Claude Haiku 5.5 on Google Cloud: A Pilot Checklist

Anthropic released Claude Haiku 5.5 on 7 October 2026 for high-volume, latency-sensitive work. Google Cloud lists the model as generally available in Gemini Enterprise Agent Platform, with support for computer use, web search, batch predictions, prompt caching, function calling and the memory tool.

That combination makes Haiku 5.5 a credible candidate for repetitive agent steps such as classification, extraction, routing and narrowly scoped subagent work. A production decision still needs evidence from the organisation's own tasks. Start with one bounded workflow and measure whether the model meets its acceptance bar at the expected volume, context size and processing location.

What changed on 7 October

Haiku 5.5 supports a one-million-token context window, up to 128,000 output tokens and adaptive thinking with a default medium effort setting. The model ID is claude-haiku-5-5 on Google Cloud and the Claude API. Anthropic describes it as its fastest model at each model's standard speed, although Opus models in Fast Mode can run faster.

Use the launch benchmarks to form a hypothesis. Anthropic reports 72.4% on the offline subset of OSWorld 2.1 and 46.4% on FrontierCode 1.1 Main. Those controlled evaluations do not establish accuracy, latency or savings for a company's own workflow, so deployment approval should come from the pilot results.

Google Cloud documents model availability through US and Europe multi-regions plus a global endpoint. Its ML-processing table lists the US multi-region, Europe multi-region and asia-southeast1. Teams in the UK, Saudi Arabia, the UAE and the United States should map those documented options to their data-location requirements instead of assuming that a nearby country has a local processing region.

Choose a task that can fail clearly

The first pilot should have a high request count and a short definition of success. Good candidates include:

• assigning incoming requests to a known queue;
• extracting required fields from a standard document set;
• classifying records against a fixed taxonomy;
• retrieving one defined fact for a larger agent;
• producing a compact summary with mandatory source references.

Avoid starting with a task whose correctness depends on an open-ended policy judgment. The pilot needs examples where reviewers can agree on the expected output and identify harmful false positives or false negatives.

Build a holdout set from recent, approved examples. Include ordinary cases, boundary cases and inputs with missing or conflicting information. Keep the holdout separate from prompt development so the final score reflects unseen work.

Write the acceptance test before the prompt

Use metrics that reflect the task's failure cost. A routing workflow may need per-queue precision and recall because a strong overall accuracy can hide poor performance on a rare, sensitive queue. An extraction workflow needs field-level exactness, unsupported-value checks and a measure of human correction time.

| Pilot measure | What it answers | Example release rule | |---|---|---| | Accepted-task rate | How often is the complete result usable? | At least the existing production baseline | | Critical error rate | Does the model cross a business or security boundary? | Zero on the holdout set | | P95 latency | Will the workflow meet its response target under load? | Within the product's service objective | | Cost per accepted task | What does a usable result cost after retries and review? | Lower than the current route | | Escalation rate | How much work moves to a stronger model or person? | Within the staffed review capacity |

Release rules should use the organisation's existing baseline rather than a generic percentage. Record the model ID, effort setting, prompt version, endpoint and test-set version with every result so a later change can be compared properly.

Treat 100,000 input tokens as a price boundary

Google Cloud's published pricing for Haiku 5.5 uses a 100,000-token breakpoint. For the listed standard tier, input and output cost $0.10 and $0.50 per million tokens in the lower tier, then $0.50 and $2.50 at long-context rates. A footnote says queries at or above 100,000 tokens use the long-context rates for all input and output, so keep routine requests safely below the boundary and verify billing for edge cases. Non-global pricing shown on the same page is higher; check the selected endpoint's SKU when calculating the budget.

The threshold affects the request architecture as well as the budget. Capture the input-token distribution from representative runs and report the share that crosses it. A workflow that regularly sends an entire document collection may cost very differently from one that retrieves the relevant sections first.

Prompt caching can lower cost when requests reuse stable context. Google Cloud lists cache-hit pricing of $0.01 per million tokens for prompts within 100,000 tokens and $0.05 above it on the standard table. Measure actual cache-hit rates and include cache writes in the calculation. If each request changes most of the prefix, the workflow will rarely receive the advertised cache discount.

Pick the endpoint from the governance requirement

Google Cloud lists higher default throughput for the global endpoint: 3,000 queries per minute, 30 million input tokens per minute and 3 million output tokens per minute. The documented multi-region limits are 1,500 queries, 15 million input tokens and 1.5 million output tokens per minute.

The global endpoint may suit workloads that value throughput and flexible routing. A multi-region or documented processing location may be necessary where policy restricts where data can be processed. Make that decision with the security, privacy and platform teams before performance testing, because moving endpoints later can change price, quota and governance evidence.

Run load tests against the selected endpoint and request quota changes before launch if the expected peak approaches the documented limit. Use the payroll, billing or month-end burst in the test instead of the daily average.

Add an explicit escalation route

Haiku 5.5 should hand work to a stronger model or a person when the acceptance check fails. The escalation trigger belongs in code or policy, where the team can observe it. Useful triggers include a missing required field, conflicting sources, a restricted action, an output that fails schema validation or an input outside the tested task family.

Start with the lowest effort setting that passes the holdout set and load test. Compare it with medium, the documented default, before changing the model. A higher effort setting may recover enough difficult cases to reduce escalation, but it can also add latency and token use.

The same operating metric used in our model-routing guide applies here: count the model, tools, retries and review needed for an accepted result. Together, those costs describe the workflow more accurately than token price alone.

A go or no-go review for the pilot

Approve the route only when the evidence package answers these questions:

1. Does the holdout set represent current production inputs and important edge cases?
2. Do all critical criteria pass across repeated runs?
3. Does cost per accepted task stay within budget on both sides of the 100,000-token boundary?
4. Does the chosen endpoint satisfy processing-location and audit requirements?
5. Can the stronger-model or human review queue absorb the measured escalation rate?
6. Are prompt, model, effort, endpoint and dataset versions present in the trace?

If the pilot passes, increase traffic in stages and keep the previous route available for rollback. If it fails, narrow the task before adding more prompt instructions. Teams planning a broader Google Cloud agent programme can use the same acceptance evidence when defining scope for a Gemini Enterprise implementation.

Sources

• Anthropic, "Claude Haiku 5.5," 7 October 2026
• Anthropic, "Claude Haiku 5.5" model documentation
• Google Cloud, "Claude Haiku 5.5 on Google Cloud"
• Google Cloud, Agent Platform pricing