Claude Haiku 3 retires April 19: it's not just a model ID swap

1. temperature and top_p cannot both be set

Claude 3 accepted both sampling parameters. Haiku 4.5 requires picking one. If your code sets both, the request fails. The fix is one line: remove top_p.

// Fails with claude-haiku-4-5
client.messages.create({
  model: "claude-haiku-4-5-20251001",
  temperature: 0.7,
  top_p: 0.9,   // remove this
  ...
});

2. Two new stop reasons

Haiku 4.5 can return refusal and model_context_window_exceeded as stop reasons. If your code switches on stop_reason and has a default/else branch for unknown values, add explicit handling. A silently ignored refusal produces empty or partial output with no obvious error.

3. Trailing newlines in tool parameters are preserved

Claude 3 stripped trailing newlines from tool parameter values. Haiku 4.5 passes them through. Parsers expecting trimmed strings will now receive values like "value\n" instead of "value". Add explicit .trim() calls or update parsers before migrating.

4. Rate limits are separate

Haiku 4.5 has its own rate limit tier, not shared with Haiku 3. If you had elevated quota approved on Haiku 3, it does not carry over automatically. Check the Anthropic Console and request matching limits on Haiku 4.5 before flipping production traffic.

You need to stay on Anthropic (compliance, tooling, existing evals)

Go to Haiku 4.5 with batch processing at $0.50/$2.50. That is double the cost of Haiku 3, not quadruple. You also get the 64k output limit, better reasoning, and a knowledge cutoff 18 months more recent. Handle the four breaking changes before you flip the switch.

Cost is the priority and you have no provider lock-in

Test Gemini 2.5 Flash-Lite at $0.10/$0.40. It costs 60% less than Haiku 3 and has a 1M context window. For classification and structured extraction, quality is competitive. If it misses your bar, GPT-4o mini at $0.15/$0.60 is the next step up.

You need agentic coding or computer use

Haiku 4.5 at standard pricing is worth it here. 73.3% on SWE-bench Verified is competitive with models priced much higher. Haiku 3 could not do any of this - you are paying for an entirely new capability set, not a marginal quality bump.

When exactly does Claude Haiku 3 retire?

Anthropic's deprecation history table lists April 19, 2026. The models overview page shows April 20 in some places. Plan for April 19 - after that date, API requests return errors with no fallback.

Does the API redirect to a newer model or fail outright?

Requests fail. Anthropic's documentation says retired models return errors, not redirects. There is no automatic fallback to claude-haiku-4-5.

Is this just a model string change?

No. Haiku 4.5 is a Claude 4-generation model with API-level breaking changes: temperature and top_p cannot both be set, two new stop reasons were added, trailing newlines in tool parameters are now preserved, and rate limits are tracked separately.

What is the cheapest Anthropic replacement?

Claude Haiku 4.5 with batch processing at $0.50/$2.50 per million tokens - double Haiku 3's price, not four times. Batch delivery takes up to 24 hours.

What is the cheapest replacement across all providers?

Gemini 2.5 Flash-Lite at $0.10/$0.40 per million tokens - 60% less than Haiku 3 on input. GPT-4o mini at $0.15/$0.60 is also cheaper. Both work well for classification and extraction workloads.

What is the Vertex AI shutdown date?

August 23, 2026 - four months after the Anthropic API retirement. The Vertex model ID is claude-3-haiku@20240307.