Skip to content

feat(server): offer Claude Fable 5.2, Sonnet 5.5 and Haiku 5.5 in the model picker - #1002

Merged
edspencer merged 2 commits into
mainfrom
feat/fable-5-2
Nov 14, 2026
Merged

edspencer merged 2 commits into
mainfrom
feat/fable-5-2

Conversation

@edspencer

@edspencer edspencer commented Nov 14, 2026 •

Copy link
Copy Markdown
Owner

Adds the three newest Claude models to the catalog. Figures are from the models overview and pricing pages, checked 2026-11-14:

Model ID Context Input / output per MTok Cache read Supersedes
Fable 5.2 claude-fable-5-2 1M $8 / $40 2.5% → $0.20 Fable 5.1 ($10/$50)
Sonnet 5.5 claude-sonnet-5-5 1M $2 / $10 5% → $0.10 Sonnet 5 (same rate, 10% cache)
Haiku 5.5 claude-haiku-5-5 1M $0.10 / $0.50 10% → $0.01 Haiku 4.5 ($1/$5, 200K)

Every superseded model stays in the catalog, so existing per-project pins keep resolving, and the default is unchanged (Opus 5.5).

Two things worth a reviewer's attention

1. Haiku 5.5 is a 1M-context model, unlike the 200K Haiku 4.5 it supersedes. The picker's hint said "Opus/Fable/Sonnet: 1M; Haiku: 200K", which this makes false, so it now reads "every model here is 1M except Haiku 4.5 (200K)". That is the only packages/web/src change, and the only user-visible copy change.

2. Haiku 5.5's cost estimate is pinned at its base tier, and under-estimates long chats. It is the one model priced by prompt length: a request whose prompt exceeds 100k tokens bills at $0.50/$2.50 instead of $0.10/$0.50, each request judged on its own. estimateCostUsd prices a chat's cumulative TokenTotals, which carry no per-request prompt lengths, so it cannot tell the tiers apart — a long Haiku 5.5 chat is under-estimated by up to 5×. I chose the base tier (exact for short chats) over the long tier (wrong by 5× for every short chat), documented it at the catalog entry, and pinned it with a test so a future change is deliberate. The fix needs a TokenTotals shape change, so it is #1003 rather than folded in here.

Visual evidence

Project → Settings → Claude. Real instances, both built from source and served on this box; no state injected beyond creating the workspace. Left-hand column is the change; everything else in frame is a control.

Before (origin/main, a5af152):

before

After (this branch):

after

The hint rewrites, and Offered models gains Fable 5.2, Sonnet 5.5 and Haiku 5.5. Permission mode, Max turns, Docker sandbox, Drive mode and Max spawn depth are unchanged in the same frame.

Control — the Max spawn depth block, cropped from each build. These two PNGs are byte-identical (md5 7d6d06411de39ac8ff9b04aa6030301f), so the layout did not shift under the new rows:

Before After
control before control after

Build provenance. The changed copy lives in a code-split chunk, not the top-level bundle, so I grepped each served build for both strings: the old Opus/Fable/Sonnet: 1M is present in before's ProjectView-C1qr6nlt.js and absent from after; the new every model here is 1M is present in after's ProjectView-BvSCdn7s.js and absent from before.

Verification

  • npm run typecheck clean. npm test: server 2727/2727, web 2315/2315.
  • /api/models on the two live instances — before: Opus 5.5 | Opus 5 | Opus 4.8 | Fable 5.1 | Fable 5 | Sonnet 5 | Haiku 4.5; after: the same plus Fable 5.2, Sonnet 5.5 and Haiku 5.5 in catalog order.
  • Selecting Fable 5.2 in Settings and saving persisted model: claude-fable-5-2, confirming server-side validation accepts a new id.
  • No real turn was run on any new model: the QA instances are credential-free by design.

Not in this PR

SWEEPER_DEFAULT_MODEL still pins Haiku 4.5. Haiku 5.5 is cheaper and larger, so moving the sweeper is probably right — but it changes behaviour for every existing instance on upgrade, so it wants its own decision.

🤖 Generated with Claude Code

Fable 5.2 (claude-fable-5-2, released 2026-11-09) joins the catalog above
Fable 5.1: 1M context, $8/$40 per MTok (20% under 5.1), and the same 0.025x
cache-read rate. Fable 5.1 and Fable 5 stay so existing pins keep resolving.
The default model is unchanged (Opus 5.5).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@cloudflare-workers-and-pages

cloudflare-workers-and-pages Bot commented Nov 14, 2026 •

Copy link
Copy Markdown

Deploying paddock with  Cloudflare Pages  Cloudflare Pages

Latest commit: 8ece885
Status: ✅  Deploy successful!
Preview URL: https://952f47e3.paddock-7u2.pages.dev
Branch Preview URL: https://feat-fable-5-2.paddock-7u2.pages.dev

View logs

Sonnet 5.5 (claude-sonnet-5-5): 1M context, $2/$10, 0.05x cache reads.
Haiku 5.5 (claude-haiku-5-5): 1M context, $0.10/$0.50 at its base tier.

Haiku 5.5 is a 1M-context model, unlike the 200K Haiku 4.5 it supersedes,
so the picker hint no longer claims Haiku means 200K.

Haiku 5.5 is also the one model priced by prompt length ($0.50/$2.50 over a
100k-token prompt). estimateCostUsd prices cumulative totals, which carry no
per-request prompt length, so it is pinned at the base tier and a long chat
is under-estimated. Documented at the catalog entry, pinned by a test, and
tracked in #1003 rather than folded in here.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@edspencer edspencer changed the title feat(server): offer Claude Fable 5.2 in the model picker feat(server): offer Claude Fable 5.2, Sonnet 5.5 and Haiku 5.5 in the model picker Nov 14, 2026
@edspencer
edspencer merged commit 532a43d into main Nov 14, 2026
10 checks passed
@edspencer
edspencer deleted the feat/fable-5-2 branch November 14, 2026 15:17
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