Introducing Fable 5.1 in Devin Desktop, Devin CLI, and Devin Cloud.
The cost of Fable intelligence just dropped by 54%. In fact, Fable 5.1 now costs less than Opus 5 on end-to-end tasks.
The savings come from new cache pricing, and we updated Devin to take advantage of this. Devin is now smarter and cheaper across the board, building on top of these Fable improvements. Overall, in Devin users should see 10-25% savings on real work from this update.
Cheaper Frontier Intelligence
FrontierCode 1.1 Extended · cost per task
Devin Fusion: Affordable Frontier Intelligence
Devin’s Fusion Harness combines frontier models for planning & review with cost-effective models for execution to deliver frontier intelligence at lower cost. Read more about the architecture here.
With Fable 5.1, we find it becomes an even more powerful default option. We highly recommend Devin Fusion to anyone who wants frontier performance without burning through their wallet:
FrontierCode 1.1 ExtendedScore vs CostThinking level: medium
Why Token-Based Pricing can be Misleading
When you look at the sticker price of these models, you may walk away with the impression that Fable is far more expensive. But when you look at the cost of these models on actual end-to-end work, you see a very different story:
A higher token price does not mean a higher cost per task
List price per million output tokens next to the measured cost of a FrontierCode 1.1 Extended task, thinking level medium
This is why, at Cognition, we think it’s misleading to frame costs in terms of token pricing. We prefer to measure and talk about costs in terms of cost per completed task.
Why is there such a discrepancy? There are two major factors here:
- Token efficiency. Smarter models are sometimes more token-efficient by making fewer and smarter tool calls. We found that on FrontierCode, Fable 5.1 takes 33% fewer tokens to complete the same tasks relative to Opus 5.
- Cached token pricing. Prompt-caching reads are now $0.25/M tokens in Fable 5.1, which is 75% below the standard cache-read rate and cheaper than Opus 5’s cached tokens. Harnesses that lean heavily on cached tokens, such as Devin, therefore see a cost advantage relative to Opus.
The Math of Cached Input Token Pricing
A typical FrontierCode task on Fable 5.1 reads about 3 million cached tokens, writes about 21 thousand output tokens, and reads only about 70 thousand tokens of uncached input. Opus 5 reads about 4.5 million, writes 26 thousand, and sends 85 thousand. Almost everything an agent reads is context it has already seen: the repository, the task, its own earlier turns. Over 95% of the tokens on both models are cache reads.
So the headline input and output prices only account for a fraction of the cost.
Where the cost of a task actually goes
At the standard cache-read rate, most of what Fable 5.1 spends on a task goes to re-reading context it has already seen. Cut that rate by three quarters and the same task drops from about $5.00 to $2.68. Opus 5 reads its cache at twice Fable’s new rate, so despite the cheaper sticker it lands above Fable.
ZDR for Fable
All of these cost and intelligence improvements are only available to customers with Fable access. Previously, Fable models were not available under ZDR (zero data retention) agreements.
Eligible customers can now use Claude Fable 5 and Fable 5.1 with zero data retention under a time-bound exemption while Anthropic rolls out Enterprise Frontier Safeguards. In parallel we’re giving Anthropic feedback on Enterprise Frontier Safeguards, which will give eligible Claude customers the option to deploy Anthropic’s most capable models while storing their data in cloud infrastructure controlled by Cognition or by the customer. We’ll share more on what that means for customers when it launches.
Fable 5.1 on the FrontierCode Leaderboard
We publish a live model leaderboard for FrontierCode, a benchmark that measures code mergeability, at cognition.com/frontiercode.
On this benchmark, Fable 5.1 does not strictly score higher as thinking increases. The reason is our mergeability criteria. One of the criteria is scope. It works against any diff touching files outside of what the task requires, even if the diff is correct or useful. This could be an extra docstring, updated docs, a new CI job where an existing one could have been reused, etc. This is why Fable 5.1’s score peaks at medium and falls below Fable 5 at higher reasoning efforts. Its pass rate continues to rise with effort, but at higher reasoning efforts it more often adds unrequested changes outside of the task.