[OsamaAnsar]

Description

computeTokenUsage computes inputTokens.noCache as prompt_tokens - cached_tokens and ignores prompt_tokens_details.cache_write_tokens. When cache writes are reported, noCache still contains the written tokens, so noCache + cacheRead + cacheWrite is larger than total.

Before (the fixture already used in compute-token-usage.test.ts, plus the example from OpenRouter's usage-accounting docs):

prompt_tokens=100 cached=20 cache_write=30 -> {"total":100,"noCache":80,"cacheRead":20,"cacheWrite":30}  noCache+read+write = 130
prompt_tokens=194 cached=0  cache_write=100 -> {"total":194,"noCache":194,"cacheRead":0,"cacheWrite":100} noCache+read+write = 294

After:

prompt_tokens=100 cached=20 cache_write=30 -> {"total":100,"noCache":50,"cacheRead":20,"cacheWrite":30}  sum = 100
prompt_tokens=194 cached=0  cache_write=100 -> {"total":194,"noCache":94,"cacheRead":0,"cacheWrite":100} sum = 194

The existing tests asserted noCache: 80 for the first case (compute-token-usage.test.ts, usage-accounting.test.ts, stream-usage-accounting.test.ts); I changed those expectations to 50.

Fix

noCache = Math.max(0, prompt_tokens - cached_tokens - (cache_write_tokens ?? 0)) in src/utils/compute-token-usage.ts. doGenerate and doStream of both the chat and completion models go through this one function, so no other call site needed changing. When cache_write_tokens is absent or null the result is unchanged (prompt_tokens - cached_tokens).

Why this is the intended contract

LanguageModelV4Usage.inputTokens.noCache in vercel/ai is documented as "The number of non-cached input (prompt) tokens used." (packages/provider/src/language-model/v4/language-model-v4-usage.ts). For the same prompt_tokens_details.cache_write_tokens field, @ai-sdk/openai computes (packages/openai/src/chat/convert-openai-chat-usage.ts, line 37, at vercel/ai 3216f06):

noCache: promptTokens - cachedTokens - (cacheWriteTokens ?? 0),

The total - cacheRead formula here comes from #409, which followed @ai-sdk/openai-compatible; that converter has no cache-write field (cacheWrite: undefined), so the formula was right for it but does not carry over once cacheWrite is populated.

What I could not confirm

I assumed prompt_tokens already includes cache-write tokens. OpenRouter's docs describe cache_write_tokens as "prompt tokens written to the cache" and their usage example (prompt_tokens: 194, cache_write_tokens: 100) is consistent with that, but I did not find a sentence stating it outright, and I have not made a live API call. If some upstream provider reports prompt_tokens excluding writes, this would under-count noCache for it.

Clamping

The reference @ai-sdk/openai converter does not clamp noCache (it does clamp text at 0). I clamp at 0 so inconsistent upstream numbers (cache tokens larger than prompt_tokens) cannot produce a negative count. Happy to drop the clamp to mirror the reference exactly.

Checklist

• I have run pnpm stylecheck and pnpm typecheck
• I have run pnpm test and all tests pass (I ran the two underlying commands, vitest --config vitest.node.config.ts --run and vitest --config vitest.edge.config.ts --run, on Node 22: 35 files / 576 tests pass in each; not the pnpm test script itself)
• I have added tests for my changes (if applicable)
• I have updated documentation (if applicable) (not applicable, no docs mention this)

Changeset

• I have run pnpm changeset to create a changeset file (I added .changeset/exclude-cache-writes-from-no-cache.md by hand in the same format as the existing patch changesets)

Tests

New cases in src/utils/compute-token-usage.test.ts: cache write only, noCache + cacheRead + cacheWrite = total, absent/null cache_write_tokens (unchanged behaviour), and cache tokens exceeding prompt_tokens (clamped to 0). Reverting only src/utils/compute-token-usage.ts makes 6 tests fail (4 in compute-token-usage.test.ts, plus the two updated usage-accounting / stream-usage-accounting expectations); restoring it makes all pass.

Not tested: no live OpenRouter calls and no e2e tests (pnpm test:e2e needs an API key).