reconstructs the answer, the system has paid for two searches.
The acceptance rule can often be narrower than a second model’s broad review. Check that the cited page exists, that the unit and period match, and that the extracted value appears in the permitted source. Use deterministic checks where appropriate, and retain expert review for semantic ambiguity. A small model should not be asked to invent certainty when the document is unclear.
Context size can reverse the apparent saving
Transferring every conversation turn to every child is easy to implement, but it weakens the reason to delegate. A lookup may receive a long strategic discussion that neither helps locate the fact nor belongs in the child’s data scope.
The 100K breakpoint gives teams a concrete reason to inspect the distribution of assembled prompts. The migration guide also says the same text produces approximately 30% more tokens than on Haiku 4.5, depending on content. Reusing an old token estimate can therefore misclassify the price band or truncate a previously sufficient output budget. Migration guide
I would test compact task context against full-history inheritance while preserving the evidence needed for the task. Include an adversarial counterexample: a relevant constraint exists only in earlier history. Compression that saves tokens by deleting a decisive qualification is a failed partition.
Measure the parent and child together. Child input, reasoning, output, parent rereading, validation, and fallback all belong in the bill. A reduced child invoice can conceal a more expensive parent loop.
Adjustable effort changes the routing problem
The meaningful comparison is no longer simply small versus large. It includes small-model effort, context size, verification, and escalation. A high-effort child might solve a bounded ambiguity cheaply; it might also spend too long on a task that needed a deterministic lookup.
Build a development set with labeled uncertainty: straightforward extraction, conflicting values, missing evidence, and tasks requiring broader synthesis. Tune a policy on that set, then freeze it before evaluating held-out work. Let abstention be a useful result when the source cannot support a reliable answer.
Avoid using the child’s confident wording as the only escalation signal. Evaluate observable features such as conflicting source fields, missing required evidence, failed validation, or a task outside the declared scope. The policy should have an explicit cost for asking the parent to redo a task.
This proposed router is an application choice. It does not alter current permissions or make a model’s judgment sufficient authority for external action. Children inherit the applicable restrictions, and their retries consume the original task’s budget.
A controlled experiment for delegation value
Compare four systems on the same frozen inputs and acceptance rubric:
| System | Question |
|---|---|
| Parent executes all work | What is the direct-execution baseline? |
| Parent plus Haiku with full context | Does cheap inference help without context redesign? |
| Parent plus Haiku with bounded context and validation | Does a better partition earn its integration cost? |
| Same bounded design with another eligible model | Is the benefit specific to Haiku or to the handoff? |
Within the Haiku arms, vary effort under equal total task budgets. Include requests on both sides of the pricing threshold and representative image inputs. Randomize assignment, reset application state, and keep retries visible.
Measure complete artifact acceptance, evidence accuracy, abstention, parent correction time, end-to-end latency, and total spend. Report the fraction of child results used unchanged, revised, or discarded. That utilization measure reveals whether the parent actually trusts the partition enough to benefit from it.
For long task chains, test correlated errors. Two children reading the same ambiguous table may agree for the same wrong reason. Agreement is evidence about consistency, not independent confirmation. Blind reviewers should inspect accepted artifacts without seeing model labels or launch testimonials.
Migration has behavioral and platform details
The migration guide requires more than changing the model name: it covers adaptive thinking, selecting content blocks by type, removing sampling parameters, replacing assistant prefill, and handling safeguard refusals. It says Haiku 5.5 has no server-side fallback and documents platform-specific computer-use and structured-output differences. Migration requirements
The practical implication is to separate integration failures from capability failures. A parser that mistakes a thinking block for the answer is not evidence that the model cannot extract a field. A request rejected because its parameters are obsolete is not a reasoning benchmark.
Run compatibility fixtures before the quality study, then include genuine refusals and unresolved cases in operational reporting. Do not silently switch models after a refusal without recording the changed execution and ensuring that the new route remains permitted.
I could not retrieve the linked Haiku system card during this review because the source exceeded the retrieval tool’s size limit. Consequently this article makes no assessment of its detailed alignment measurements or benchmark configuration beyond the accessible launch documentation.
The business opportunity is more selective automation
The most promising application may be work that was too small to justify a large-model call: locating evidence, normalizing a record, or preparing a bounded candidate for review. A lower cost expands the set of economically testable tasks.
It can also encourage unnecessary subdivision. Every handoff adds context construction, waiting, validation, and failure recovery. The correct number of children is the number that improves accepted work under the actual budget, including human attention.
My adoption criterion would be a stable improvement in total cost or completion time with no unacceptable loss in evidence quality. Report the partition and its applicability conditions with the result. “Haiku worked well” is less reusable than knowing that it reliably handled a specific source format, ambiguity level, and validation rule.
Dated source ledger
Sources opened October 19, 2026:
- Haiku 5.5 announcement: October 7 release; model positioning and associated pricing/SDK changes.
- Model overview: living specifications and two pricing bands.
- Migration guide: living integration requirements. The system-card retrieval gap is stated above.