Claude Fable 5.1 Review: Leads Opus 5 on Every Benchmark — And Costs Less

Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026. The naming says .1 — a patch, a touch-up, a Friday afternoon release. The numbers say something else.

Fable 5.1 more than doubles its predecessor on agentic scientific research. It leads Opus 5 on every single benchmark row Anthropic published. At the same sticker price as Fable 5 — $10 per million input tokens, $50 per million output tokens — it costs roughly 25% less for typical workloads and up to 45% less for highly agentic ones.

The pricing cut is real but quiet. The headline is the cache read price: $0.25 per million tokens, down 75% from Fable 5's $1.00. For context-heavy agentic work, cache reads are most of the bill. That's where the savings come from.

And the model itself? It's the same underlying weights as Mythos 5.1. Two doors from one model: Fable 5.1 with full safeguards for general use, Mythos 5.1 with lighter safeguards for vetted cybersecurity and life science organizations.

The Benchmark Table, Row by Row

Anthropic published seven benchmark rows. Here's the table:

Fable 5.1 leads Opus 5 on all seven rows. GPT-5.6 Sol appears on five of seven and trails on all five. On computer use and reasoning, there's no Sol score at all.

A few rows deserve a closer look.

Terminal-Bench-Science 0.1 — 52.6% for Fable 5.1 versus 24.7% for Fable 5. More than double. Standard error is ±3.5 to 4.5 points per model, so the gap survives any reasonable error bar. The public leaderboard (3 trials per task, Claude Code harness) has Opus 5 at 30.0% and Fable 5 at 21.4%; Anthropic says its setup reproduces both within noise at 29.0% and 24.7%.

Terminal-Bench 4.0 — 55.8% for Fable 5.1, 60.9% for Mythos 5.1. Same model, five-point gap. The gap is tasks where Fable 5.1's cyber safeguards stepped in and Mythos 5.1's didn't. Anthropic expects the more precise safeguards shipping today to shrink the gap. For anyone buying the generally available model, 55.8% is the honest number — an 8-point lead over Opus 5 on the row closest to predicting real terminal-agent work.

OSWorld 2.0 — the fine print matters here. Task files are from the benchmark authors' August 2026 release, so these numbers aren't comparable to previously published OSWorld results. That's why no competitor score is shown. And both Anthropic models scored zero wherever their safeguards intervened. The strict-mode 41.7% is the number that says computer use still has a long way to go. The 5-point lead over Opus 5 is the number that says Fable 5.1 leads anyway.

AutomationBench — 31.4% for Fable 5.1 versus 17.1% for Fable 5. Nearly double. Footnote: Fable 5 scored zero wherever its safeguards intervened on this benchmark, so part of the doubling is subtraction on the old model's side. And 31.4% is still 31.4% — autonomous business workflows remain unsolved for everyone.

CursorBench 3.2.0 — 73.4% for Fable 5.1, the smallest gap in the table at 3 points over Opus 5's 70.0%. Coding leadership is contested now. The .1's edge is that it leads while using less of your budget.

The Zeros Behind the Table

Fable 5.1 was evaluated with its production safeguards enabled. Where safeguards intervened, both Fable 5.1 and Fable 5 scored zero on OSWorld 2.0, and Fable 5 scored zero on AutomationBench. In other interventions, cybersecurity tasks were completed by Claude Opus 4.8 and biology tasks by Claude Opus 5. Anthropic flags this as likely depressing both models' scores on the affected rows.

Treat those rows as floors, not ceilings. The raw capability is higher than the table shows.

The Safeguards Got More Precise

Same release, same day. Cyber safeguards now block 60% fewer false positives than before. Fable 5.1 can be used to discover software vulnerabilities — though still not to develop exploits for them. Penetration testing, exploit generation, and binary-based vulnerability scanning still route to Opus models.

In biology, safeguards fire 85% less often on benign medical and elementary biology questions. Research-grade life sciences work moves to Mythos 5.1 through the new Life Sciences Verification Program, built with the US government. Anthropic expects to open enrollment for scientists soon.

One more compliance note: as a model released after August 2, 2026, Fable 5.1's outputs carry Anthropic's statistical text watermark — the EU AI Act watermark, invisible without their detection API.

The Price Cut That Isn't a Price Cut

Sticker price unchanged: $10 per million input tokens, $50 per million output tokens. The cut is underneath, in cache reads.

Cache reads now cost $0.25 per million tokens — 75% less than Fable 5. Over four weeks of August 2026 usage at default effort, indexed cost comes out to roughly 75 for typical workloads (25% cheaper) and roughly 55 for highly agentic workloads (45% cheaper), with Fable 5 at 100 on both scales.

The market noticed fast. Walden Yan at Cognition said "with the new cache read pricing a Fable-class model is finally economical for the workloads we'd kept on Opus, starting with code review." Cognition moved Devin's Opus 5 traffic to Fable 5.1 on launch day.

Effort settings matter too. Fable 5.1 defaults to High effort in Claude Code, Medium in Claude Cowork and on Claude.ai. Anthropic claims Low and Medium effort already match Fable 5's full-effort results at substantially lower cost.

The Root-Cause Stories

Benchmarks are one thing. The qualitative jump shows up in specific incidents.

Millennium — an investment firm had a crash that hit roughly once in a million runs, unexplained by their engineers and every model they tried, including Fable 5, for four to five years. Fable 5.1 disassembled an external vendor library, matched it against the core dump, and traced the crash to a bug in that library. Damien, a senior portfolio manager there, said the time it would have taken to conduct that analysis is hard to justify.

MongoDB — Ron Sanzone, a staff software engineer, said Fable 5.1 built a complex prototype in about three days: initial research across all of their services code and documentation, a novel and extensible design, then hours of unattended work with strong verification loops. He'd wake up to the next phase finished, with a full visual walkthrough and clear evidence of success.

Rakuten Medical — Felix Giovanni Virgo, a principal AI engineer, said they asked Fable 5.1 to review a clinical research project that three other frontier models had signed off on. It found a gap none of them had seen, insisted on testing it further, and proposed a completely new hypothesis — turning a dataset they'd written off into a new research direction in one afternoon.

Ramp — Dwight Temple, a senior ML engineer, described a 38-hour unattended run on a machine learning problem that diagnosed a prior result as a label artifact, made the correction, kicked off six parallel experiments that ran overnight, and returned with a result and next steps. Given an open-ended prompt to find the highest-leverage unowned problem, it surfaced an unowned alert tied to a production outage, pulled the logs, and prescribed the fix.

These are Anthropic's chosen quotes. They're not independently verified. But they're specific — named people, named companies, specific incidents — which is more than you usually get from a model launch announcement.

The Science Pitch

Anthropic spent the announcement on the science angle, and the results are the kind that don't fit a leaderboard.

Fable 5.1 was used to design novel protein binders targeting three proteins — EGFR, Nipah G, and 15-PGDH — from Adaptyv Bio's protein design competitions. The hit rate for the designs that were tested was nearly 50%. For Nipah G, the comparison is against de novo designs targeting the receptor-binding site on the G head, where the best achieved roughly 8–12 nM; Fable 5.1's designs reached comparable affinity.

That's the strategic story Anthropic is telling with Mythos-class models: not just coding tools, but research instruments. Protein design, GPU kernel optimization (2.5x improvement in one case), and a 2-3 km resolution map of Venus. The coding benchmarks are the entry point; the science is the pitch.

Jensen Hua, Anthropic's head of hardware, separately published a Model Hardware Standard research preview — a different thread, but part of the same month's output from the company.

Fable 5.1 vs Claude Opus 5: Which One Do You Actually Want?

If you've been using Opus 5 as your daily driver, the Every team's vibe check title says it plainly: "Anthropic Is So Back (Again)." Dan Shipper's first message to the team was "I can actually understand what it's saying." On their Slack assistant it used less than half the tokens of Opus 5 with comparable results.

Kieran Klaassen's verdict: "this model is Fable for everyone" — a Fable you can use in the loop instead of only for long hauls. Dan Shipper: "It's friendly Fable. Fable-level intelligence, Opus-level price, Sonnet-speed. In our tests it was about twice as fast as Opus 5 and used half as many tokens."

The counter-evidence is real too. Asked for 1,000 words it wrote 1,288. At Extra-high effort it sometimes kept working after being interrupted. On Every's tests the one job where it trailed was the X post, where GPT-5.6 Sol led every Anthropic model. And on Hacker News, where the launch thread passed 400 points in its first hour, Anthropic's own Felix Rieseberg said the underrated upgrade is prose — it "sounds a lot less stereotypically like other Claude models" and follows style instructions more reliably.

So the answer depends on your workload:

• Coding, terminal work, agentic research: Fable 5.1 leads Opus 5 and costs less. The case is strong.
• Knowledge work where Opus 5 is already good enough: the GDPval gap is 29 points, about 1.6%. Marginal.
• Computer use in production: strict-mode 41.7% tells you it's not solved yet, but Fable 5.1 leads anyway.
• Cost-sensitive agentic workloads: the cache read pricing makes Fable 5.1 the economic choice over Opus 5 for the workloads Cognition was keeping on Opus.

Claude Mythos 5.1: The Same Model, Restricted

Mythos 5.1 is Fable 5.1 with lighter safeguards, available only to vetted organizations through the Cyber Verification Program and the Life Sciences Verification Program — currently US-only, with Anthropic coordinating with the US government to expand access to a broader set of domestic and international partners.

On Terminal-Bench 4.0, Mythos 5.1 scores 60.9% versus Fable 5.1's 55.8%. The gap is the safeguard-intervened tasks. Anthropic expects today's safeguard update to shrink it.

For anyone not in the trusted access programs, Fable 5.1 is the model. Mythos 5.1 is a restricted variant of the same thing.

The Breaking Changes for Fable 5 Users

Three breaking changes if you're migrating from Fable 5:

1. Forced tool use is no longer supported. A tool_choice of type any or tool returns a 400, including on the token counting endpoint.
2. Thinking blocks are bound to the model that produced them. Fable 5.1 reads earlier models' blocks, but no earlier model reads Fable 5.1's, and unreadable blocks are dropped unbilled.
3. Editing an earlier turn invalidates every later thinking block — enforced for accounts created on or after August 31, 2026. Mythos 5.1 does not run this check.

Five additions are backward compatible: per-message effort changes mid-conversation, turn-scoped system messages (clear_at: "next_user_message"), progress updates between tool calls via thinking.display: "updates" (all three in beta), the lower cache read price, and content provenance — Anthropic's statistical text watermark on generated text and signed C2PA Content Credentials on files retrieved through the Files API.

Fable 5.1 is available to all customers on the Claude API as claude-fable-5-1, plus Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Both Fable 5.1 and Mythos 5.1 carry 30-day data retention, are not available under zero data retention unless expressly authorized, and are Covered Models.

Fable 5 and Mythos 5 move to legacy status.

What This Means in the Context of September 2026

Read Fable 5.1 alongside this month's other releases and a pattern emerges.

September 1: Anthropic ships Fable 5.1 and Mythos 5.1 — leads Opus 5 on every benchmark, 45% cheaper for agentic work.

September 3: OpenAI ships GPT-6 Astra — 99.9% on ARC-AGI-3, 100% on ExploitBench, Greg Brockman calls it AGI.

September 10: OpenAI ships the Agents API in public beta — the Codex harness as a managed product.

The model layer and the harness layer are both moving. Anthropic's bet is on the model — better benchmarks, better pricing, science as the long-term pitch. OpenAI's bet is broader — better models plus a managed harness plus GPT-Live-1 plus financial services verticalization, all in the same week.

For developers choosing between them right now: Fable 5.1 is the better Claude, and it's cheaper than the Claude you were probably using. GPT-6 Astra is the more capable model on the benchmarks that matter for agentic work, and the Agents API is the harness to run it with. The choice between the two stacks is real, and it's not just about which model scores higher on which benchmark.

Published Tuesday, September 15, 2026. Sources: Anthropic's "Introducing Claude Fable 5.1 and Claude Mythos 5.1" announcement (September 1, 2026); the Claude Fable 5.1 and Mythos 5.1 system card; Vellum's benchmark walkthrough; the Every vibe check; Anthropic's Enterprise Frontier Safeguards announcement; coverage by AI Weekly, Vellum, MarkTechPost, and Hacker News. Follow AIPress for ongoing coverage of AGI, ChatGPT, Claude, OpenAI, Anthropic, and the AI industry.