Claude Haiku 4.5 as Default: A Token Receipt

01 / Make Haiku the first attempt, not the final authority

A default model should handle ordinary work without turning every request into an expensive investigation. My recommendation: start Claude Haiku 4.5 on bounded tasks with clear acceptance checks, then escalate failures. Think ticket classification, extracting fields from supplied text, drafting short replies, and writing small patches that tests can verify. Treat these as candidates to evaluate, not guaranteed capabilities on your data.

The useful question is not whether the model can answer something impressive. It is whether its accepted answers cost less after retries, review, and escalation. Here is a receipt you can reproduce instead of a benchmark you have to trust.

02 / Receipt one: 10,000 short support drafts for $35

Anthropic lists Claude Haiku 4.5 at $1 per million ordinary input tokens and $5 per million output tokens. Those are provider list prices, not a verified CheapRelay quote. Use them as a transparent reference; substitute the relay's displayed rates before budgeting actual usage. ([platform.claude.com](https://platform.claude.com/docs/en/about-claude/pricing))

Assume 10,000 requests, each using 2,000 input tokens and 300 output tokens. This is an illustrative workload, not a measured production average.

• ■ Input: 10,000 × 2,000 = 20 million tokens; charge: $20.
• ■ Output: 10,000 × 300 = 3 million tokens; charge: $15.
• ■ Total reference token bill: $35.
• ■ Average: $0.0035 per request, or 0.35 cents.

At these rates, another 100 output tokens on every request adds $5 across the run. Asking for a short answer is therefore a budget decision, not just a style preference. Cap verbosity without cutting fields your downstream application needs. ([anthropic.com](https://www.anthropic.com/claude/haiku?m=1&utm_source=openai))

03 / Receipt two: repeated context changes the bill

Anthropic lists five-minute cache writes at $1.25 per million tokens and cache reads at $0.10 per million for this model. Cache writes replace the ordinary input charge for those tokens; do not count both. ([platform.claude.com](https://platform.claude.com/docs/en/about-claude/pricing))

Consider 100 requests sharing an identical 8,000-token prefix, plus 500 fresh input tokens and 300 output tokens per request. Assume the prefix is eligible, caching is enabled, and all 99 subsequent calls hit the live cache. Anthropic documents a default five-minute lifetime that refreshes when cached content is used. A repeated prompt alone does not prove a billed cache hit. ([platform.claude.com](https://platform.claude.com/docs/en/build-with-claude/prompt-caching))

• ■ Initial prefix write: 8,000 tokens × $1.25 per million = $0.0100.
• ■ Prefix reads: 99 × 8,000 tokens × $0.10 per million = $0.0792.
• ■ Fresh input: 50,000 tokens = $0.0500.
• ■ Output: 30,000 tokens = $0.1500.
• ■ Cached total: $0.2892; uncached equivalent: $1.0000.

That difference belongs to these assumptions, not every workload. Check usage records for actual cache writes and reads before claiming the saving.

04 / Define enough with checks you can run

• ■ Classification: require an allowed label and audit ambiguous cases.
• ■ Extraction: validate the schema and verify extracted values against the source.
• ■ Support drafts: supply the policy and require approval before sensitive replies.
• ■ Small code changes: run tests and inspect the diff before merging.

Do not use a cheap default as permission to automate irreversible decisions. Route missing evidence, repeated validation failures, security-sensitive changes, and sprawling investigations to a stronger model or human reviewer.

05 / Add retries before declaring the default cheaper

Return to the $35 receipt. If 1,000 requests each need one identical retry, add $3.50: the total becomes $38.50. That hypothetical 10% retry rate is not a measured failure rate. Record accepted tasks, total tokens, escalations, and review time; compare cost per accepted result rather than cost per call.

CheapRelay offers pay-as-you-go access with no monthly fee or subscription, billing real tokens including cache reads and writes. Its shared key can simplify testing a default-and-escalation policy, but confirm route-specific pricing and caching behavior. Sign up for CheapRelay, run a representative sample, and keep Haiku as the default only where the acceptance checks and final receipt justify it.