Summary
- Cognition has added Anthropic's Fable 5.1 to its coding agent Devin, and a change in caching cut per-task costs 54% below Fable 5, making it cheaper than even Claude Opus.
- Devin's own execution system, the Fusion harness, scored nearly the same as Fable 5.1 on the FrontierCode 1.1 Extended benchmark (63.2 vs. 63.6) at $1.43 per task — 47% cheaper than Fable 5.1.
- The announcement lands amid an ongoing round of price cuts from OpenAI and Anthropic as both companies push back against competition from Chinese AI models.
Devin's coding agent gets a new brain
Cognition announced on September 1 (local time) via its X account that it has integrated Anthropic's new model, Fable 5.1, into its coding agent Devin. Thanks to a change in how caching works, the cost of running Fable-tier intelligence dropped 54%, making it cheaper to use than even Claude Opus, the company said. A FrontierCode 1.1 Extended benchmark table released alongside the announcement shows Fable 5 scoring 62.8 on mid-difficulty tasks at $5.84 per task, while Fable 5.1 scored 63.6 at just $2.68.
To unpack that a bit: Fable is the name Anthropic uses for its Claude-family models, and Cognition is the company that runs Devin, an agent service built on top of those models to handle coding work. Devin is a software engineering agent that plans, writes, and executes code on its own once given a task. This announcement is essentially about swapping in a newer version of the model that does the actual decision-making inside Devin — and cutting costs in the process.
What Fable 5.1 is, and why this fits a bigger price war
Anthropic priced Claude Opus 5 at half the cost of its top-tier model Fable 5 back in August. At the time, Fable 5 was priced at $10 per million input tokens and $50 per million output tokens, with Opus 5 set at half that. Now, with Fable 5.1 cutting prices again, the gap between the two models has narrowed further. OpenAI made a similar move around the same period, cutting both input and output prices for GPT-5.6 Luna by 80%. SiliconData's token price index reportedly dropped about 25% since mid-July. Behind these cuts is a growing number of companies switching over to Chinese models like Moonshot and DeepSeek.
Cognition's own harness outpaces the external model
Alongside the model swap, Cognition also revealed details about Devin's execution system, called the "Fusion harness." A harness is the framework that governs how a model calls tools and sequences code execution. On the same FrontierCode 1.1 Extended benchmark, the Fusion harness scored 63.2 at $1.43 per task — nearly matching Fable 5.1's score of 63.6, but at 47% lower cost.
| System | Benchmark score | Cost per task | Notes |
|---|---|---|---|
| Fable 5 | 62.8 | $5.84 | Previous version |
| Fable 5.1 | 63.6 | $2.68 | 54% lower than Fable 5 |
| Devin Fusion harness | 63.2 | $1.43 | 47% lower than Fable 5.1 |
Looking at the three numbers together, it's clear Cognition is squeezing costs from two directions at once: one reduction comes from Anthropic's own caching changes to the model, and the other comes from Cognition optimizing how it runs that model.
Part of a broader cost-cutting push
This announcement fits into a broader trend of OpenAI and Anthropic focusing on driving down the cost of running agents — similar to a case that cut document-processing costs by a factor of 18. Coding agents call a model dozens or even hundreds of times to complete a single task, so even small per-task savings add up fast across total operating costs. Cognition rolling out its own harness to match external model performance at a lower price is, in that sense, another attempt to chip away at the cost of all those repeated calls.
Editor's take
What stands out here isn't the price cut itself, but the fact that Cognition chose to put its own harness performance side by side with Fable 5.1. For agent companies like Cognition that build on top of external models, profit margins are constantly at the mercy of the model provider's pricing decisions. That's pushed companies in this position to invest not in the model itself, but in harness technology — figuring out how to call the model less often and cache smarter. These numbers show that kind of investment can actually bend the cost curve.
Anyone who's put a coding agent to work in a real team will recognize this pattern. First you bolt on the latest model and check the performance. Then you move into the phase of tightening up caching and tool-call sequencing to get the same results for less. For development teams here that have adopted Devin or similar coding agents, the real savings probably won't come from just reading the price sheet every time a model swap is announced — it's worth checking how often your own workflow triggers repeated calls in the first place.
In the coming weeks, it's likely other coding agent companies will follow suit, publishing their own benchmarks to show off the cost efficiency of their execution systems.
Comments