8 things that changed in Claude Fable 5.1 (vs Fable 5), ranked by size

Anthropic shipped Claude Fable 5.1 on September 1, 2026, its first update since Fable 5 launched on June 9. The context window is the same 1 million tokens. Max output is the same 128,000 tokens. Base input and output pricing are the same $10 and $50 per million tokens. So what actually changed? Eight things, and they range from "notice it in your invoice next week" to "notice it in a lab that maps Venus."

This is the same underlying model released two ways. Claude Fable 5.1 is generally available with Anthropic's standard safeguards. Claude Mythos 5.1 is the restricted-access sibling, gated behind Anthropic's Cyber Verification Program and Life Sciences Verification Program, a partnership with the US government for vetted cybersecurity and life-sciences organizations. We cover the model access split separately in how to access Claude Fable 5 and Mythos 5; this post is about what's different in the 5.1 update itself, ranked by how much it actually moves the needle for a team already running Fable 5 in production.

The 8 changes at a glance

• Cache read pricing dropped 75%: from $1.00/MTok to $0.25/MTok, the single biggest lever if you run agentic workloads with repeated context.
• Terminal-Bench-Science 0.1 more than doubled: 24.7% to 52.6%, the largest single benchmark jump in the release.
• A new sibling model, Mythos 5.1: restricted access, and it beats Fable 5.1 on agentic coding.
• Anthropic's own recommendation shifted: Opus 5 is now the default pick for most workloads, not Fable 5.1.
• Humanity's Last Exam (no tools) rose: 57.8% to 60.9%.
• CursorBench 3.2.0 rose: 70.5% to 73.4%.
• OSWorld 2.0 (partial) rose: 72.9% to 77.9%, the biggest jump among the general benchmarks.
• Named scientific and real-world results Anthropic cited at launch: protein design, Venus mapping, genome analysis costs, and one investment firm's root-cause debugging story.

Source for all benchmark and pricing numbers: Anthropic's launch announcement and the model overview docs. We've also added Fable 5.1 to our own AI model comparison hub if you want to see it stacked against every other current model on the same page.

1. Cache read pricing dropped 75%

Cache reads went from $1.00/MTok on Fable 5 to $0.25/MTok on Fable 5.1. Everything else on the price sheet held steady: input is still $10/MTok, output is still $50/MTok, batch API is still 50% off base rates. Cache writes are $12.50/MTok for the 5-minute tier and $20/MTok for the 1-hour tier.

Why this is #1 on the list: it's the only change on this page that hits your bill without you changing a single line of code. Anthropic states this brings typical-workload costs down about 25% and highly agentic workload costs down up to about 45% versus Fable 5 at identical base prices. That "up to 45%" number specifically targets teams running long agent loops with repeated context, exactly the workload we build for clients doing custom workflow automation. If your agent re-reads the same system prompt and tool definitions on every step, a 75% cut on cache reads compounds fast.

2. Terminal-Bench-Science 0.1 more than doubled

24.7% to 52.6%. No other benchmark in Anthropic's comparison table moved anywhere close to this. Terminal-Bench-Science measures scientific computing and research-coding tasks in a terminal environment, and the jump lines up with the scientific use cases Anthropic highlighted at launch (more on those in #8).

Worth noting: Fable 5.1 still isn't the top scorer on every science-adjacent benchmark. On Terminal-Bench 4.0, the general agentic-coding test, Mythos 5.1 posted 60.9% against Fable 5.1's 55.8%. That gap is why item #3 exists.

3. Mythos 5.1 is a real sibling model, not a rebrand

Fable 5.1 and Mythos 5.1 are the same underlying model with different safety filters. Fable 5.1 ships generally available with Anthropic's standard safeguards. Mythos 5.1 is restricted to organizations vetted through Anthropic's Cyber Verification Program (cybersecurity) or Life Sciences Verification Program, run in partnership with the US government.

The reason to know this if you're not eligible for either program: Mythos 5.1 outscores Fable 5.1 on Terminal-Bench 4.0 (60.9% vs 55.8%). If you've read that Fable 5.1 "won" every benchmark Anthropic published, that's not quite right, its restricted sibling wins one of them. For most teams this is academic since Mythos access requires formal vetting, but if you're evaluating vendor claims, it's the kind of detail worth getting right.

4. Anthropic now points most workloads to Opus 5, not Fable 5.1

This is a positioning change, not a benchmark number, and it's easy to miss if you only skim the comparison table. Anthropic's own docs now recommend Opus 5 for most workloads, reserving Fable 5.1 for "demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5 at higher effort still fall short."

That's a shift from how Fable 5 was positioned at its June launch, as the default top-tier pick. If you're choosing between the two today, that's the actual decision criterion Anthropic wants you using, not "which model is newer." We break down the full pricing and use-case tradeoff in Fable 5 vs Opus 4.8 and Opus 5 vs Fable 5 if you want the side-by-side.

5. Humanity's Last Exam rose from 57.8% to 60.9%

This is the no-tools reasoning benchmark, a 3.1-point gain. It's a smaller jump than the science benchmark but a meaningful one on a test that's specifically designed to resist memorization and reward genuine reasoning depth.

6. CursorBench rose from 70.5% to 73.4%

A 2.9-point gain on the coding-assistant benchmark version 3.2.0. Smallest of the five general benchmark deltas Anthropic published, but it's the one most directly relevant if your team uses Fable 5.1 or 5.1-based agents inside an IDE workflow rather than a pure terminal agent loop.

7. OSWorld 2.0 (partial) rose from 72.9% to 77.9%

A 5.0-point gain, the second-largest jump on this list after the science benchmark. OSWorld tests an agent's ability to operate a real computer desktop environment: opening apps, navigating UI, completing multi-step tasks outside a sandboxed terminal. This is the benchmark most directly tied to computer-use agent work, which is the category behind a growing share of the AI agent development projects we scope for clients right now.

8. The named scientific and real-world results Anthropic cited at launch

Anthropic didn't just publish benchmark tables. It named specific outcomes:

• Protein design hit close to a 50% success rate across 12 targets, with binding affinities roughly 10 times higher than prior approaches.
• Venus surface mapping resolution improved from 10-20 kilometers to 2-3 kilometers.
• Computational biology work saw GPU speedups of up to 2.5 times, with 30-60% cost savings on genome-wide analyses.
• Millennium, an investment firm, reported that Fable 5.1 identified the root cause of a rare system crash that its own engineers and other models had failed to solve over several years (reported by 9to5mac, Zac Hall, Sep 1 2026).

These are Anthropic's and third-party reported figures, not independently reproduced by us, and we're citing them as claims rather than facts we've verified ourselves. Still, they're the concrete evidence behind item #2's science benchmark jump, and they matter more to a research team than any single percentage point on a leaderboard.

Should you switch from Fable 5 to Fable 5.1?

If you're already on Fable 5 through the API, AWS Bedrock, Google Cloud, or Microsoft Foundry, the practical answer is: switch the API identifier to claude-fable-5-1 and re-run your evals. Nothing about integration changed, context window, max output, and base pricing are identical, so this isn't a migration, it's a model swap.

Where it's worth checking your evals specifically:

• Agentic workflows with repeated context: the cache pricing cut alone can justify the switch on cost grounds before you even factor in the accuracy gains.
• Scientific or research coding tasks: the Terminal-Bench-Science jump is large enough that it's worth a direct A/B on your own task set, not just trusting the aggregate number.
• Computer-use agents: the OSWorld gain is the second-largest jump published, relevant if your agents interact with real UIs rather than terminals or APIs.
• General reasoning or coding where Opus 5 already clears your bar: per Anthropic's own updated guidance, don't default to Fable 5.1 here. Test Opus 5 at higher effort first; it's priced lower ($5/$25 per MTok vs $10/$50).

If you're weighing Fable 5.1 against non-Anthropic alternatives entirely, our Fable 5 alternatives roundup still covers the field accurately. The pricing and benchmark deltas in this post add to that comparison, they don't replace it.

Standing up an evaluation harness against your own tasks, rather than trusting a vendor's benchmark table, is exactly the kind of groundwork an AI strategy consulting engagement covers before a model swap goes into production. Book a discovery call if you want a second set of eyes on the eval before you flip the switch.

FAQ

What's new in Claude Fable 5.1 compared to Fable 5? Cache read pricing dropped 75% (from $1.00/MTok to $0.25/MTok), five benchmarks improved (most notably Terminal-Bench-Science 0.1, which more than doubled from 24.7% to 52.6%), and Anthropic released a restricted-access sibling model, Mythos 5.1. Context window, max output, and base input/output pricing are unchanged.

Is Claude Fable 5.1 more expensive than Fable 5? No. Base input ($10/MTok) and output ($50/MTok) pricing are identical. The only pricing change is cache reads, which got 75% cheaper. Anthropic states this makes typical workloads about 25% cheaper overall and highly agentic workloads up to about 45% cheaper.

What is Claude Mythos 5.1? The same underlying model as Fable 5.1, released with restricted access instead of Fable 5.1's standard safeguards. Access requires vetting through Anthropic's Cyber Verification Program (cybersecurity organizations) or Life Sciences Verification Program (life-sciences organizations), run in partnership with the US government.

Should I use Fable 5.1 or Opus 5? Per Anthropic's own current guidance, Opus 5 is the default for most workloads. Reserve Fable 5.1 for demanding reasoning or long-horizon agentic work, or cases where Opus 5 at higher effort still underperforms on your specific evals. See our Opus 5 vs Fable 5 comparison for the full breakdown.

Did the context window or output limit change in Fable 5.1? No. Both are unchanged from Fable 5: a 1,048,576 token (1M) context window and 128,000 max output tokens.

Where can I access Claude Fable 5.1? Claude Code, the Claude Platform (Enterprise), the API (identifier claude-fable-5-1), AWS Bedrock, Google Cloud, and Microsoft Foundry. Full setup steps are in how to access Claude Fable 5 and Mythos 5.

What's the knowledge cutoff for Claude Fable 5.1? June 2026, per Anthropic's model overview docs. Anthropic states it won't retire the model sooner than September 1, 2027.

Is Mythos 5.1 better than Fable 5.1 at coding? On Terminal-Bench 4.0 specifically, yes: Mythos 5.1 scored 60.9% against Fable 5.1's 55.8% in Anthropic's own comparison table. Mythos 5.1 isn't generally available, though, so this only matters if you already qualify for restricted access.