What is Claude Fable 5.1?

Claude Fable 5.1 is Anthropic's new flagship Claude model, released on September 1, 2026 as the successor to Fable 5. It is available via the Claude API, Amazon Bedrock, Google Vertex AI and Microsoft Foundry. Anthropic says it is the world's most advanced model for coding and knowledge work, topping all 8 public benchmarks it was evaluated on.

How much does Claude Fable 5.1 cost?

The base price is unchanged from Fable 5: $10 per million input tokens and $50 per million output tokens. The headline change is cache reads, cut 75% from $1.00 to $0.25 per million tokens. Cache writes are $12.50/M (5-minute) and $20/M (1-hour); batch mode is half price at $5/$25.

Why does the cache-read price cut matter?

Long agentic sessions re-read cached prefixes (system prompts, codebases, tool definitions) on every turn, so cache reads dominate cost. Anthropic's example: an agent re-reading a 100K-token context 50 times costs $5.00 in cache reads under Fable 5 pricing but only $1.25 under Fable 5.1. Typical workloads get ~25% cheaper and heavily agentic workloads up to ~45% cheaper.

What are Fable 5.1's benchmark scores?

On Terminal-Bench 4.0 (agentic coding) Fable 5.1 scores 55.8%, up from 42.0% for Fable 5 and above Opus 5's 52.3% and GPT-5.6 Sol's 37.3%. On Terminal-Bench-Science (agentic scientific research) it scores 52.6%, more than double Fable 5's 24.7%. SWE-bench Verified is 95.0%, CursorBench 73.4%, LiveCodeBench 90.52% and HLE 59.1%. All scores are Anthropic self-reported with production safeguards enabled.

What is Claude Mythos 5.1?

Mythos 5.1 is the same underlying model as Fable 5.1 but wrapped in stricter safety guardrails, and it is restricted to verified cybersecurity and life-science organizations. On Terminal-Bench 4.0 it actually scores higher (60.9%). Regular developers and consumers cannot access it.

Does Fable 5.1 cost less than GPT-5.6 Sol or Opus 5?

Not on the sticker: Fable 5.1 is $10/$50 while GPT-5.6 Sol dropped to $4/$20 in August 2026. The savings come from context-heavy agent loops where the $0.25/M cache read is far cheaper than competitors, so the total bill for long-running agents can come out ahead.

Who should upgrade to Claude Fable 5.1?

Developers and teams doing serious coding, long-running AI agents, or knowledge work — especially Claude Code users and enterprise customers who want Enterprise Frontier Safeguards (data on your own cloud). Casual chat users should stick with lighter models; the $50/M output means a single complex real coding task can still cost $20-50.

Claude Fable 5.1 review 2026 — Anthropic's coding flagship with 75% cheaper cache reads

Claude Fable 5.1 Review 2026: The Coding King That Also Got 75% Cheaper to Run

Data verified: September 2, 2026. All benchmark figures are Anthropic self-reported with production safeguards enabled, not independently verified.

On September 1, 2026, Anthropic shipped a new flagship: Claude Fable 5.1, the successor to Fable 5, alongside a restricted sibling called Mythos 5.1. The model topped all 8 public benchmarks it was evaluated on — pushing agentic coding (Terminal-Bench 4.0) from 42.0% to 55.8% and more than doubling agentic scientific research to 52.6%. Anthropic calls it "the world's most advanced model for coding and knowledge work."

But the sticker price didn't move. The real headline is the 75% cut in cache-read pricing — from $1.00 to $0.25 per million tokens — which is what makes long-running, tool-heavy agent sessions genuinely cheaper. In this review we separate the marketing from the measurable: what Fable 5.1 actually improves, what it costs in real agent workloads, where it still hurts, and whether you should upgrade.

The Verdict in One Sentence

Fable 5.1 is the strongest coding and agentic model Anthropic has shipped — a benchmark-topping upgrade that also cuts long-session costs by 25–45% via a 75% cheaper cache read — but it keeps the same premium $10/$50 sticker price, its scores are self-reported, and a single complex real coding task can still cost $20–50.

Core Specs at a Glance

Spec	Claude Fable 5.1
Release date	September 1, 2026 (successor to Fable 5)
Model ID	`claude-fable-5-1` (Bedrock: `anthropic.claude-fable-5-1`)
Context window	1,048,576 tokens input (1M) · 128K max output
Pricing	$10/M input · $50/M output (unchanged); cache read $0.25/M (−75%)
Thinking mode	Always on; effort levels low / medium / high / xhigh / max
Availability	Claude API, Amazon Bedrock, Google Vertex AI, Microsoft Foundry
Restricted sibling	Mythos 5.1 (same model, stricter guardrails, cybersecurity/life-science orgs only)

Benchmarks: Where It Actually Moved

Anthropic reports Fable 5.1 topped all 8 public benchmarks. The meaningful gains concentrate in long-horizon agentic work:

Benchmark	Fable 5.1	Fable 5	Opus 5	GPT-5.6 Sol
Terminal-Bench 4.0 (agentic coding)	55.8%	42.0%	52.3%	37.3%
Terminal-Bench-Science (research agents)	52.6%	24.7%	29.0%	22.4%
SWE-bench Verified	95.0%	—	—	—
CursorBench 3.2	73.4%	70.5%	—	—
LiveCodeBench	90.52%	—	—	—
HLE	59.1%	—	—	—

The Terminal-Bench-Science jump from 24.7% to 52.6% is the most striking number — it signals that Fable 5.1 genuinely got better at long, self-directed work (research a problem, run tools, iterate), not just at answering single-turn questions. For comparison, our earlier review of GPT-5.6 put Sol at 37.3% on the same agentic coding bar, well behind here.

Honesty note: these are Anthropic's own numbers, run with production safeguards on, and not yet independently reproduced. Treat them as directional.

Pricing: Same Sticker, Cheaper Cache

The sticker price is unchanged, which makes this a price-neutral upgrade in capability but a real cost cut in practice:

Item	Fable 5.1	Fable 5
Input	$10 / 1M	$10 / 1M
Output	$50 / 1M	$50 / 1M
Cache read	$0.25 / 1M (−75%)	$1.00 / 1M
Cache write	$12.50 / 1M (5-min) · $20 / 1M (1-hr)	Same
Batch	$5 / $25	Same

Why cache reads dominate: an agent loop re-reads its system prompt, codebase and tool definitions as a cached prefix on every single turn. Anthropic's worked example:

Agent re-reads a 100K-token context 50 times → cache cost $5.00 under Fable 5, $1.25 under Fable 5.1.
Net effect: typical workloads ~25% cheaper; heavily agentic workloads up to ~45% cheaper.

That reframing — from "capability race" to "long-running agent economics" — is exactly the industry shift we flagged when Cohere Command A+ and GPT-5.6 landed. Anthropic reportedly eyes an IPO (valuations floated up to $2 trillion) as early as late September, so cost-per-task matters to its growth story.

Effort Levels: The Practical Cost Lever

Fable 5.1's thinking mode is always on — you can't disable it. Instead you pick an effort level: low, medium, high, xhigh, or max.

Default: high in Claude Code, medium in Claude Cowork and on Claude.ai.
Low / medium can already match or beat Fable 5 on many tasks, per Anthropic — the cheap path for routine work.
High → max trades more tokens for deeper reasoning; per-task cost scales roughly from a few dollars up to $30+ depending on effort.

For teams this is the real budget dial: run simple things at low/medium, escalate to high/max only when a task actually needs it.

Enterprise & Safety: Safeguards and Anti-Distillation

Enterprise Frontier Safeguards — your data lives on your own cloud infrastructure; positioned squarely at government and large-enterprise buyers.
Anti-distillation — protections against training rival models on Claude outputs (edited conversation history invalidates preserved thinking; thinking blocks are bound to the producing model).
Breaking API changes to be aware of: forced tool use (tool_choice: any/tool) now returns a 400; preserved thinking is tied to the model that produced it.

As a real-world signal, Cognition said it moved Devin's Opus 5 traffic to Fable 5.1 on launch day after it matched or edged Fable 5 at lower cost — the strongest third-party validation available so far.

Scorecard

Dimension	Score	Note
Coding & agentic capability	9.5/10	Top of Terminal-Bench 4.0 & Science; SWE-bench 95%
Long-session cost efficiency	9/10	Cache read −75% → agent tasks up to −45%
Context & output	9/10	1M in / 128K out, no long-context surcharge
Sticker price	6/10	$10/$50 unchanged; pricier than GPT-5.6 Sol on paper
Transparency	6.5/10	Self-reported benchmarks, no independent reproduction yet
Accessibility	7.5/10	Every major cloud; but Mythos 5.1 restricted

Pros & Cons

What Fable 5.1 does really well

Benchmark-topping agentic coding: Terminal-Bench 4.0 55.8%, SWE-bench 95%, LiveCodeBench 90.5%.
Genuinely cheaper long sessions: cache reads at $0.25/M make big-context agent loops economical.
Research agents more than doubled (24.7% → 52.6% on Terminal-Bench-Science).
Effort dialing: low/medium can match Fable 5 for less — a real cost lever.
Same sticker, more model: no price increase for a clear capability jump.
Enterprise-ready: Frontier Safeguards, data on your cloud, every major cloud platform.

Where it falls short

Still premium-priced: $10/$50 sticker; GPT-5.6 Sol is $4/$20 after its August cut.
Single-task cost is real: one complex real coding run can still hit $20–50 even with good cache hits.
Self-reported benchmarks with production safeguards — no independent verification yet.
Breaking API changes may require adapting existing tool-use/thinking code.
Overkill for chat; the high cost makes it wrong for casual or lightweight workloads.

Frequently Asked Questions

Should I switch from Fable 5 to Fable 5.1?

Yes if you run coding or agent workloads — it's strictly better on benchmarks and the cache cut lowers your long-session bill. If you only send short single-turn prompts, the benefit is mostly the improved raw capability, not the cost math.

Is Fable 5.1 the best coding model right now?

By Anthropic's own numbers, yes — it tops the public agentic-coding and general-knowledge bars it was tested on, ahead of Opus 5 and GPT-5.6 Sol. Independent third-party reproduction hasn't caught up yet, so treat it as "best available evidence," not settled truth.

How do I make Fable 5.1 cheaper to use?

Maximize cache hits on long contexts (reuse cached prefixes), and pick the lowest effort level that solves the task — low/medium already beats Fable 5 on many tasks. Batch mode halves all prices if you can tolerate latency.

Can I get Mythos 5.1?

Only if you're an organization verified in cybersecurity or the life sciences, for which it's restricted. Everyone else gets Fable 5.1.

Where is Fable 5.1 available?

Claude API (model ID claude-fable-5-1), Amazon Bedrock, Google Vertex AI, Microsoft Foundry, the Claude app, and Claude Code (it's the default after update, running at high effort).

Bottom Line

Claude Fable 5.1 is the model to use for serious coding and agentic work in September 2026. It tops the benchmarks, and the 75% cache-read cut quietly makes long-running agents the most economical they've been on Anthropic's platform. It is not a bargain model — $10/$50 is unchanged and GPT-5.6 Sol undercuts it on sticker — so the calculus favors teams who live in long, context-heavy sessions.

If that's you, this is a rare "more capable and cheaper to run" upgrade. If you're mostly chatting or sending short prompts, Fable 5.1's power is wasted — keep using a lighter model and revisit when you actually need an agent that works all day.

Score: 8.7/10 — Recommended for serious coders and agent builders; price-sensitive or light users should look at lower tiers.

搜索此博客

ring