Price Claude at Anthropic's list rates on the day they applied: Opus 5.5, Fable 5.1 and Sonnet 5.5 rows, Sonnet 5 at $2 / $10, and 1-hour cache writes

[JasonYeYuhe]

What people saw

The two newest Claude models were charged their predecessors' prices. The Mac's Claude table ended at Opus 5, Sonnet 5 and Fable 5, so Claude Opus 5.5 and Claude Fable 5.1 went through the family fallback and borrowed the newest row in their family. Cache hits are most of a Claude Code day, and that is where the borrowed rates were furthest off:

model	Anthropic list (input / cache hit / output, per 1M)	charged until now
claude-opus-5-5	$4 / $0.20 / $20	$5 / $0.50 / $25 (Opus 5's)
claude-fable-5-1	$10 / $0.25 / $50	$10 / $1 / $50 (Fable 5's)
claude-sonnet-5-5	$2 / $0.20 / $10	$3 / $0.30 / $15 (Sonnet 5's row)
claude-sonnet-5	$2 / $0.20 / $10	$3 / $0.30 / $15

Since #623 those borrowed figures carry "≈", so the first "≈" a Mac showed came from Claude.

Sonnet 5 was charged a rate it never had. It launched at $2 / $10, "introductory" through 2026-08-31, with an increase to $3 / $15 scheduled for 2026-09-01. On 2026-08-10 Anthropic made $2 / $10 the standard price and cancelled the increase. The table held the cancelled $3 / $15 for every day.

Every cache write was charged at the 5-minute rate. Anthropic charges a 5-minute cache write 1.25x input and a 1-hour write 2x. Claude Code splits its writes between the two in every response (usage.cache_creation.ephemeral_5m_input_tokens / ephemeral_1h_input_tokens), and on this Mac, over the last 45 days, the 1-hour share of the writes was 16% on Opus 5.5, 44% on Fable 5.1, 56% on Opus 5 and 94% on Opus 4.8. The scanner ignored the split.

The 200K long-context tier was computed per token kind. On a tiered row, the scanner charged the first 200K input tokens, the first 200K cache reads, the first 200K writes and even the first 200K output tokens at standard rates, and only the excess of each at the premium. Anthropic's rule is per request: a request whose prompt, cache reads and writes included, is over 200K pays the premium on every token, output included. It was also applied to Sonnet 4.6 forever, although its 1M window has been at standard pricing since 2026-03-13, and never to Opus 4.6, which paid the premium before that date.

What changed

The table (ClaudePricingTable.swift, new, CLIPulseCore, shared with iOS like CodexPricingTable). Every row was read against Anthropic's pricing page, https://platform.claude.com/docs/en/about-claude/pricing, on 2026-10-03, and is written in the page's column order: base input, 5-minute write, 1-hour write, cache hit, output.

• New rows: claude-opus-5-5 (released 2026-09-22), claude-fable-5-1 (2026-09-01), claude-sonnet-5-5 (2026-09-28), and claude-mythos-5-1 / claude-mythos-5 (Project Glasswing, Fable's prices). Dates are from the API release notes, https://platform.claude.com/docs/en/release-notes/api.
• claude-sonnet-5 at $2 / $10, cache hits $0.20.
• Cache hits are 0.1x input, except 0.05x on Opus 5.5 and 0.025x on Fable 5.1 and Mythos 5.1, as the page lists.
• Every other row is unchanged in its standard prices.

Dated rates. ClaudePricingTable.superseded holds the 200K tier for Opus 4.6 (1M beta from 2026-02-05) and Sonnet 4.6 (from its release on 2026-02-17) until 2026-03-13, when "the 1M token context window is out of beta for Claude Opus 4.6 and Sonnet 4.6, at standard pricing" (release notes). The instant is 00:00 UTC, the convention of CodexPricingTable. The premium rates (Opus 4.6 $10 / $37.50, Sonnet 4.6 / 4.5 / 4 $6 / $22.50, with the caching multipliers on top) come from the page as it stood then: https://web.archive.org/web/20260301150515/https://platform.claude.com/docs/en/about-claude/pricing. The page lists no tier for any current model ("Claude 4.6 and later models ... include the full 1M token context window at standard pricing"). Sonnet 4 and 4.5 keep theirs: their 1M beta was retired on 2026-04-30, after which a request over 200K is an error.

There is no dated rate for Sonnet 5: $2 / $10 is the only rate it ever had.

Each response is priced at its own time, with its 1-hour writes. parseClaudeFile reads the line's instant (the same parse that gave its day) and cache_creation.ephemeral_1h_input_tokens, and Pricing.claudeCostUSD takes both (cacheCreation1hInputTokens:, at:; both default to the old behaviour). The 1-hour share is clamped to the writes it is part of. A log without the split is charged at the 5-minute rate, as before.

The tier is per request. ClaudePricingTable.requestCostUSD puts the whole response on the long-context rates when input + cache reads + cache writes is over 200K. The old per-kind split was written for pricing a day's sum; the scanner prices each response as it reads it. The two places that still price a sum now use aggregateCostUSD, which never takes the tier, so a busy day cannot land on long-context rates: the fallback for a cached day row with no stored cost (entriesFromClaudeCache), and a live session's total (sumClaudeTotals, found by review).

Cache. costUsageCachePricingVersion 6 → 7: every Mac re-reads its Claude logs once, as #624's bump already does, so stored costs move to the new table. Pricing.claudeRatesFingerprint() (every row, every dated rate, and how a fixed list of names resolves) is pinned next to the version in ClaudePricingTableTests, as codexRatesFingerprint() is for Codex, so a future Claude price change without a bump fails the build. Codex is untouched (its fingerprint pin passes unchanged).

Effect on this Mac. A read-only script over this Mac's Claude Code logs for the last 30 days (aggregates only, last line per response, the app's rules approximated) puts the Claude estimate 18% lower under the new table than under the old: Opus 5.5 −41%, Fable 5.1 −25%, Sonnet 5 −30%, while Opus 5 and Opus 4.8 rise about 5% from their 1-hour writes. For the What's New decision on disclosing that Claude estimates rise (#624, mostly in output tokens): with this PR the Claude direction is mixed and depends on the models used, and on this Mac the net is down. So What's New should say that Claude estimates change (some models lower, some higher), not that they rise.

Not in this PR.

• Fast mode (Opus 5.5 $8 / $40, Opus 5 and 4.8 $10 / $50). No local log shows usage.speed other than standard, and Price Codex at OpenAI's list rates on the day they applied, and mark a borrowed rate with "≈" #623 left Codex's Fast multipliers out too.
• Data residency (inference_geo, 1.1x) and the Batch API discount: neither applies to Claude Code's interactive requests as logged.
• The desktop app (cli-pulse-desktop) has its own Claude table and is not changed here.
• Days the Mac archived before this version keep the costs they were recorded with. The routine read re-prices the days it reads, and those reach the cloud and the iPhone as the Mac uploads them.
• Vertex-style names (claude-opus-4-5@20251101) still do not normalize; that is unchanged.

How it is tested

• ClaudePricingTableTests (new, 20 tests)
  • Every row of current, all five columns, against the page; a row added to the table without being listed in the test fails.
  • The caching multipliers on every row and tier: 5-minute 1.25x, 1-hour 2x, cache hit 0.1x except Opus 5.5 (0.05x) and Fable / Mythos 5.1 (0.025x).
  • Only Sonnet 4 and 4.5 keep a tier today, at 2x input and 1.5x output.
  • The dated boundary: the 2026-03-13 00:00 UTC instant, and Opus 4.6 and Sonnet 4.6 one second before it and at it. No date means today's rate; every model with earlier rates has a row today, oldest first.
  • 1-hour writes: $5 / $8 / $6.50 for 1M 5-minute, 1-hour and half-and-half writes on Opus 5.5, clamped to the writes.
  • A sum never takes the tier.
  • Opus 5.5, Fable 5.1, Mythos 5.1 and Sonnet 5.5 resolve to their own rows (not approximate), including anthropic., Bedrock us.anthropic.…-v1:0 and dated spellings.
  • End to end through the scanner: an Opus 5.5 response with split writes costs $1.084 (it was $1.38) and is not approximate; a log without the split is charged 5-minute writes; two 301K-prompt Sonnet 4.6 requests either side of 2026-03-13 cost $0.186 and $0.093; a Fable 5.1 cache row stores $0.25 for 1M cache hits.
  • A response as Claude Code writes it, over several lines with the same ids where the last one counts (added after the 10-07 review): an Opus 5.5 response over three lines, only the last carrying the 1-hour split, costs $1.084, not the $0.904 of its first line's pricing; a two-line Sonnet 4.6 response with a 301K prompt before 2026-03-13 costs $0.231, not today's $0.123; and the same Opus 5.5 response with its first line in one read ($0.704) and the rest in an incremental one costs $1.084. Pending.line no longer has a default, so each place that makes a Pending names the line that prices it, or .alreadyPriced.
  • A cached day row with no stored cost, and a live session whose rows hold none, are priced as sums: 300K Sonnet 4.5 tokens cost $0.90, not $1.80.
  • The fingerprint pin, and a test that a repriced row, a repriced cache hit, a new tier, a new row and a moved dated end each change the fingerprint.
• ClaudeTieredAndCodexEdgeCostTests: the Claude half rewritten to the per-request rule (300K input pays $1.80, not $1.20; cache reads count toward the threshold; output does not decide it; exactly 200K is standard and one token more moves every token), with the archived page cited.
• Updated: Claude5PricingTests (Sonnet 5 at $2 / $10; an unknown claude-opus-6 now borrows Opus 5.5 and claude-sonnet-9 Sonnet 5.5; claude-opus-5-1 still takes Opus 5 over Opus 4.8), CodexPricingTableTests (the Claude fallback target and its $4), CostUsageCacheIOTests (version 7).

run	where	result
filtered swift test (the suites above plus ClaudePricingOpus47Tests, CostUsageScannerAccountingTests, UnpricedCoverageTests, ApproximatePriceCoverageTests, ArchiveTokenBasisTests, CodexTokenAccountingTests)	this Mac, acc987fe	218 tests, 0 failures
CLIPulseCore full suite, plain	this Mac, acc987fe	4,037 tests, 0 failures (5 skipped)
ClaudePricingTableTests after the review fix	this Mac, 831f01a8	18 tests, 0 failures
filtered swift test after the 10-07 review fix (ClaudePricingTableTests, CostUsageScannerAccountingTests, Claude5PricingTests, ClaudeTieredAndCodexEdgeCostTests, CostUsageCacheIOTests, ClaudePricingOpus47Tests)	this Mac, 7ff00eab	104 tests, 0 failures; SwiftLint on both changed files, 0 violations
CLIPulseCore full suite, plain	this Mac, 7ff00eab	4,041 tests, 0 failures (5 skipped)
gates: no internal docs, duplicate sources, copyright, hardcoded UI strings, personal-info mask, Apple strings parity (1,785 keys x 6, unchanged); SwiftLint on the new files	this Mac	all pass, 0 violations
CLIPulseCore unit tests	CI, 8a5cff14 (Swift CI run 37120631692)	plain 4,039 tests, 0 failures (5 skipped), and 0 failures in the Japanese, ROC and Buddhist calendar passes; -DDEVID_BUILD 4,077, 0 failures (7 skipped). All 18 ClaudePricingTableTests ran and passed in all five passes
CLIPulseCore unit tests	CI, this head 7ff00eab (Swift CI run 37500794263)	plain 4,041 tests, 0 failures (5 skipped), and 0 failures in the Japanese, ROC and Buddhist calendar passes; -DDEVID_BUILD 4,079, 0 failures (7 skipped). Both new tests ran and passed in all five passes. Every build and package job of the full matrix passed

Negative controls (throwaway nc/ branches on CI, deleted after; each mutation has a test that only it trips):

branch	mutations	failed tests (unit-test job)
nc/p156-claude-prices-a (run 37119510475)	Sonnet 5 back to $3 / $15; Opus 5.5 cache hit at 0.1x; the 2026-03-13 date a day late; the tier decided by uncached input only	15 of 4,038: Claude5PricingTests.testPublishedRatesPerMillionTokens and test_publishedRatesPerMillionTokens (Sonnet 5); test_cacheRatesFollowThePagesMultipliers and test_aScannedOpus55ResponseIsChargedItsOwnRates ($1.284) (Opus 5.5); test_theLongContextDateIsTheOneAnthropicPublished (date); test_sonnet_cache_reads_count_toward_the_threshold, test_sonnet_a_prompt_of_exactly_200K_pays_standard_rates and test_aScannedSonnet46RequestPaysTheTierOnlyBeforeMarch13 ($0.093) (tier); the fingerprint pin
nc/p156-claude-prices-b (run 37119524297)	the scanner ignores 1-hour writes; the scanner prices without the event time; the fingerprint leaves out dated rates; a cached day's sum priced as one request; the rules version left at 6	6 of 4,038: test_aScannedOpus55ResponseIsChargedItsOwnRates ($0.904); test_aScannedSonnet46RequestPaysTheTierOnlyBeforeMarch13 ($0.093); test_aTableChangeChangesTheClaudeFingerprint; test_aCachedDayRowWithoutAStoredCostIsPricedAsASum ($1.80); testCurrentPricingVersionIsSeven; the pin
nc/p156-claude-prices-c (run 37120002017)	a live session's day rows priced as one request; Fable 5.1 cache hit back to $1	5 of 4,039: test_aSessionWhoseRowsHoldNoStoredCostIsPricedAsASum ($1.80); test_cacheRatesFollowThePagesMultipliers, test_claudeCacheRowsCarryTheirResponsesCost ($1.00), test_publishedRatesPerMillionTokens and the pin (Fable 5.1)
nc/p647-fix-a (run 37500820418, after the 10-07 review)	the line that replaces an earlier line of a response priced as .alreadyPriced (without its time or 1-hour split), the reviewer's mutation	3 of 4,041: test_aResponseWrittenOverSeveralLinesIsPricedFromItsLastLine ($0.904 for Opus 5.5, $0.123 for Sonnet 4.6) and test_aResponseWhoseLastLinesArriveInALaterReadIsRepricedFromThem ($0.904)
nc/p647-fix-b (run 37500827434)	a line that replaces an open row carried from an earlier read keeps that row's priced flag	1 of 4,041: only test_aResponseWhoseLastLinesArriveInALaterReadIsRepricedFromThem ($0.904)
this Mac only	the replacing line keeps the earlier line's pricing (line: existing.line)	both new tests ($0.904 each); 79 other tests pass

Review. Gemini through agy-review on the diff, with the pricing page, the release notes and the archived page: one finding, fixed in 831f01a8 with a test (NC-C above): sumClaudeTotals, which gives a live session its cost, still priced a cached day row with no stored cost as one request, so a busy session could have landed on the tier. Codex CLI was out of quota.

Second review, 2026-10-07, on this head (8a5cff14): Gemini through agy-review again, on the same four questions (every rate and date against the page, the archived page and the release notes; event time and the 1-hour split on every path of parseClaudeFile, including rows carried from an earlier incremental read; any other place that prices Claude, or text that is now false; whether the tests catch a wrong rate, a lost split, a lost time and a per-kind tier): APPROVE, no findings. Codex CLI was tried again and was out of quota again. A repo-wide search found no other Claude price table (the Companion, the backend and Android carry none) and no user-facing string that names the old rates or the 1.25x write.

Third review, 2026-10-07, on 8a5cff14: APPROVE (non-blocking). Its one medium finding was a test gap: the branch where a later line of a response replaces an earlier one, which prices nearly every real response, had no test of its 1-hour writes or event time. That is fixed in 7ff00eab, with the tests and negative controls above. The low and informational findings are answered in the PR comment of the same day.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[JasonYeYuhe]

Review follow-up (2026-10-07), head 7ff00eab

The 10-07 review approved 8a5cff14 with one medium finding and five low or informational ones. One commit, 7ff00eab, fixes the medium one. The rest are answered below. This PR is not merged.

Fixed: the branch that prices nearly every real response is now tested (medium)

Claude Code writes a response over several lines and the last one counts. So almost every real response is priced in the branch where a later line replaces an earlier one. Before this fix, the end-to-end tests wrote one line per response and never reached that branch. Dropping line: pricing there still compiled and passed every test.

• Pending.line has no default now. Every place that builds a Pending names the line that prices it. The two places that carry a row already priced say .alreadyPriced. Leaving the argument out is now a compile error.
• test_aResponseWrittenOverSeveralLinesIsPricedFromItsLastLine, one read:
  • An Opus 5.5 response written over three lines, where only the last line carries the 1-hour split, costs $1.084. Priced from its first line it would be $0.904.
  • A two-line Sonnet 4.6 response with a 301K prompt, dated before 2026-03-13, costs $0.231. At today's rates it would be $0.123.
  • The response counts once, with the output of its last line.
• test_aResponseWhoseLastLinesArriveInALaterReadIsRepricedFromThem, two reads:
  • The first read sees only the response's first line and charges $0.704. The test also checks that this read leaves the response open.
  • The second read is incremental, picks up the remaining lines, and charges $1.084.

Negative controls. Each mutation changes only the replace branch in parseClaudeFile.

mutation	where	failed
the replacing line priced as .alreadyPriced, the reviewer's mutation	CI nc/p647-fix-a, run 37500820418	3 of 4,041 tests: test_aResponseWrittenOverSeveralLinesIsPricedFromItsLastLine ($0.904 and $0.123) and test_aResponseWhoseLastLinesArriveInALaterReadIsRepricedFromThem ($0.904)
the replacing line keeps the earlier line's pricing (line: existing.line)	this Mac	both new tests ($0.904 each); 79 other tests pass
a line that replaces an open row carried from an earlier read keeps that row's priced flag	CI nc/p647-fix-b, run 37500827434	1 of 4,041 tests: only test_aResponseWhoseLastLinesArriveInALaterReadIsRepricedFromThem. The row is stored without a cost, so the day falls back to pricing its sum with every write at the 5-minute rate, which gives $0.904

Run locally, the first and third mutations failed exactly as on CI. Both nc/ branches are deleted.

CI on 7ff00eab: Swift CI run 37500794263 (full matrix) passed every job. CLIPulseCore ran 4,041 tests with 0 failures and 5 skipped, again under the Japanese, ROC and Buddhist calendars, plus 4,079 with 0 failures under -DDEVID_BUILD. Both new tests passed in all five passes. The push run 37500785984 also passed. CI Gate timed out after 40 minutes waiting for the queued Swift CI run and was re-run once Swift CI finished; attempt 2 passed. All 32 checks are green (2 skipped, as on every push run); the PR is blocked only by REVIEW_REQUIRED and is 0 commits behind main. On this Mac: the full suite ran 4,041 tests with 0 failures, the filtered suites 104 with 0 failures, and SwiftLint found 0 violations.

Not changed in this PR

• What's New wording (low). The 1.56 What's New is not written yet. The requirement is recorded in the 1.56 release notes: in all 7 store locales it should say that Claude estimates change, with some models lower and some higher, not that they rise. The PR body's "Effect on this Mac" paragraph now says the same.
• A Claude line in the Mac's "figures changed" note (low). Decision: not in 1.56; What's New covers the disclosure.
  • The note's "counted the old way through …" line comes from the archive's record of which days had their Codex slice retaken (Count Codex cached input once in the usage history, and say on the Mac that Codex figures changed #622). Recount the older Codex days of the usage history under the 1.56 rules, on the Mac and in this Mac's cloud rows #649 is reworking that record.
  • A Claude line would need its own record of which Claude days were re-priced.
  • Count a Claude response once from its last line, and skip proxy estimates #624 already changes Claude figures in 1.56 without a note, so a single Claude note would have to cover both changes.
  • What users will see in the meantime: archived Claude days older than the roughly 31-day routine read keep the old, borrowed prices. The Usage Dashboard and the iPhone history will show a step in Claude cost at that boundary.
• Coordination with Port two of CodexBar's subagent and fork rules, and say on the Mac what is still simplified #652 (low). A trial git merge-tree against 7ff00eab still conflicts only in the doc block next to the version constants in CostUsageCache.swift. Whichever of the two PRs merges second rebases. It then re-runs test_claude_rate_changes_come_with_a_rules_version_bump and the Codex fingerprint pin on the merged tree. Neither pin should move, because 7ff00eab changes no rate and no name resolution.
• A device's day and provider replace their cloud rows, so a renamed model stops counting twice (v0.84, not applied) #648 / v0.84 (low). No change here, as the review suggests. Keep A device's day and provider replace their cloud rows, so a renamed model stops counting twice (v0.84, not applied) #648's 404 fallback and apply v0.84 as planned.
• Desktop parity (info). Recorded as a desktop follow-up: port ClaudePricingTable to src-tauri/src/pricing.rs. That means the rows, the cache-hit multipliers, 1-hour writes, the per-request tier and the 4.6 dated rates, ideally with a fixture shared with the Mac. It does not block 1.56.
• Public repo data (info). The PR body no longer quotes this Mac's 30-day dollar totals; only the percentages remain. Haiku 3.5 can get its own row later, with its own Claude rules version bump. Until then claude-3-5-haiku-* stays unpriced.

🤖 Generated with Claude Code