Fix a Haiku 5.5 400 error

A Haiku 5.5 400 error usually means the service rejected the submitted request. Begin with the response's named field and the request your application actually serialized. Repeatedly sending an unchanged invalid body is unlikely to help, and switching models before inspecting the payload can hide the compatibility problem instead of fixing it.

This checklist targets request validation failures. If the response reports a different status, follow that status rather than forcing it into this diagnosis. A browser message saying "request failed" is not enough evidence to classify the upstream problem; inspect your server's sanitized error record first.

Locate the Haiku 5.5 400 error boundary

Determine which service produced the response. Your application, a gateway and Anthropic can each reject a request. Record the HTTP status, provider error type, request identifier and timestamp where available. Keep the endpoint host and selected model identifier, but redact credentials and private source content.

Compare the body sent over the wire with the object you intended to send. An SDK wrapper may inject defaults, translate field names or retain settings from another model. The code at the call site can look correct while the final serialized request contains an unsupported argument.

Do not publish a full request dump in a support forum. Replace customer material with a short invented input that still reproduces the failure. Preserve structural details such as message roles and field types; those details often matter more than the original prose.

Reduce to a minimal request

Start from the native API quickstart or your gateway's own minimal example. Keep the provider, credential and model route unchanged while removing optional features. Use one short user message and a reasonable output allowance. This isolates request compatibility from document size and tool complexity.

If the minimal version succeeds, restore optional fields one at a time. Save each failing case as a regression fixture. If it still fails, inspect the exact error before changing the endpoint or account: an invalid model identifier and an invalid request parameter require different corrections.

For a Haiku 5.5 400 error after an upgrade, compare against the model-specific migration guide. A generic API example can describe fields accepted by other models without establishing that this model accepts the same combination.

Check migrated settings

The migration guide identifies old manual thinking-budget configuration and assistant prefill as compatibility hazards. Remove the old pattern and use the documented model-specific request shape. Turning thinking off is not a general way to make every legacy field valid again.

Also inspect sampling fields added by a shared client configuration. Avoid retaining a field merely because its value looks like a harmless default. Presence can matter independently of value. The actual serialized body is the evidence that determines whether a wrapper has removed an unsupported setting.

Check structured-output configuration separately. An old parameter name, unsupported schema construction or incompatible feature combination can fail before generation begins. Reduce the schema to a small documented example, then restore business fields gradually. Do not treat a schema rejection as proof that the model cannot perform the underlying extraction task.

Review message and tool structure

A Haiku 5.5 400 error involving conversation history can come from a malformed sequence rather than the newest user message. Inspect the final message role, tool-result identifiers and whether the adapter preserved required response blocks. Summarizing history into visible text is not always equivalent to replaying the protocol objects.

For tools, check names, input schemas and result associations. A result referring to a nonexistent tool call should be repaired in the application state, not papered over with a prompt asking the model to ignore it. Keep a synthetic fixture representing the exact invalid sequence so the repair remains testable.

Validate basic JSON types before submission. A string where an integer is expected, an empty required array or an unexpected null can be introduced during form parsing. Local schema validation provides faster feedback than sending every malformed request upstream and waiting for rejection.

Separate size problems from parameter problems

If the response names request size or context capacity, measure the complete request rather than only the latest input box. History, tool definitions and reference material all contribute. The token-count guide explains why character length is only a rough application constraint, not the model's tokenizer count.

Do not truncate arbitrary bytes from a JSON body to make it smaller. Reduce source material at a meaningful boundary, preserve complete message objects and revalidate the task. A smaller but corrupted request creates a second error while obscuring the original one.

An incomplete generated answer is a different case. If the provider accepted the request and returned an output-limit stop, use the output-limits guide. Increasing the input allowance will not resolve an answer that ran out of output budget.

Verify the fix without masking the defect

After correcting a Haiku 5.5 400 error, rerun the smallest reproduction and then the original synthetic workflow. Confirm both transport acceptance and the expected result shape. A successful status alone does not prove that removing a parameter preserved the behavior your application needed.

Add a test at the layer that introduced the invalid field. If a shared adapter injected it, test that adapter rather than only the page that happened to expose the failure. Keep automatic retries disabled for this invalid-request case so future regressions remain immediate and visible.

Escalate with a sanitized request, SDK version, provider route and request identifier when the documented minimal example still fails. That evidence is more useful than a screenshot of a generic error toast. It also lets support investigate without access to your secret or your customer's original document.