Issue 25414 reproduction on main 184d915d87, Ubuntu fs-24, 2026-10-26.
Runtime: vp run dev --home-dir /tmp/issue-25414/t3-home --no-seed
Ports: web 5733, server 13773. Browser viewport 1280x800.
Fixture: python3 seed-usage.py /tmp/issue-25414 (same fixture seeded before Usage first read).
UI: /usage, Cost, 30 days, exclude OpenCode to show only fixture data.
Codex rollout: gpt-6-sol, input_tokens=300000, cached_input_tokens=280000,
output_tokens=1000, total_tokens=301000.
Rate fixture: input=2e-6/cache read=2e-7/output=1e-5; above_272k rates
input=4e-6/cache read=4e-7/output=1.5e-5.
Observed UI: $0.11, 301K processed, 280K cached, 20K uncached, 1K output.
Observed real server UsageSummary response: costUsd=0.106; input=.04,
cacheRead=.056, output=.01. Full sanitized bucket in before-runtime.json.
Expected above-tier estimate: .08 + .112 + .015 = .207.
No provider credentials or paid request required; this follows reporter's rollout-fixture variant.

Red tests at 93b6571d6e:
UsageService.test.ts, prices a Codex turn above 272K prompt tokens at the long-context rates
AssertionError: expected 0.106 to be close to 0.207 +/- 1e-9
usagePricing.test.ts, prices a request above the model's long-context threshold at its above-tier rates
Claude >100K: expected 0.0375 to be close to 0.07250000000000001
Priority >272K: expected 0.212 to be close to 0.41400000000000003