[lucasbraum]

Two commits, in order: UPT-128 (the price table), then UPT-127 (auditable SDK cost).

UPT-128: price table

Source: https://platform.claude.com/docs/en/about-claude/pricing, read 2026-10-20 (cited at the top of the Anthropic table in pricing.ts).

Model	Input	Output	Cache read	5m write	1h write	Change
claude-sonnet-4-6	3	15	0.30	3.75	6	none
claude-sonnet-5-5	2	10	0.20 → 0.10	2.50	4	cache read is 0.05x input, per the page's footnote 2
claude-opus-5-5	4	20	0.20	5	8	none (also 0.05x)
claude-fable-5-1	10	50	0.25	12.50	20	none (0.025x)
claude-haiku-4-5-20251001	1	5	0.10	1.25	2	none

There is one change: Sonnet 5.5 cache reads. docs/model-pricing.md is updated to match.

• pricing.test.ts pins every Anthropic cell to a single literal expected table (toEqual). Setting Sonnet 5.5 back to 0.20 fails it. OpenAI and Gemini are not re-checked and not pinned (out of scope).
• sdkCostEstimate (UPT-94) does not move. It prices uncached token equivalents only (cachedInputTokens: 0), so the cache-read rate never enters it. cost-estimate.test.ts passes unmodified before and after (recorded runs: $0.23 for alanagoyal; $0.14 and $0.27 for the unit cases; all four within 1.3–2.5× of actual). SDK_ESTIMATE_MARGIN and cost-runs.json are unchanged. The actualUsd values there are the SDK's own totals with no token breakdown, so they cannot be repriced.

UPT-127: auditable SDK cost

Where tokens were merged: packages/core/src/fix/sdk-agent.ts:1058-1062 (on main) set inputTokens = input_tokens + cache_read_input_tokens + cache_creation_input_tokens and costUsd = result.total_cost_usd. The patch path did the same at packages/core/src/llm/fixer.ts:217 (billedInput), on top of adapters.ts:89, which already adds cache reads to input.

• sdkSpend(result, model) reads result.modelUsage, which the SDK documents as the field for token and cost accounting (it covers subagents and compaction, which usage does not). For each model it takes uncached input, output, cache reads and cache writes.
  • TTL split: modelUsage has none. result.usage.cache_creation has one, but for the main loop only. Its 1-hour writes are assigned to the session model; any other model's writes are priced as 5-minute writes.
• costUsd = tokens × pricing.ts rate, per class and model, rounded up to the nanodollar. sdkReportedCostUsd keeps the SDK's total_cost_usd separate, and each row also carries its own SDK figure.
• report.json and --json gain llm.models[]: inputTokens, cacheReadTokens, cacheWriteTokens, cacheWriteHourTokens, outputTokens, rates, costUsd, sdkReportedCostUsd. The same fields appear on each llm.tasks[] record.
  • costUsd and inputTokens keep their names. inputTokens now means uncached input only, documented in docs/model-pricing.md and the types.
• --max-cost counts max(computed, SDK-reported) (chargedUsd) for the feedback-round budget, the typical-session check and the Action's per-run budget.
  • Test: feedback.test.ts "counts the higher…" runs computed $0.05 / SDK $0.10 and the reverse. Both stop exactly where $0.10 sessions do. Making chargedUsd ignore the SDK figure fails the test.
• A gap above 5% (of the higher figure) adds a line to llm.warnings with both numbers, printed to stderr. It is never in the PR body: notes feed the PR, warnings do not.
• Unknown model: no guessed rate. The row has no rates and is marked unverified: true, costUsd is the SDK's figure, llm.costUnverified is set, and the run warns. If there is no modelUsage at all, the SDK total is used, also marked unverified. Tested in sdk-cost.test.ts.
• Patch executor: it did not have a breakdown before. It now returns the same single-model row per call, and assisted.ts and run.ts add rows up by model (addSpend).
• schema/uptide.run.v1.json is untouched: the run report copies named fields only, and its costUsd is now the computed figure.

Real run: hand calculation

uptide fix --only zod --max-cost 0.25 --json on a local clone of public consetto/calmcp at bdaef39^ (zod ^3.24.1 → 4.6.7). Origin was removed from the clone, and nothing was pushed or commented. Fix ran in its temporary clone and verification passed.

• UTC window: 2026-10-20T16:58:27Z → 2026-10-20T16:59:53Z
• Model: claude-sonnet-5-5, one SDK session, 31 turns, no feedback rounds. Preflight estimate $0.15 for 10 sites.

"inputTokens": 24,
"outputTokens": 6581,
"costUsd": 0.1622007,
"sdkReportedCostUsd": 0.1622007,
"models": [{
  "model": "claude-sonnet-5-5",
  "inputTokens": 24, "outputTokens": 6581,
  "cacheReadTokens": 256402, "cacheWriteTokens": 28281, "cacheWriteHourTokens": 0,
  "rates": { "input": 2, "output": 10, "cacheRead": 0.1, "cacheWrite": 2.5, "cacheWriteHour": 4 },
  "costUsd": 0.1622007, "sdkReportedCostUsd": 0.1622007
}]

Class	Tokens	$/MTok	Token × rate
uncached input	24	2.00	48.0
cache read	256,402	0.10	25,640.2
cache write 5m	28,281	2.50	70,702.5
cache write 1h	0	4.00	0
output	6,581	10.00	65,810.0
total			162,200.7 / 1e6 = $0.1622007

Computed $0.1622007 = SDK-reported $0.1622007, a 0% gap. At the old $0.20 cache-read rate the computed figure would have been $0.1878409, 15.8% above the SDK. So the SDK does use Sonnet 5.5's $0.10 rate, and the UPT-128 fix is right. These numbers are pinned in sdk-cost.test.ts, with the arithmetic in the comment.

Verification

• pnpm test: core 2460 passed / 37 skipped (205 files), cli 538 passed (48 files)
• pnpm typecheck: 3/3 tasks successful
• pnpm exec biome check .: no errors; the 2 warnings and 1 info are the same as on main

Follow-up commit (66f17f1)

CI's "No private material" step failed both test jobs. .changeset/auditable-sdk-cost.md matched PRIVATE_DOC (audit[^/]*\.md) in scripts/public-tree.mjs:41, so it is renamed sdk-cost-per-model.md. Locally that step exits early without UPTIDE_PRIVATE_DENYLIST, so it never ran. Run with a dummy denylist, it now reports clean. The same commit drops the unverified OpenAI/Gemini snapshot and states the 5-minute assumption for non-session-model cache writes in docs/model-pricing.md.

Open questions (answered)

1. The SDK enforces maxBudgetUsd inside a session with its own table. If that table ever undercharged, one session could pass the cap by Uptide's figures. max() only protects what comes after the session. Should the session's budget be scaled by the observed computed/SDK ratio?
2. 1-hour cache writes from a model other than the session model are priced at the 5-minute rate, because the SDK gives no per-model TTL split. Today's run had one model and no 1-hour writes.
3. The 5% gap is measured against the higher of the two figures. Say if you would rather measure it against the SDK figure.
4. The OpenAI and Gemini rows were not re-checked (out of scope; they are snapshot-pinned only).

🤖 Generated with Claude Code

[tenki-reviewer]

Review complete. No blocking issues — approved ✅; 3 nitpicks below.

🧹 Nitpicks (3) — 🟢 3 low

• 🟢 modelCost rates ignore the long-context price tier (pricing.ts:195) — modelCost builds rates from priceFor(provider, model) with no token count (packages/core/src/llm/pricing.ts:195, 199-205), while costUsd is computed via usageCost, which prices with priceFor(provider, model, billedInput) and switches to the long tier above the threshold (packages/core/src/llm/pricing.ts:127-128).
• 🟢 Failure paths drop models and costUnverified from the task record (feedback.ts:121) — feedbackPatchRunner forwards the provider's per-model cost breakdown and its unverified-cost flag only on the success path (packages/core/src/fix/feedback.ts:131-132); the answer.failure return at line 121 and the catch-block return at lines 147-153 forward costUsd/unreportedCostUsd but omit models and costUnverified, even when the fixer reported usage before failing.
• 🟢 addSpend duplicates identical cost warnings every merge (pricing.ts:226) — addSpend appends b.warnings wholesale without deduplication (packages/core/src/llm/pricing.ts:226), so the same per-session warning text is copied again on every merge.

---

Reworks LLM cost accounting in the fix pipeline around the SDK's own reported spend, adds cache-aware pricing and a per-model cost breakdown, and corrects the Sonnet cache-read rate.

Files	Change
packages/core/src/llm/pricing.ts, pricing.test.ts	New pricing engine: modelCost, addSpend, cache rates, long-context tier, warning collection.
packages/core/src/fix/sdk-agent.ts, sdk-cost.test.ts	SDK agent reads provider-reported spend and per-model usage, merging it into the spend accumulator.
packages/core/src/fix/feedback.ts, types.ts	FixReport.llm extends Spend; feedback forwarding carries per-model rows and costUnverified on success.
packages/core/src/fix/run.ts, assisted.ts, action/run.ts	Budget gating switches to chargedUsd (max of computed and SDK totals).
packages/core/src/llm/fixer.ts	Fixer emits token usage consistent with new uncached-input semantics.
packages/cli/src/cli.ts	CLI prints SDK-reported cost and accumulated warnings to stderr.
docs/model-pricing.md, docs/models.md, .changeset/*	Document the pricing table, report schema changes, and release notes.

Three low-severity issues remain: modelCost rows ignore the long-context tier, failure paths drop the per-model breakdown, and repeated warnings are not deduplicated.

_{Reviewed commit: b95d7fa}