Claude Fable 5.1: the highest index score measured, and the cache cut that decides your bill
Every figure below is traceable: each comes from a named, dated public source or from Scalarware's own live catalogue, listed in full at the end.
TL;DR
- Claude Fable 5.1 launched 1 September 2026 and scores 66 on the Artificial Analysis Intelligence Index at max effort — the highest figure that index has recorded.
- Sticker price is unchanged: $10 input / $50 output per million tokens, matching Fable 5 and GPT-6 Astra.
- Cache reads dropped 75%, from $1.00 to $0.25 per million. Cache writes are $12.50 per million.
- Per task it is 20% more expensive than Fable 5 ($3.76 vs $3.14) because it emits roughly 1.7x more output tokens — the cache cut only wins it back if your hit rate is high.
- The gains concentrate in long, tool-using work: Terminal-Bench-Science is reported at 52.6% against Fable 5's 24.7%. On CursorBench the gap to Opus 5 is 3.4 points, on a model costing twice as much.
Anthropic released Claude Fable 5.1 on 1 September 2026. It scores 66 on the Artificial Analysis Intelligence Index at max effort — the highest measured — ahead of Opus 5 (63), Fable 5 (62) and GPT-5.6 Sol (61). Sticker price is unchanged at $10 / $50 per million tokens, but cached reads fell 75% to $0.25.
The score
| Model (max effort) | AA Intelligence Index |
|---|---|
| Claude Fable 5.1 | 66 |
| Claude Opus 5 | 63 |
| Claude Fable 5 | 62 |
| GPT-5.6 Sol | 61 |
| Grok 4.6 | 61 |
Three points clear of the next model is a real margin on this index. But the composite hides where the movement actually happened.
Where the gains actually are
The largest published improvements are in long, tool-using work rather than single-turn answers. On Terminal-Bench-Science the reported jump is 24.7% to 52.6% against Fable 5 — more than double. That is the number most outlets led with, and it is the one that justifies the model for agentic workloads.
The number fewer outlets ran: on CursorBench the gap to Claude Opus 5 is reported at 3.4 points, on a model that costs twice as much. Read the index as evidence about long agentic runs, which is where the separation is real, rather than as a uniform ranking across all work.
The economics, which are not what the sticker suggests
| Measure | Fable 5 | Fable 5.1 |
|---|---|---|
| Sticker / 1M in–out | $10 / $50 | $10 / $50 |
| Cache read / 1M | $1.00 | $0.25 (−75%) |
| Cache write / 1M | — | $12.50 |
| Cost per Intelligence Index task (max) | $3.14 | $3.76 (+20%) |
Per task it is more expensive, because it generates roughly 1.7x more output tokens. The cache cut pulls the other way and is reported to save around $1.40 per task on agentic workloads.
Which effect dominates is a property of your workload, not of the model. A long-running agent that re-reads a large stable context across many turns wins outright. A one-shot call that never hits cache pays the 20% and receives nothing back. Cache writes at $12.50 per million also mean caching a context you only read once is a loss — the break-even is roughly a second read.
Working out which side you land on
You do not have to estimate this. The API returns token usage per call, so a single run tells you your real cache ratio:
import os
from openai import OpenAI
client = OpenAI(
base_url="https://scalarware.com/api/v1",
api_key=os.environ["SCALARWARE_API_KEY"],
)
result = client.chat.completions.create(
model="anthropic/claude-fable-5.1",
messages=[
{"role": "system", "content": LARGE_STABLE_CONTEXT},
{"role": "user", "content": "Next step?"},
],
)
usage = result.usage
print("prompt tokens ", usage.prompt_tokens)
print("completion tokens", usage.completion_tokens)
# $10 / $50 per 1M, cache reads $0.25 per 1M
cost = (usage.prompt_tokens * 10 + usage.completion_tokens * 50) / 1_000_000
print(f"uncached cost ${cost:.5f}")Run the same prompt twice. If the second call's input is dominated by a context you resend unchanged, Fable 5.1's cache pricing is working for you and the 20% per-task premium is already paid back.
Comparing it against the alternatives on one key
Every model below is live on the Scalarware catalogue at 0% markup on language-model inference, behind one OpenAI-compatible endpoint. Switching is a one-string change.
| Model | Model id | Input / 1M | Output / 1M |
|---|---|---|---|
| Claude Fable 5.1 | anthropic/claude-fable-5.1 | $10.00 | $50.00 |
| Claude Opus 5 | anthropic/claude-opus-5 | $5.00 | $25.00 |
| GPT-5.6 Sol | openai/gpt-5.6-sol | $2.00 | $10.00 |
| Meta Muse Spark 1.3 | meta/muse-spark-1.3 | $1.25 | $4.25 |
curl https://scalarware.com/api/v1/chat/completions \
-H "Authorization: Bearer $SCALARWARE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "anthropic/claude-fable-5.1",
"messages": [{"role": "user", "content": "Plan the migration, then execute step one."}]
}'When Fable 5.1 is worth $10 / $50
- Yes: long agentic runs, heavy tool use, a large context re-read across many turns. The index lead is concentrated here and the cache cut compounds here.
- Probably not: single-turn chat and classification. You pay the 20% output premium and never touch the cache discount.
openai/gpt-5.6-solat $2 / $10 scores 61 on the same index. - Check first: coding assistants. The CursorBench gap to Opus 5 is 3.4 points at twice the price — measure on your own repository before paying the difference.
- Cost-first:
meta/muse-spark-1.3at $1.25 / $4.25 is an eighth of the input price. Meta claims near-parity on agentic and coding work; that claim is cheap to test.
Fable 5.1 is days old. Vendor benchmark tables are self-reported and independent evaluation is still arriving. Treat this as the published evidence as of 4 September 2026.
Frequently asked questions
- What does Claude Fable 5.1 score on the Artificial Analysis Intelligence Index?
- 66 at max effort — the highest score the index has recorded, ahead of Claude Opus 5 (63), Claude Fable 5 (62), GPT-5.6 Sol (61) and Grok 4.6 (61).
- How much does Claude Fable 5.1 cost?
- $10 per million input tokens and $50 per million output on Scalarware, at 0% markup. Cached reads are $0.25 per million — a 75% cut from Fable 5's $1.00 — and cache writes are $12.50 per million.
- Is Fable 5.1 cheaper than Fable 5?
- Not per task. It costs about $3.76 per Intelligence Index task against Fable 5's $3.14, roughly 20% more, because it emits about 1.7x more output tokens. The 75% cache-read cut saves roughly $1.40 per task on agentic workloads, so whether you come out ahead depends on your cache hit rate.
- What is the model id for Claude Fable 5.1 on Scalarware?
- anthropic/claude-fable-5.1, called through the OpenAI-compatible endpoint at https://scalarware.com/api/v1/chat/completions.
- Where does Fable 5.1 improve most over Fable 5?
- Long, tool-using work. Terminal-Bench-Science is reported at 52.6% against Fable 5's 24.7%. On CursorBench the gap to Claude Opus 5 is only 3.4 points, so the advantage is not uniform across all tasks.