feat: price AI Gateway 1h cache writes - #30872

Bills a Claude cache write with a 1-hour TTL at its model's 1-hour price, 2x the input price, instead of the 5-minute cache write price.

Stack context

AI Gateway prices every cache write at the 5-minute price and ignores the long-context prices that models.dev publishes, so Claude usage and long prompts are under-priced. On dogfood, Oct 1-14 recorded $7,535 against a list price of about $9,290. The 1-hour cache rule closes about 86% of that gap, and the tiers close almost all of the rest. The stack, bottom to top:

1. chore: add 1h cache write and long-context tier prices to the price book #30915: adds the 1-hour cache write and long-context tier prices to the price book.
2. feat: price AI Gateway long-context requests at their tier prices #30916: bills long prompts at their long-context tier prices.
3. feat: price AI Gateway 1h cache writes #30872 (this PR): bills 1-hour cache writes at the 1-hour price.

Why

Claude bills a 1-hour cache write at 2x the input price and a 5-minute write at 1.25x, but AI Gateway billed both at the 5-minute price. On dogfood the 1-hour rule closes about 86% of the gap above. This PR first carried both pricing rules; it now holds only the 1-hour rule, on top of #30916.

Providers

• Claude on Anthropic (including Claude Platform), Bedrock (InvokeModel and the Mantle /anthropic route), and Copilot goes through the Messages API, which reports the 1-hour part of each cache write. Those writes now use the 1-hour price.
• OpenAI models, on OpenAI, Azure, Copilot, or Bedrock Mantle (openai.*), have no TTL-dependent cache price. GPT-5.6 and later bill a cache write at 1.25x the input price with a single 30m TTL, and earlier models do not charge for cache writes. One cache write price already covers them, so they do not change.
• Claude through OpenRouter or Vercel goes through the OpenAI-compatible path, which does not report the 1-hour part, so those cache writes keep the cache write price.

What changed

• aibridge sends the 1-hour part of each cache write on the token usage record (aibridged API v1.5), and structured token_usage logs include it as cache_write_one_hour_input_tokens.
• Database (migration 000629): ai_model_prices.cache_write_1h_price, and on aibridge_token_usages the 1-hour token count and the 1-hour price that applied. A check keeps the 1-hour count within the cache write count.
• Cost: 1-hour writes use the 1-hour price and the other cache writes use the cache write price. A model or tier without a 1-hour price uses the cache write price that applies.
• API, SDK, and CLI carry cache_write_1h_price. It is optional, and omitting it clears it. The CLI adds --cache-write-1h-price.
• Model form: a read-only "Cache write (1h)" price when the row has one, and the tier lines include it.
• Docs: cost-controls.md states the rule and tells admins to set cache_write_1h_price on custom prices.

What to scrutinize

• A tier without a 1-hour price falls back to the tier's cache write price, not to the base 1-hour price (pricesForPrompt).
• usageFromRequest clamps a 1-hour count above the cache write count instead of rejecting it. The count comes from the provider, and a rejected record would drop the whole usage row.

Generated files

All from make gen: coderd/aibridged/proto/aibridged.pb.go (protoc); coderd/database/queries.sql.go, querier.go, models.go, dump.sql, and check_constraint.go (sqlc and the schema dump); coderd/apidoc/* and docs/reference/api/schemas.md (swag); site/src/api/typesGenerated.ts.

Validation

• The go test and site checks listed on feat: price AI Gateway long-context requests at their tier prices #30916 pass on this head, with the new 1-hour cases.
• Each new test was checked to fail when the behavior it covers is removed.
• Before the split, dogfood UAT of the combined change passed four rounds. The cost of all 29 new usage rows matched a recomputation (13 had 1-hour writes and 12 used a tier). A custom price row and usage rows recorded before the upgrade did not change. Through the CLI and the API, a missing or null cache_write_1h_price clears the price, and a 0 is kept.
• After the split, a smoke UAT on dogfood passed. It upgraded a test deployment from main to 6e7e4a67b68: migrations 000628 and 000629 applied, custom price rows and earlier usage rows did not change, and the cost of all 10 new usage rows matched a recomputation. On this head (ef70012cde3), which only adds log fields, live 1-hour writes were billed correctly, and the structured token_usage logs from both log sources carried prompt_tokens and cache_write_one_hour_input_tokens.