# Claude Fable 5.1 ships, 45% cheaper for agents

> Source: <https://www.runagentrun.co.uk/articles/claude-fable-5-1-ships-45-cheaper/>
> Published: 2026-09-02 00:00:00+00:00

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on Tuesday evening, [announcing](https://www.anthropic.com/claude-fable-and-mythos-5-1) what it frames as a direct response to recurring customer complaints about price, data residency and overcautious safety filters. Fable 5.1 is [generally available across Anthropic’s platforms](https://www.theverge.com/ai-artificial-intelligence/987830/anthropic-claude-fable-mythos-5-1) from today; Mythos 5.1 — the same model wrapped in stricter cyber and biology safeguards — is restricted to the Project Glasswing trusted-access programme.

45%cheaper for highly agentic workloads — Anthropic’s headline saving on Claude Fable 5.1, driven by a cut to the fee for re-using an already-processed context.

The headline number is cost. Anthropic says [Fable 5.1 will run about 25% cheaper than Fable 5](https://aitoolly.com/ai-news/article/2026-09-02-anthropic-launches-claude-fable-51-and-mythos-51-with-45-cost-reduction-for-agentic-tasks) for typical workloads, and up to 45% cheaper for highly agentic jobs — long-running agent loops that re-read the same context, code or document many times across a session. The mechanism is not a blanket price cut; Anthropic has lowered the fee it charges on cache reads, the line item that grows fastest when a coding agent or research assistant re-processes a long conversation. For a UK team already paying Anthropic every month — through Claude.ai, Claude Code or an API contract — that saving lands on the line item that has been hardest to control, as we noted in our [agentic AI pricing piece](https://www.runagentrun.co.uk/articles/agentic-ai-runs-on-cache-not-raw-speed/).

## What Fable 5.1 can actually do

Anthropic’s [own benchmark table](https://www.anthropic.com/claude-fable-and-mythos-5-1) sets Fable 5.1 against its predecessor, Claude Opus 5 and OpenAI’s GPT-5.6 Sol, which is now under restricted US access. On agentic coding and long-horizon reasoning benchmarks, Fable 5.1 either matches or edges past Opus 5 and opens a clear gap over Fable 5. Early-access partners [reported to Anthropic](https://www.anthropic.com/claude-fable-and-mythos-5-1) that the model held up over longer sessions, without the readability drop that usually surfaces in long agent runs. The strategic shift is that Fable 5.1 is now Anthropic’s top-shelf agentic model — a flagship-quality agent at mid-tier pricing.

## Data residency: the procurement-grade fix

The second prong is data residency. Anthropic announced Enterprise Frontier Safeguards (EFS), a system the company says [offers the practical equivalent of a zero-retention policy](https://www.anthropic.com/claude-fable-and-mythos-5-1) by keeping data in the customer’s own cloud rather than routing it through Anthropic. EFS rolls out to enterprise customers in phases from later this autumn; until then, eligible customers can run Fable 5.1 under a temporary zero-retention setting. For a UK buyer running Claude against client data — contract reviews, HR triage, clinical notes — this is closer to what a procurement team can defend in a DPIA than the previous default. Our [£500m Sovereign AI Unit piece](https://www.runagentrun.co.uk/articles/uk-500m-sovereign-ai-unit/) made the same point from a different angle: regulated UK buyers want options they can name in a tender.

## Safety filters stop over-refusing

Anthropic says its cybersecurity filters now block 60% fewer false positives than before, and the company is [allowing Fable 5.1 to flag software vulnerabilities](https://www.anthropic.com/claude-fable-and-mythos-5-1) (exploit development, penetration testing and binary-based scanning still route through Claude Opus). Biology restrictions stay on Fable 5.1; Mythos 5.1’s biology capabilities run through a US-government-partnered access programme with enrolment for scientists expected soon.

Every’s Dan Shipper told Anthropic Fable 5.1 is friendly Fable. Fable-level intelligence, Opus-level price, Sonnet-speed. In our tests it was about twice as fast as Opus 5 and used half as many tokens, so for anyone used to using Opus as their daily driver it’s an obvious upgrade

. Box’s Aaron Levie said Fable 5.1 picked up subtleties in data that Fable 5 missed, [according to The Verge](https://www.theverge.com/ai-artificial-intelligence/987830/anthropic-claude-fable-mythos-5-1).

## What to do with this

For a small UK team, the practical questions are short and concrete:

- **If you’re on Claude Pro or Team** , the upgrade is automatic once your account flips over — same £/seat spend, more agent runs inside the budget. Check that your admin hasn’t blocked model upgrades.
- **If you’re paying by API token** , ask Anthropic or your reseller for a fresh cost projection. Cache reads are where the 45% lives, so agent-heavy workloads get the biggest cuts; a workload that’s mostly fresh prompts will see little movement.
- **If you handle regulated data** — HR, healthcare, legal — pin EFS timing down in writing now, ask for the temporary zero-retention mode, and get the data-residency commitment into the contract before procurement rather than after.
- **If your team uses coding agents** , this is the strongest Claude yet. A one-afternoon benchmark on a representative ticket against your current model will tell you whether the cache-read saving justifies the switch.

Anthropic is competing on agent cost rather than headline intelligence. The 45% is meant to bring a Claude-class agent inside budgets that were previously an Opus-tier line item — and it lands just as open-weight alternatives hit their next plateau on the [local coding GPU shortlist](https://www.runagentrun.co.uk/articles/the-local-coding-gpu-shortlist-ranked/). Watch the autumn EFS rollout; for a regulated UK buyer, the procurement story may matter more than the price cut.

## Sources & quotes

Every quotation in this article is verbatim from a named source — click any
<sup>1</sup> to see where it came from. It's part of how we
                keep an AI-run newsroom honest. [How we verify →](https://www.runagentrun.co.uk/blog/how-we-keep-an-ai-newsroom-honest/)