SDSignal Desk

Anthropic launches Claude Fable 5.1 and says it’s up to 45 percent cheaper for agentic work

Sep 1, 2026, 3:01 PM · The Verge

Image: The Verge

Fable 5.1 is Anthropic answering the three complaints that actually show up on invoices: price, refusals, and where the data lives. The 45 percent figure is for cached agentic work, not a list-price holiday.

Why it matters

Anthropic launched Claude Fable 5.1 across its platforms and Mythos 5.1 for Project Glasswing participants. The company says Fable 5.1 is stronger than Fable 5, typically about 25 percent cheaper, and up to 45 percent cheaper on complex agentic tasks because of lower prices on cached data already processed and stored.

That is a customer-ticket release. The Verge notes the models are aimed at complaints about price, data retention, and overzealous safeguards. Enterprise Frontier Safeguards, due later this fall, store data on the customer's cloud rather than Anthropic's and are sold as 'complete privacy.'

From the desk

Price is the headline because agent workloads burn tokens. A cache discount that can reach 45 percent is how you keep a coding agent online without a CFO intervention. It is not the same as a 45 percent cut on every call. Anyone who quotes the high number without 'cached, agentic' is doing marketing.

The early-access quotes are the usual friendly chorus. Dan Shipper called it the strongest coding model Every has used, now fast and 'speaks like a normal person.' Aaron Levie said a Box agent caught subtleties Fable 5 missed on the same test. A user posted that Mythos 5.1 on low reasoning matched the prior model on Max. Treat those as directional. They are not a bakeoff.

Safeguards are the cultural tell. Fable 5.1 has 'more precise' filters that Anthropic says are less likely to block basic biology questions; Mythos 5.1 keeps the old biology restrictions. Anthropic is also letting Fable 5.1 identify software vulnerabilities, while still sending penetration testing, exploit generation, and binary scanning to Opus. That is a product manager drawing a line the safety team can live with and the security buyer can still use.

Signal Desk's read: this is Anthropic competing on cost-to-complete-a-task and on enterprise data gravity, not on a new capability class. If Frontier Safeguards actually keep prompts off Anthropic disks, that is the feature regulated buyers have been waiting for. If it is a contractual label on the same logs, the 45 percent will be the only thing that ships.

Context

Project Glasswing is Anthropic's limited access track for Mythos 5.1. Opus remains the sink for the sharper cyber tasks. The release sits in a week when OpenAI is absorbing lawsuit and delay news, which makes a cheaper, less naggy Claude an easy switch pitch.

Who feels it

Developers
Re-benchmark coding agents on cached-token bills, not list price. Check whether biology and vuln-id refusals actually dropped in your prompts.
Enterprises
The fall privacy controls are the procurement item. Ask where logs live, in which cloud account, and who can still see them.
Anthropic
Mythos limited to Glasswing keeps the frontier SKU scarce. Fable is the volume play.

What to watch

  1. Independent cost-per-resolved-ticket numbers on Fable 5.1 versus Fable 5, with cache hit rate disclosed.
  2. The Enterprise Frontier Safeguards rollout date and the actual data-flow diagram.
  3. Whether Opus remains required for the cyber tasks Anthropic still will not give Fable.

Read the original

Continue at the source.

The Verge

Companies: Anthropic