SWE 2 pushes programming models toward the cost wall, cutting spending by 64 % while chasing advanced performance

CogniCtion launched its latest coding model, SWE 2, pushing it to the so-called Pareto frontier between efficiency and cost, achieving a success rate of 50.0 % on the FrontierCode 1.1 Main test, coming within one percentage point of the Fable 5.1 model, which scored 50.9 % but at a cost 64 % lower. Subsequent training of the model was performed on the base model Kemi K3, which has 2.8 trillion parameters, marking the first expansion of reinforcement-learning techniques to multi-trillion-parameter models within the company and surpassing the earlier SWE 1.7 and Grok 4.6 models on both accuracy and price.

The core innovation in the training architecture came from a reinforcement-learning algorithm that trains all levels of logical effort in a single training pass.The company employed a reward function that imposes a linear cost penalty calculated precisely according to the Pareto curve slope for each effort level, and that accounts for the monetary inference cost in dollars together with the execution time of the code path, which prevents the model from cutting expenses by abandoning complex problem solving. The company supported this approach with a baseline weighted by path length to reduce gradient variance and stabilize training without additional compute cost, while tripling the reinforcement-learning environments, using quantization kernels from the NVFP4 and FP8 families, and a draft model to increase decode throughput and lower memory consumption despite the large size of the base model relative to previous generations.

This engineering shift translates directly into model behavior in production environments. Test data showed that the model eliminated excessive exploration of code repositories and reading of unnecessary files before execution. In the medium-effort mode, the model achieved higher scores than its predecessor while reducing the number of steps by 58 % and lowering average cost by 81 %, recording the first real code modification after only 18 steps compared with an average of 48 steps for the previous model. The model also demonstrated discipline in hypothesis validation by running code and building comprehensive tests instead of relying on theoretical responses, and showed flexibility in reconstructing data from alternate sources such as conversation logs when API interfaces failed. It was made publicly available in the desktop DevEn application and command-line interface, paving the way for a browser-based release and the Fusion platform.

This development materially changes the calculations of technology leaders and development teams in the Gulf, Egypt and the Levant, as tech firms, banks and digital-innovation centers move to integrate coding agents for automating development and maintenance tasks. The most prominent operational obstacle for these teams locally has been the rapid depletion of cloud inference budgets, caused by agents looping through hundreds of exploratory steps before writing a single line of code. Shifting to models that regulate cost penalties during training and reach solutions in fewer than half the usual steps reduces the operational bill for development environments and enables emerging and midsize regional companies to rely on independent agents without fearing inflated interface costs.

This shift also requires a reset of coding-agent evaluation standards in regional tech organizations.The deciding metric is no longer limited to raw success rates on generic benchmarks; it is now anchored in economic efficiency combined with effort and cost per completed task. With multiple logical effort levels provided within the model itself, from a medium tier for everyday tasks to advanced tiers for complex architectural work, regional engineering teams need to train their engineers in delegation governance and in selecting the appropriate effort level for each programming task, thereby avoiding waste of compute resources and ensuring the safety of generated code and the reliability of its review before deployment in production environments.