v0.3.5 — Claude 5.5 pricing and output-token fixes

This release corrects the cost figures again, mostly downward. v0.3.4 billed the Claude 5.5 models at older rates and missed a third of the output tokens. On the maker's logs, all-time spend went from $42,141 to $35,447, and October from $16,276 to $8,976.

Updating from v0.3.4: the update button in v0.3.4 does nothing, so download and run BurnmeterSetup.exe once. This release fixes the button for the next update. The first launch re-reads all logs once, which takes about 5 minutes on a large history.

Downloads: Windows BurnmeterSetup.exe. macOS Burnmeter.dmg is attached by CI a few minutes after publishing; it is unsigned (right-click → Open).

Fixed

• Claude 5.5 models were billed at older rates. Opus 5.5 was priced as Opus 5,
  so its cache reads, the bulk of a Claude Code bill, cost 2.5x. Sonnet 5.5 cache
  reads cost 2x, and Haiku 5.5 was billed at Haiku 4.5 rates (10x). Model ids now
  match exactly. Haiku 5.5 prompts over 100K tokens bill at its higher tier, fast
  mode bills at 2x, and an unknown version is still flagged. On the maker's logs,
  October fell from $16,276 to $8,976.
• Output tokens were under-counted by a third. Claude Code logs a response's
  usage as it streams. Only the last line has the full count, and v0.3.4 kept the
  first. The lines of one response are now merged at their highest values. On the
  maker's logs, output tokens rose from 122.8M to 179.2M. With both fixes,
  all-time spend went from $42,141 to $35,447.
• Days, weeks and months follow the local clock. "Today" used to reset at UTC
  midnight, which is 03:00 at UTC+3, so it disagreed with the activity heatmap.
• A device that stopped syncing kept adding to the fleet. A Mac last synced in
  July still added its July burn rate, "today" and month to "All devices". An old
  snapshot now adds nothing to those three, and its card and chips show "—".
• One-click update did nothing in the installed app. It ran
  burnmeter-server.exe -m burnmeter _update, which exited at once, and the
  dashboard still said "✓ Updated" after 150 s. The app now downloads and runs the
  release installer for the version the dashboard offered. /api/health reports
  progress and failure, and a failed update can be retried from the pill.
• Scanning another log tree wiped the shared parse cache. This happened when
  running the tests or passing a second --projects-dir, and the next launch then
  re-read every log, which takes minutes on a big history. Entries for other trees
  are now kept, and an unchanged cache is not rewritten. The Codex cache works the
  same way.
• Tool counts grew with uptime. A build served from the parse cache added each
  multi-line response's tool calls again, which inflated the per-tool numbers and
  grew the cache file on every rebuild.
• The Chats "goal" column showed "Tool loaded." or image notes. Claude Code
  also logs request copies, meta lines and compaction summaries as user messages.
  The goal is now the first message the user typed. The model in use now comes
  first in the models column.
• A forked or resumed session counted its tool errors twice. Errors are now
  de-duplicated per tool call.
• Cache savings price each call at the rates it actually paid, including fast
  mode and Haiku 5.5's long prompts.
• Smaller fixes:
  • Model names keep their minor version ("Opus 5.5", not "Opus 5").
  • Model names and timestamps from synced devices are escaped before they reach
    the page.
  • The dashboard skips its charts instead of breaking when Chart.js can't load.
  • A log line stamped outside the OS clock's range no longer takes down the report.
  • The cache card counts cache writes as input, so its hit rate matches the rest
    of the dashboard. It read 100.0% next to 98.2% on the Insights page.
  • Chat costs and durations are no longer cut off: "$3,33…" now reads "$3,339",
    and a chat resumed over weeks shows its span in days.
  • The pace next to the burn gauge now uses the needle's own measure. It used to
    project a block's first minutes over all 5 hours, and could read "10.4×" next
    to a needle in the yellow.

Changed

• Charts work offline. Chart.js 4.4.7 is now bundled, so the dashboard loads
  nothing from a CDN.
• The statusline is instant while Burnmeter runs. It asks the app instead of
  rebuilding the report: 0.5 s, against about 65 s on the maker's history. Without
  the app it reads only the last 26 hours, in about 9 s. It then blanks what
  26 hours can't show: the pace dot (⚪), "P90+", and the block when that day has
  no 5-hour pause. The line is now in English:
  🟠 Opus 5.5 $4.20/h · block 4h24m · 0 err/24h · cache 97% · today $32.
• The parse cache version was bumped, so the first launch after upgrading does one
  full re-read.