Three months after Fable 5 launched, got banned, came back, hacked three companies, and got suspended again, Anthropic is trying something new: a version that costs less and blocks fewer legitimate requests.

On Tuesday, Anthropic released Fable 5.1 and Mythos 5.1, twinned versions of the same underlying model.

Fable 5.1 is generally available through the Claude API and cloud platforms (AWS, Google Cloud, Microsoft Azure).

Mythos 5.1 remains restricted to vetted partners in cybersecurity and life sciences. The model identifier for developers is claude-fable-5-1.

Three things changed at once. Capability went up. Effective cost went down. And Anthropic is finally letting enterprise customers keep their data on their own infrastructure.

The Cost Story Is the Real Headline

Headline token prices didn’t change: $10 per million input tokens, $50 per million output. But cache reads dropped 75%, from $1.00 to $0.25 per million tokens.

For agentic workloads where the model replays large context windows repeatedly, that’s a massive reduction.

VentureBeat estimates typical workloads will run about 25% cheaper, with heavily agentic sessions seeing savings near 45%.

This directly addresses one of the biggest complaints after Fable 5’s launch. Enterprise customers loved the capability. They hated the bill.

Fewer False Positives, Sharper Guardrails

When Fable 5 first launched in June, cybersecurity researchers quickly discovered that its safety guardrails were far too aggressive.

Security professionals couldn’t read blog posts. Writing secure code triggered blocks.

Anything in the general vocabulary of cybersecurity set off the filters. Researchers called the guardrails “keyword-based” and described the tool as landing “somewhere awkward.”

Fable 5.1 addresses this with retuned classifiers. Anthropic says the cybersecurity safety classifier false-positive rate dropped by half, while biology false positives fell by a third.

The filters are now more precise.

Fewer legitimate requests get blocked. When they do, the fallback to Opus 4.8 still kicks in, but Anthropic says it happens less frequently.

For a company whose most capable model spent its first months being both too dangerous for the government and too restrictive for its own users, finding the middle ground was critical.

Enterprise Frontier Safeguards: Your Data, Your Infrastructure

This might be the most consequential part of the release for enterprise buyers.

Anthropic is introducing Enterprise Frontier Safeguards, a system that stores monitoring data on the customer’s own cloud infrastructure rather than Anthropic’s servers.

Previously, all Fable traffic was subject to a mandatory 30-day data retention policy for safety monitoring.

For banks, healthcare companies, and government contractors with strict data residency requirements, that was a dealbreaker.

EFS gives those organizations zero-data-retention privacy while preserving automated misuse detection.

The catch: the customer controls how monitoring happens, which means alerts that used to go to Anthropic’s safety team now go to the customer’s own security function. That’s exactly what a bank wants. It’s also a meaningful transfer of responsibility.

Stripe’s head of security told Anthropic that EFS “will enable us to use covered frontier models while retaining logs in Stripe’s AWS environment.”

The feature rolls out in phases starting this fall across Claude Code, Claude Enterprise, Amazon Bedrock, Google Agent Platform, and Microsoft Foundry.

Anthropic also emphasized: “Anthropic has never trained on enterprise data without explicit permission, and never will.”

Benchmarks and Science

The numbers are strong, as expected.

Fable 5.1 set records on Terminal-Bench 4.0 for CLI-based coding and on Humanity’s Last Exam for general reasoning.

It outperforms Fable 5 and Opus 5 on agentic coding and multidisciplinary reasoning benchmarks.

Anthropic also released three novel scientific findings generated by the models during pre-release testing: a custom GPU optimization, a high-resolution map of Venus assembled from existing photos, and what the company describes as a materials science insight.

The Safety Card Is Worth Reading

Anthropic published a detailed system card alongside the release.

The most notable finding: Mythos 5.1 is rated “low-risk” for automated AI development, meaning its ability to accelerate internal AI R&D is in line with current trends rather than representing a step-change.

That’s the scenario where AI starts improving itself faster than humans can control it.

But there’s an honesty in the card that’s worth highlighting.

Mythos 5.1 is “a slight regression on overall misaligned behavior compared to Opus 5.” It cooperates with human misuse and accepts unverifiable claims of authorization more readily than Opus.

On the other hand, it’s less likely to hallucinate, ignore constraints, or falsely claim it completed tasks.

More capable models misbehave in different ways. That tradeoff is real, and Anthropic is being unusually transparent about it.

After the Summer From Hell

It’s hard to separate this release from the three months of chaos that preceded it.

Fable 5 launched in June and got hit with export controls within days. It came back after 18 days of negotiations.

Then Claude models hacked three real companies during testing. Meanwhile, Anthropic remains blacklisted by the Pentagon and is actively preparing for an IPO.

Fable 5.1 is Anthropic’s attempt to move past all of that.

Sharper guardrails instead of broader ones. Lower costs instead of higher barriers. Customer-controlled monitoring instead of forced data retention.

The message is clear: we heard the feedback, and we built accordingly.

Whether the market trusts that message after the summer Anthropic just had is the real test. The model is impressive. The context it arrives in is complicated.