fix(server): usage prices Claude Mythos 5.2 from its Bedrock rate#33721

Problem

Usage shows Claude Mythos 5.2 (claude-mythos-5-2) as Unpriced, even after Refresh usage forces a rate refetch. LiteLLM has no bare claude-mythos-5-2 row yet. It lists the model only under Bedrock IDs: anthropic.claude-mythos-5-2 at $8/M input and $40/M output, plus global., us./eu. (+10%), us-gov. and bedrock/us-gov-*/ (+20%). parseRateTable only derives a bare alias by dropping the text before the last /, so no row ever becomes claude-mythos-5-2.

Reproduced on main 6c9adfb273 with the web dev server (fresh database), using a synthetic Claude transcript with 100K input and 20K output tokens per model. Usage → Cost → Breakdown shows claude-mythos-5-2 as Unpriced, while claude-mythos-5-1, claude-fable-5-1 and claude-fable-5-2 are priced.

The issue also reports Fable 5.2 as unpriced. A rate cache predating the bare Fable 5.2 row could explain that report. Today's table prices Fable 5.2 directly (it shows $1.60 on both sides of the evidence below), and Refresh usage already refetches rates, so this PR does not change the 24h rate refresh.

Change

Rate-table keys without a / now also get a bare alias by dropping Bedrock's anthropic. prefix. Only the base Bedrock ID qualifies. Regional and GovCloud IDs (us., eu., global., us-gov.) keep their own names, so their markups never compete with the base rate. The Bedrock base row ranks like any other cloud row: an exact bare row still wins, a first-party anthropic/ row still beats it, and disagreeing candidates of equal priority still leave the name unpriced.

Scope and approval

Fixes pingdotgg/t3code#33715

Verification

• New test prices a bare name listed only under Bedrock IDs at the base anthropic rate in apps/server/src/usage/usagePricing.test.ts parses copies of the seven real Mythos 5.2 rows in both orders. It fails on main (expected null to match object { inputCostPerToken: 0.000008, … }) and passes here. vp test run apps/server/src/usage: 148 passed; vp run check:changed passes on the final head.
• Regression diff of old and new parseRateTable over a pinned copy of LiteLLM's full model_prices_and_context_window.json (4,893 rows, SHA-256 42677c2c…0710325): 0 rates changed, 0 removed, 16 aliases added (claude-mythos-5-2 plus 15 Bedrock-versioned or legacy Claude names such as claude-opus-4-1-20250805-v1:0). Snapshot, scripts and report.

Evidence

Before = main 934c7f8d8e, after = 0957a043af (this branch before its last main merge, which touched no usage code). Each side runs its own dev server on a fresh database (T3CODE_DEV_SEED=0), reads the same synthetic Claude transcript, and loads the same pinned LiteLLM snapshot from the rate cache, so no refresh hits the network. Fixtures: A four models with 100K input + 20K output each; B two Mythos 5.2 records with cache read/write, one fast; C two Mythos 5.2 records, one carrying a reported costUSD. Fixture matrix.

Native steps, identical on both sides (driver-elapsed seconds before/after; playback runs about 2 s behind): M1 Cost/30d → scroll 7/8 → bottom 18/20 → Tokens/30d 35/38 → 24h 45/47 → Cost/24h 55/57. M2 Cost/30d → scroll 9/11 → bottom 21/25 → Tokens/30d 36/47 → 24h 45/56. On both sides, switching to 24h triggers the app's existing transcript rescan: "Scanning provider transcripts…" replaces the chart for under a second (playback 47.2 s before / 49.3 s after on iOS, 47.2 s / 58.3 s on Android, each under 0.5 s) before the 24h figures render. This PR does not change it.

W1: before shows claude-mythos-5-2 Unpriced and $5.60 total; after shows $1.60 and $7.20 total

Environment. Web: headless Chrome 151 on Linux (agent-browser), no device emulation, DPR 1, font scale 100%, reduced motion, clock frozen at 2026-11-10 06:50 UTC. Browser time zone Europe/Paris on both sides for W1–W6 (hourly labels read 4 AM for 03:50 UTC records). The W4 quality pair was retaken with UTC held through CDP and --disable-lcd-text, because the blurred backdrop recomposited text and lit 1–2/255 noise in the header. Native: iPhone 17 Pro, iOS 27.2, 402 pt, font scale 1.0, dark; Pixel 9, Android 16, 411 dp, font scale 1.0, dark. Both use the dev client from vp run mobile:install, and the client clock is frozen identically on both sides.

What else changed. Everything that depends on price changed together: headline and provider totals, model cost cells and shares, ranking and bars, chart scales and tooltips, the unpriced notice (25% → gone in A), cost-by-type and cost-by-speed sections (appear in B), cache savings ($0 → $0.78 in B), and the model dialog's statistics, footer, and so its height and shadow. In C the reported $1.60 is kept and the other record adds $1.60, so the total is $3.20 and "Other" splits into Input/Output. Token and session counts are unchanged everywhere. On native, the same mid-page swipe can stop at different scroll limits, and bottom-anchored views shift earlier sections by the height of the new Cost section (about 50 pt in M1 details30, 262 pt in model30). Top-of-page views share one scroll offset. Nothing outside pricing changed.

Rebuild: seed.py, servers.py, driver.py, browser-init.js, browser-launch.sh, W4 retake, Android, iOS; full manifests: web, Android, iOS.

AI-assisted; verified with the tests and tools listed above.