Many LLM cost decisions rest on half remembered rules. This post checks six common beliefs against provider documentation fetched on 19 October 2026. Details change, so treat each fact as a pointer to the page you should confirm, not a permanent rule.
Myth 1: A large max_tokens raises my rate limit usage
Fact: the rate limits page says max_tokens does not factor into output token limit calculations, because only generated tokens count. It still matters for truncation, and thinking counts toward it.
Myth 2: Caching always saves money
Fact: a write costs more than normal input in the caching docs, with a five minute write listed at 1.25 times and a one hour write at 2 times the base price. A prefix used once, or a cache that expires between requests, costs more than no caching. Prompts under the model minimum are not cached at all, and no error is returned.
Myth 3: Counting tokens uses my quota
Fact: the token counting docs say it is free, subject to a requests per minute limit by tier, and that it has limits independent of message creation. It does not apply caching, and it is an estimate.
Myth 4: Batch results arrive in order, within the hour
Fact: the batch docs say most batches finish within an hour, but the guarantee is that a batch expires if it does not finish in 24 hours. Results can come back in any order, so match by custom_id. Cache hits inside batches are best effort.
Myth 5: Omitting a parameter keeps my behavior when I switch models
Fact: the effort docs say the default level differs by model, and that on one newer model omitting the setting runs one level lower than it did on the previous model. Set it explicitly and re-test.
Myth 6: A spend cap returns a retryable error
Fact: the spend cap response is an HTTP 429 with a dedicated error code and no retry-after header, and retries fail until access resumes. A limit you set yourself returns HTTP 400 instead. Handle each case differently.
What these have in common
Each myth comes from treating cost as one number. Real cost depends on input and output tokens, cache behavior, model defaults, and how requests are shaped. The fix is the same each time: log the usage fields, read the current documentation, and test on your own workload.
A habit worth adopting
Keep a short internal page listing the assumptions your cost model depends on, each with a link to the provider page and the date you last checked it. Review it when you change models or when billing surprises you.
Quick check
Which myth costs teams the most?
It depends on the workload, but assuming caching is working without checking the usage fields is common. Verify the cache read tokens in real responses.
Where should I confirm these facts?
On your provider's current pricing, caching, rate limit, and batch pages. Numbers and defaults change between model releases.
Sources
- Rate limits, Claude docs
- Prompt caching, Claude docs
- Token counting, Claude docs
- Batch processing, Claude docs
- Effort, Claude docs
Last reviewed: 19 October 2026.
