Skip to content

release 0.47.0 — 2026-09 pricing (Fable 5.1 cache reads, Sonnet 5, 1M windows, fast mode off the wire) + the docs the UI arc shipped without - #108

Merged
lroolle merged 2 commits into
mainfrom
feat/pricing-2026-09
Sep 2, 2026
Merged

release 0.47.0 — 2026-09 pricing (Fable 5.1 cache reads, Sonnet 5, 1M windows, fast mode off the wire) + the docs the UI arc shipped without#108
lroolle merged 2 commits into
mainfrom
feat/pricing-2026-09

Conversation

@lroolle

@lroolle lroolle commented Sep 2, 2026

Copy link
Copy Markdown
Member

Release PR for 0.47.0. Two commits on top of main (#104 #105 #107 merged today).

Pricing follows the 2026-09-01 docs (src/pricing.ts, src/summarize.ts, src/cost.ts)

  • Cache reads on Claude Fable 5.1 / Mythos 5.1 are 0.025x input ($0.25 per MTok; Fable 5 / Mythos 5 stay $1) - the embedded offline table now says so. models.dev (the live source) carried the change within the day, so live pages were already right.
  • Sonnet 5's $2/$10 launch price is the standard price; the planned $3/$15 increase was cancelled.
  • Long context on Claude 4.6+ is standard-rate across the full 1M window - what cctrace always assumed. The window's offline fallback follows the same rule: Opus 4.6+/5, Sonnet 4.6/5, Fable/Mythos = 1M; earlier tiers and every Haiku = 200k; an unversioned family = 0 (no %, never a wrong one).
  • Two modifiers are read off the wire in ONE place (pairRates) so the per-pair cost and the cache-bump arithmetic agree: fast mode - the response's usage.speed is "fast" (Opus 5 / 4.8; a silently downgraded request says "standard" and is not fast) - doubles every rate with the cache multipliers on top; US-only inference (inference_geo: "us") is 1.1x on every class. The cost tooltip names the modifiers. No existing trace carries either field.
  • Tests: pricing (5.1 cache read, Sonnet 5, pairRates, costTitle), cost (a fast-mode step bills 2x; a downgraded one does not), summarize (the two fields), context (the window fallback table). One grammar test that pinned Opus 4.6 at 200k now accepts the 1M window.

Release surfaces

Verified: bun test 900 pass / 0 fail, make build clean.

🤖 Generated with Claude Code

lroolle and others added 2 commits September 1, 2026 21:31
…M windows, fast mode and us-inference off the wire

The embedded offline table (models.dev is the live source and already
carried these) follows the 2026-09-01 pricing docs: cache reads on
Claude Fable 5.1 / Mythos 5.1 are 0.025x input ($0.25; Fable 5 stays
$1), Sonnet 5's $2/$10 launch price is the standard price. Long
context on Claude 4.6+ bills at standard rates across the full 1M
window — what cctrace always assumed — and the window fallback now
says 1M for Opus 4.6+/5, Sonnet 4.6/5 and Fable/Mythos (a 431k prompt
on Fable 5 read "100% of context used" against a wrong 200k), 200k
for earlier tiers and every Haiku, 0 for an unversioned name.

Two modifiers are read off the wire in ONE place (pairRates), so the
per-pair cost and the cache-bump arithmetic price the same request the
same way: fast mode — the response's usage.speed is "fast" (Opus 5 /
Opus 4.8; a silently downgraded request says "standard" and is not
fast) — doubles every rate with the cache multipliers on top; US-only
inference (inference_geo: "us") is 1.1x on every token class. The cost
tooltip names the modifiers. No existing trace carries either field.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Version 0.47.0. README (en + zh) replay bullet rewritten for the
trajectory bar (turn blocks, both-way sync; the loop-row diagram is
gone), the Context deck names the inspector, an insights bullet, the
2026-09 pricing note; docs/web-ui.md: the inspector, the bar, replay,
the cost modifiers; the version badge's fresh-off-the-wire list.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@lroolle
lroolle merged commit e55ef13 into main Sep 2, 2026
1 check passed
@lroolle
lroolle deleted the feat/pricing-2026-09 branch September 2, 2026 04:34
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant