LIVE
+++ Amsterdam's Wonderful lands $550M at $5B valuation +++ Pentagon Adds ChatGPT and Grok to GenAI.mil +++ US government backs OpenAI's fair-use case against the NYT +++ Runway's Solaris generates app interfaces in real time +++ OpenAI Delays Astra Model After Hugging Face Hack +++ OpenAI hit with 30 new suits over Tumbler Ridge attack ++++++ Amsterdam's Wonderful lands $550M at $5B valuation +++ Pentagon Adds ChatGPT and Grok to GenAI.mil +++ US government backs OpenAI's fair-use case against the NYT +++ Runway's Solaris generates app interfaces in real time +++ OpenAI Delays Astra Model After Hugging Face Hack +++ OpenAI hit with 30 new suits over Tumbler Ridge attack +++
All news ›
AI IN LIFE AI IN LIFENEWS
DAILY
DE EN
BUSINESS

Anthropic Cuts Fable 5.1 Costs by Up to 45 Percent

Anthropic says cached requests run about 25 percent cheaper, with savings reaching 45 percent on complex agent workloads. Some safety filters loosen.

Anthropic Cuts Fable 5.1 Costs by Up to 45 Percent

Symbolic image: a technician slides an accelerator tray into an open server rack in a data center as status lights blink across the chassis.

Anthropic has released Fable 5.1, which the company says costs about 25 percent less than its predecessor when cached data is reused and up to 45 percent less on complex agentic tasks.

At a glance

  • Fable 5.1 is available on all platforms as of the September 2, 2026 report.
  • Reused cached data makes requests roughly 25 percent cheaper than the prior version.
  • Anthropic cites savings of up to 45 percent on complex agentic tasks.
  • Mythos 5.1 stays limited to Project Glasswing firms Amazon, Google and Microsoft.
  • Enterprise Frontier Safeguards are slated for fall 2026, keeping data on customer clouds.

Anthropic has shipped Fable 5.1, a new version of its flagship model, and the headline change is price. Requests that reuse cached data cost about 25 percent less than on the previous version, according to the company. For complex agentic tasks, Anthropic puts the saving at up to 45 percent. The model is available on all platforms, t3n reports.

The two numbers are not the same claim

The 25 percent figure applies to calls that reuse cached input. The 45 percent figure describes long, multi-step agent runs, where the effect compounds across many calls. Teams whose traffic looks like neither case should expect something in between, or nothing at all.

The available reporting does not give absolute per-million-token prices. That means the numbers can only be read as a ratio against the earlier model, and the saving has not been measured independently so far.

Guardrails: sharper, and looser in places

Anthropic says it retuned the safeguards in Fable 5.1 so they trigger more precisely, and describes the model as less restrictive than its predecessor on basic biology questions. The model can now also be used to identify software vulnerabilities. Certain cybersecurity tasks are still routed to Opus models rather than handled directly.

The loosening is therefore selective, not wholesale. On Mythos, the biological restrictions are unchanged from the previous version.

Mythos 5.1 stays behind a closed door

Mythos 5.1 remains available only to the companies in Anthropic's Project Glasswing program: Amazon, Google and Microsoft. Anthropic classifies it as low-risk with respect to automated AI development. For everyone outside that group, Fable 5.1 is the model that matters this week.

Benchmarks and the rest of the announcement

Anthropic says both models set new bests across several benchmarks, including Terminal-Bench 4.0 for command-line coding work and Humanity's Last Exam for general reasoning. Those scores come from the vendor and have not been independently reproduced. The company also published three scientific results, among them GPU optimizations and a high-resolution map of Venus.

Separately, Anthropic announced Enterprise Frontier Safeguards, which would let business customers keep their data on their own cloud servers instead of Anthropic's. Rollout is planned for fall 2026.

What it changes for buyers

For teams running agents over long chains of calls, the pricing shift is the practical takeaway. Whether the 45 percent figure shows up on your invoice depends on how much of your input is genuinely reused between calls. Comparing one week of your own billing before and after the switch is a firmer answer than the vendor's range.

◈ AI-GENERATED REPORT · SOURCES LINKED

FAQ

How much cheaper is Fable 5.1?

Anthropic cites roughly 25 percent lower cost for requests using cached data and up to 45 percent for complex agentic tasks. The available reporting does not state absolute prices per million tokens.

Is Fable 5.1 available now?

Yes. As of the September 2, 2026 report, Fable 5.1 is available on all platforms. Mythos 5.1 remains restricted to the Project Glasswing companies Amazon, Google and Microsoft.

Which safety filters did Anthropic relax?

Anthropic says its safeguards now trigger more precisely and are less restrictive on basic biology questions. Fable 5.1 can also identify software vulnerabilities, while certain cybersecurity tasks still go to Opus models.