共有:

Anthropic Flagship

Claude's Top Tier
Just Became Fable 5.1.

On September 1, 2026, Anthropic announced Claude Fable 5.1 as its new flagship. Reports put its per-token rate at roughly double Opus 5's, yet Anthropic itself claims up to a 45% cost cut. We unpack the seeming contradiction.

AI Navigate Editorial2026.09.087 min read

Haiku Sonnet 5 Opus 5 Fable 5.1 NEW FLAGSHIP
01
The Handover

The Top Spot Moves From Opus to Fable

This isn't a routine version bump — the identity of "the flagship" itself shifted.

On September 1, 2026, Anthropic announced Claude Fable 5.1 and Claude Mythos 5.1. Fable 5.1 (API model ID: claude-fable-5-1) now sits a tier above Opus 5 (API model ID: claude-opus-5), which had effectively been the top model until last month.

Anthropic has periodically shipped a higher-tier "Fable" line above its regular Opus/Sonnet/Haiku triad. This release is the latest of that line, and it matters less as a raw performance bump than as a change in default routing logic: the model you reach for on your hardest task is now Fable 5.1, not Opus 5.

02

Fable 5.1 By The Numbers

Up to 45%
Effective cost cut on equivalent tasks (Anthropic's own figure)
Capped at 67
GPT-6 Astra's coding score — Fable 5.1 scored higher
2026.09.01
Fable 5.1 / Mythos 5.1 official announcement date
03
The Paradox

How "Pricier" and "Cheaper" Are Both True

Per-token rate and per-task cost are not the same number.

Anthropic's official announcement claims both improved coding performance and up to a 45% cost cut, while reported API rates put Fable 5.1 at roughly double Opus 5's per-token price. The two figures aren't actually contradictory — they're likely measuring per-token rate versus effective cost per completed task. A higher per-token rate can still lower total spend if it takes fewer retries and fewer output tokens to reach the same quality answer.

A September 7, 2026 GIGAZINE report on benchmarking firm Artificial Analysis found that GPT-6 Astra, despite a top-tier coding score, capped out at 67 points and fell short of Fable 5.1. Fable 5.1's exact score wasn't disclosed, but it keeps beating rival flagships in head-to-head comparisons. Separately, a technical breakdown on Zenn reports that Fable 5.1's "low effort" mode alone matches the previous Fable 5's maximum-effort output — one plausible source of the claimed cost savings.

04
Who Benefits

Who It Helps, And How

Engineers

Try the low-effort mode on lower-difficulty tasks first and measure actual cost against your Opus 5 logs — no need to flip every workload at once.

Procurement / Leadership

Don't judge this on the rate card alone — pull one to two weeks of real usage data before revisiting your plan. Anthropic's 45% figure assumes it reproduces on your own workload.

Individuals / Light Use

If Opus 5 already covers your needs, there's no rush to switch. The context window is unchanged from the old Fable 5, so long-document work sees little added benefit.


It's not that it got cheaper —
the rate and the total moved in opposite directions.


05
Caveats

What's Still Unconfirmed

The 45% cost-cut figure comes from Anthropic itself; no independent third-party verification has surfaced yet. Depending on the workload, low-effort mode may fall short on quality, forcing a return to the costlier setting. The Astra comparison also only discloses one side of the scoreboard — Astra's 67 — while Fable 5.1's own score stays undisclosed. And the "roughly double" pricing claim is still based on reporting and estimates rather than a confirmed side-by-side of official rate cards.

06

Opus 5 vs. Fable 5.1: What Changed

Opus 5Fable 5.1
Effective flagship until last monthNew top tier, above Opus
API model ID: claude-opus-5API model ID: claude-fable-5-1
The pricing baselineReported at roughly double the rate
Context windowUnchanged from the old Fable 5
07
Next Steps

What To Do Next

01

Start with low-effort mode

Test Fable 5.1's low-effort mode on low-to-medium difficulty tasks and check the quality gap against Opus 5.

02

Measure a week of real cost

Compare actual per-task spend from your call logs, not the rate card, to check whether the 45% figure holds for your workload.

03

Hold off for long-context work

Since the context window is unchanged, don't rush to switch for very long documents — pick the model per use case instead.