Skip to main content

Anthropic's New AI Models Double Coding Performance, Slash Costs

Anthropic has just dropped its latest AI models, Claude Fable 5.1 and Claude Mythos 5.1, and they're making some bold claims. The company calls them "the most advanced programming and knowledge work models in the world." Both models share the same underlying architecture, but they differ in security levels: Fable 5.1 is open to everyone, while Mythos 5.1 is reserved for approved organizations in cybersecurity and life sciences through a Trusted Access Program.

Benchmarks That Speak Volumes

In head-to-head tests, Fable 5.1 left its predecessors in the dust. On the Agent Science Research Benchmark (Terminal-Bench-Science 0.1), it scored 52.6%, a massive jump from Fable 5's 24.7% and Opus 5's 29.0%. Even OpenAI's GPT-5.6 Sol only managed 22.4%. For programming, Fable 5.1 hit 55.8% on Terminal-Bench 4.0, while Mythos 5.1, with a more relaxed security setting, reached 60.9%. In knowledge work, Fable 5.1 scored 1853 on the GDPval-AA v2 test, edging out Opus 5's 1824 and Fable 5's 1723.

Cost and Performance: A Win-Win

One of the most striking updates is the price cut. Cache read costs for Fable 5.1 have plummeted by 75%, now sitting at just $0.25 per million tokens. For typical workloads, this translates to about 25% savings compared to Fable 5, and up to 45% for highly agent-based tasks. The base pricing remains unchanged: $10 per million input tokens and $50 per million output tokens. So, you're getting more power without breaking the bank.

Security Gets Smarter

Anthropic has also refined the safety mechanisms. False positives—those annoying times when the AI blocks legitimate requests—have been reduced by about 60% in cybersecurity applications. Fable 5.1 can now help discover software vulnerabilities, but it won't assist in developing exploits for them. In the biological realm, Anthropic partnered with the U.S. government to create an access program, letting researchers tap into Mythos 5.1's advanced biological capabilities. Registration for scientists is expected to open soon.

Real-World Scientific Wins

Mythos 5.1 isn't just about benchmarks; it's already making waves in science. In molecular design, it crafted high-affinity binders using open-source tools, achieving binding affinities ten times higher than the best entries in the Adaptyv Bio Protein Design Competition. On 12 targets, it hit a success rate close to 50%, whereas typical protein design hovers around 10-15%. In computational analysis, Fable 5.1 processed radar images from NASA's Magellan mission, creating a high-resolution elevation map of Venus. The resolution improved from 10-20 kilometers down to 2-3 kilometers, with a 25% boost in elevation accuracy. And in computational biology, Mythos 5.1 sped up seven open-source deep learning models by up to 2.5 times, potentially slashing GPU costs by 30-60%.

Data Retention and Compliance

Anthropic introduced Enterprise-Focused Security (EFS), which combines zero data retention with advanced protections. Customer data stays in cloud infrastructure fully controlled by the enterprise, not on Anthropic's systems. EFS will roll out to enterprise customers this fall, but eligible clients can already use Fable 5.1 under the zero data retention policy.

On the compliance front, Anthropic signed the EU AI Act's practice guidelines in July 2026. This means model outputs released after August 2, 2026, must include an invisible watermark. Fable 5.1 already does this—it's imperceptible to users, doesn't affect quality, and contains no user-specific info. The detection API is in private preview, available to regulators, law enforcement, media, and fact-checkers per EU laws.

Availability

Claude Fable 5.1 is now live on all major platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure, and developers can access it via the Claude API. Claude Mythos 5.1, however, is currently limited to approved U.S. organizations.

Image

Key Points

  • Performance Leap: Fable 5.1 doubles programming benchmark scores compared to its predecessor.
  • Cost Reduction: Cache read costs drop by 75%, making it more affordable for heavy users.
  • Enhanced Security: False positives cut by 60% in cybersecurity; new access programs for biological research.
  • Scientific Impact: Achievements in protein design, Venus mapping, and GPU optimization showcase real-world utility.
  • Compliance Ready: Invisible watermarking meets EU AI Act requirements, with a detection API in private preview.