Dongcha Beating AI breaking news:AI Cognition, the programming company, has significantly reduced the costs of its programming Agent Devin. Devin can read code, modify code, run tests, and complete development tasks independently. The usage costs of its Fusion and Normal modes have fallen 30%to 40%, while Ultra has fallen 15%to 20%, and Devin Review has fallen by up to 70%.
Fusion is Devin's multi-model collaboration mode. It assigns planning and checking to more capable models, then delegates specific execution to lower-cost models. Cognition also optimized the execution framework this time by combining multiple commands into a single call, executing independent steps in parallel, and making greater use of context caching to reduce redundant computation and model calls.
Alongside the cost reductions, Cognition's self-reported benchmark results also improved. On its proprietary FrontierCode 1.1 Extended benchmark, Devin Fusion scored 68.8 points, with an average cost per task of 0.60 USD, currently ranking first.
Original article link https://m.theblockbeats.info/flash/369473
Key points
01
Cognition optimized the usage costs of programming Agent Devin.
02
Fusion and Normal mode costs fell 30%to 40%, while Ultra fell 15%to 20%, and Devin Review fell by up to 70%。
03
Fusion is Devin's multi-model collaboration mode.
04
Cognition says Devin Fusion scored on its proprietary FrontierCode 1.1 Extended benchmark 68.8 points, with an average cost per task of 0.60 USD, currently ranking first.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
Devin is a programming Agent that can read code, modify code, run tests, and complete development tasks. This update primarily reduces redundant computation and model calls by combining commands, executing them in parallel, and using context caching.
Why it matters to readers
If these cost-reduction figures can be replicated in actual use, Devin's cost per development task could decrease.
This indicates that the pricing and operating models of programming Agents are still changing rapidly; advertised benchmark results do not mean that all users will achieve the same outcomes.
Risks and unknowns
The article does not provide Cognition's original announcement or complete pricing conditions.
FrontierCode 1.1 Extended is a proprietary benchmark built by Cognition; its test set, scoring methodology, and independent reproducibility are unknown.
The impact of the cost reductions on task success rates, response speed, and actual bills is unclear.
Related Developments
Loading event timeline…
Related concepts
Programming Agent
This term is not in the glossary yet. Browse related concepts in the glossary.
