fix: track Claude Opus 5 as a 1M-context model - #81

Problem

claude-opus-5 is reported with a 200 000-token context window, flagged as exact. Because a real Opus 5 session routinely exceeds that, the context indicator doesn't just read high — it reads over 100%, and estimated: false means the UI has no reason to soften it.

Measured on my own logs (31 060 records, via getCurrentContextInfo):

Root cause

contextWindowFor has a [1m]-suffix escape hatch and a major-version check, but the latter only matched sonnet-:

// Sonnet 5.x+ (e.g. "claude-sonnet-5") has no "-4-" segment to match
// above, so check the major version directly. (Opus 5 doesn't exist yet —
// don't guess its window size here.)
if (/sonnet-(?:[5-9]|\d\d)(?:-|\b)/.test(m)) {

That comment is now out of date, and the [1m] hatch doesn't cover the gap: Claude Code writes the bare claude-opus-5 id to its JSONL even when the session is running the 1M-context variant. Across three days of my logs:

7750 "model":"claude-opus-5"      ← no [1m] marker, ever
 723 "model":"claude-opus-4-7"
 672 "model":"claude-sonnet-5"
 637 "model":"claude-opus-4-8"

So claude-opus-5 falls through to the generic /haiku|opus|sonnet/ branch and gets 200K.

Fix

One alternation, covering Opus 5+ alongside Sonnet 5+:

if (/(?:opus|sonnet)-(?:[5-9]|\d\d)(?:-|\b)/.test(m)) {

Older models are unaffected — the (?:-|\b) anchor still prevents a date suffix from matching, so claude-3-5-sonnet-20241022, claude-sonnet-4-5-20250929 and claude-opus-4-20250514 keep their 200K window. There's a regression test pinning that.

claude-opus-5 also gets an explicit MODEL_PRICING entry. No cost figure changes — inferPricingByFamily already resolved it to the correct current-Opus tier. The entry mirrors the existing claude-sonnet-5 anchor and removes a spurious Unknown model: claude-opus-5 line from the diagnostic output, which seemed worth avoiding in an extension whose diagnostics users are asked to attach to bug reports.

Opus 5 pricing ($5 / $25 per MTok) and context window verified 2026-07-28 against https://platform.claude.com/docs/en/about-claude/models/overview — for Opus 5, as for Sonnet 5, 1M is both the default and the maximum, with no smaller context variant.

Tests

Red first, then green. The failing run was actual: 200000, expected: 1000000.

• getCurrentContextInfo reports a 1M window for Opus 5, with or without the [1m] marker
• getCurrentContextInfo keeps the 200K window for pre-4.6 Opus and Sonnet (regression guard)
• getModelPricing resolves Opus 5 from the exact table, long-context variant included

npm test → 172 tests, 171 pass, 0 fail, 1 skipped (the skip is pre-existing).

Also smoke-tested as an installed VSIX in VS Code 1.129.1 on Linux.

Notes for review

• CHANGELOG.md has an entry under 2.2.1 → Fixed. I didn't touch the READMEs: 2.2.1 has no "what's new" section in any of the seven yet, so there was nowhere consistent to put it. Happy to add one if you'd rather it landed there too.
• Not addressed here, but possibly worth its own issue: Opus 5 fast mode bills at $10 / $50 per MTok, not $5 / $25. On Opus 5 that's a speed: "fast" request parameter rather than a distinct model id, so I couldn't tell from my logs whether it's distinguishable in the JSONL at all. If it is, current pricing under-counts those turns.

🤖 Generated with Claude Code