A US coding model is built on top of a Chinese one

Image: METAL

Summary

Cognition released its SWE-2 coding model on September 10. It is post-trained on Kimi K3, a 2.8-trillion-parameter Chinese model, through a further round of reinforcement learning.
SWE-2 scores 50.0% on FrontierCode 1.1 Main, within a point of Fable 5.1's 50.9%, with 73.0% on DeepSWE 1.1 and 92.8% on Terminal-Bench 2.1. On the harder Terminal-Bench 4 it trails badly at 27.3%.
Because the base model is Chinese, the company put 145 politically sensitive questions to it in three languages and graded the answers. The Simplified Chinese pass rate came in at 95.2%, against 99.8% in English.

Cognition, the American coding-agent company, has released a new model called SWE-2, and the model underneath it came out of China. Published on September 10, SWE-2 is Kimi K3, a 2.8-trillion-parameter model, put through another round of reinforcement learning, and it scores 50.0% on the FrontierCode 1.1 Main coding benchmark. The company says that puts it within a single point of Fable 5.1 at 64% lower cost.

Line the numbers up and the model's position becomes clear. On DeepSWE 1.1 it scores 73.0%, above Fable 5.1 at 67.4% and GPT-5.6 Sol at 72.7%, and on Terminal-Bench 2.1, which measures terminal work, its 92.8% is the highest of the seven models in the company's table. The previous model, SWE-1.7, scored 42.0%, 37.7% and 81.5% on those same three tests.

The weakness the company wrote down itself sits in the same table. On the harder Terminal-Bench 4, SWE-2 manages only 27.3%, far below Fable 5.1 at 55.8% and GPT-6 Astra at 57.9%. At the hardest end of the range, in other words, a lower price does not close the capability gap.

The scores are not the only thing that moved. The Cognition team wrote in its announcement that "SWE-2 medium scores higher than SWE-1.7 while taking 58% fewer turns and costing 81% less on average." Mean steps per run fell from 127 for SWE-1.7 to 53 for SWE-2 medium, and the point at which the model makes its first real edit moved from the 48th step to the 18th.

The company traces the difference to how the model explores. Users had said the earlier model dug through a codebase for too long and spent time even on easy jobs, and the explanation now is that the new one first decides which parts actually matter. It is also more willing to find another route when the obvious one is blocked: in one case an MCP connection it needed was unavailable, so it rebuilt the data from Slack channel history it already had access to.

What is new in the training method is putting cost directly into the reward function. The formula is simple, subtracting a coefficient times cost from whether the run succeeded, and that coefficient is matched to the slope of the base model's cost-performance curve. The Cognition team wrote that this lets it "train all effort levels end-to-end during a single RL run," a different path from Kimi K3, which trains a separate expert for each combination of domain and effort level and then consolidates them into one model.

Cost also came out of the training infrastructure. The company said batching prefill requests raised throughput by 10 to 20%, and that training a draft model online lifted decoding speed. With NVFP4 and FP8 kernels plus quantization-aware training, it wrote, it reduced the mismatch between training and inference below the previous model's level even while using a base model with almost three times the parameters. It also tripled the number of reinforcement-learning environments and built a loop in which earlier SWE-2 checkpoints keep hardening its verifiers.

The effort levels behave differently from one another, which is its own feature of this model. Medium moves into action far sooner and handles simple and mid-difficulty work cheaply, while high and max plan more, sweep more of the codebase, and use more elaborate verification to close out uncertainty. The company said the three levels were raised together in one training run rather than trained separately.

Here is where the real story of this release sits. An American company put the brain of its flagship product on top of a model released by a Chinese lab. Cognition carried over the training infrastructure and recipe of its previous model, SWE-1.7, and this time moved to a base model with almost three times the parameters. The pattern of open-weight models crossing national borders to become the floor of someone else's commercial product has now reached the top tier of coding models.

That is why the company had to measure something besides performance. Cognition put 145 questions on topics treated as politically sensitive in China to the model in English, Simplified Chinese and Traditional Chinese, and checked whether the answers adopted the official Chinese position as their own. SWE-2 passed 98.0% overall, with 99.8% in English and 99.1% in Traditional Chinese, and 95.2% in Simplified Chinese.

The second evaluation looked at whether the code changes depending on who the customer is. Using Western, Pakistani, Chinese, Tibetan and Falun Gong-affiliated customer framings, the team asked for functionality that was vulnerable or open to abuse, and the company said no framing produced a statistically significant rise or fall in vulnerability. The very practice of shipping evaluations like these alongside a model announcement is a procedure the base model's nationality created.

When the price of writing code falls, what changes is not the tool but where people stand. Cutting the number of steps by more than half also means cutting the number of points where a person steps in to check. The company wrote that the new model re-derives its conclusions rather than restating them when challenged, which is also a signal that verification itself is moving from the person to the model.

METAL read the announcement in full, down to how the prices were set. Costs for every model assume list pricing including public discounts, and a footnote records that Fable 5.1 Max scored 50.3% at $12.83 per task on FrontierCode 1.1 Main, below Fable 5.1 Medium's 50.9% at $3.28. METAL reported on DeepSeek shelving its 1.6-trillion-parameter model in favor of a 552B one, and the two announcements point the same way. The contest is no longer over the smartest model but over the one that travels furthest at the same price.