Fable 5.1 in Claude Code: more power for 25% less

TL;DR Fable 5.1 shipped on September 1, 2026 and it needs Claude Code v2.1.257. Below that version the fable alias keeps handing you Fable 5 without a word. The per-token price hasn't moved ($10 input / $50 output per MTok): what dropped is cache reads, from $1.00 to $0.25 per million. Since rereading your history in a long session weighs tens of times more than writing it, my three longest sessions come out 42%, 44% and 48% cheaper at list price.

Anthropic announced the launch with a 25% cost drop on typical workloads and up to 45% on agentic work. Launch-note percentages are usually rounded optimism, so before writing any of this I applied both rates to my own sessions.

It lands on the 45%, not the 25%. And it comes down to a single line of the pricing table.

The only thing that moved

A cache read costs 10% of the input price on every Claude model. On Fable 5.1 and Mythos 5.1 it costs 2.5%. That is the entire pricing change.

Why that line is worth so much in Claude Code

Because in Claude Code that line is the bill. Every turn pays for the whole conversation again: writing your context into the cache is a one-off, rereading it is charged on every turn the session has left. The longer and warmer the session, the more that column runs the show.

Across my three longest sessions, rereads ran 37 to 58 times larger than writes:

session         turns      writes           rereads      Fable 5   Fable 5.1   drop
longest one      8212    80,726,017   3,000,633,827       $5236       $2986      42%
yesterday's      1739     8,138,384     469,689,508        $727        $374      48%
an older one      471     3,422,224     141,868,141        $237        $131      44%

There's a second effect you only see with Opus 5 next to it. Fable 5 cost exactly double Opus 5 on all four billing lines, not on average. Fable 5.1 lands 3% to 14% above it, because on the column that matters it now costs half what Opus 5 does ($0.25 against $0.50):

session        Opus 5    Fable 5    Fable 5.1
longest one     $2618      $5236        $2986   (+14% over Opus 5)
yesterday's      $363       $727         $374   (+2.9%)
an older one     $118       $237         $131   (+10.2%)

The spread isn't random: the more rereads outweigh writes, the closer it sits to Opus 5. Yesterday's session reread 58 times what it wrote and came out 2.9% above.

These are API list-price dollars, not what a subscription charges you, and they assume one-hour cache writes, which is what Claude Code uses. You can pull the same numbers from any session of your own:

jq -s '[.[] | select(.type=="assistant") | .message | select(.usage)]
       | unique_by(.id)
       | {writes:  (map(.usage.cache_creation_input_tokens // 0) | add),
          rereads: (map(.usage.cache_read_input_tokens // 0) | add),
          input:   (map(.usage.input_tokens // 0) | add),
          output:  (map(.usage.output_tokens // 0) | add)}
       | {fable_5:   ((.writes*20 + .rereads*1.00 + .input*10 + .output*50) / 1000000),
          fable_5_1: ((.writes*20 + .rereads*0.25 + .input*10 + .output*50) / 1000000)}' \
  ~/.claude/projects/<your-project>/<session>.jsonl

The unique_by(.id) isn't optional: without it you count every streamed response twice.

And the power half of the headline

The model itself moves up, and the docs say where: long agentic sessions (multi-file refactors, migrations, code review running for hours), work on documents and spreadsheets, multistep research, reading charts and tables buried in PDFs, reasoning across the full million-token window, and computer use. The docs also flag that the gap widens at higher effort levels, which is exactly where Claude Code puts it by default (high). The knowledge cutoff jumps from January to June 2026.

What pushes the other way

Two default-behaviour changes the press note skips and the docs don't. Both raise output tokens, which is the $50 line:

1. It batches fewer tool calls. Where Fable 5 fired three reads at once, Fable 5.1 may fire one per turn. Answer quality holds, but you pay for it in turns, tokens and wall-clock. Asking fixes it:

When you intend to make several independent tool calls, send them in the same message.

2. It rewrites whole files for small changes. Editing text, it leans toward redoing the entire file instead of touching the one line. The result is usually identical, but you pay the difference in output:

For small changes to a file, make a targeted edit rather than rewriting the whole file.

Both belong in your CLAUDE.md. It also answers from memory more often at low effort, so raise the level for anything that needs fresh information.

How to find out where you stand

1. Update, or it doesn't exist

claude update
claude --version
# 2.1.257 (Claude Code)

The model catalog is baked into the binary. I checked both versions on the same day, same account, same alias:

2.1.257 → canonicalModel: "claude-fable-5-1"
2.1.252 → canonicalModel: "claude-fable-5"

Both calls succeeded. Neither warned about anything. Skip the update and you stay on Fable 5, and the only place you notice is the bill.

2. Check what your alias resolves to

/model fable              # Fable 5.1 from v2.1.257 on
/model claude-fable-5-1   # pins the version and removes the surprises

best points at Fable 5.1 too. The exception is the gateway, where fable stays on Fable 5.

Reference

Official docs: What's new in Claude Fable 5.1 · Introducing Claude Fable 5.1 and Claude Mythos 5.1

Fable or Opus? I wrote my routing across the three back when Fable cost double. That arithmetic just changed. Why does the cache dominate your bill? Here's the whole mechanism, and why a turn turns slow and expensive the moment you break it. Coming from the classifier interruptions? They came back with Fable 5, and on 5.1 they drop by around 60% per session, with vulnerability identification now allowed. And the model you already had set? Opus 5 doesn't reset your effort, so whatever level you carry over is the one that applies.

Requirements

• Claude Code v2.1.257 or later (claude update).