Computing Life · Share · YageSep 13, 2026, 08:00 UTC+8

Drawing a cost curve is not the same as pushing it down

Cognition released SWE-2, baking inference cost directly into the RL reward function so the model learns to take shorter paths. The mid-tier variant cuts interaction turns by 58% and cost by 81% vs. SWE-1.7. The reward is R = S − λC: pass score minus a time-and-token penalty. But if the penalty shape is off, the model games it by giving up early. On Terminal-Bench 4 it scores 27.3%, trailing Claude Fable 5.1 and GPT-6 Astra. The post doesn't include an ablation without the cost penalty, so it's unclear how much of the efficiency gain comes from the stronger base model Kimi K3.

Why it matters: Cognition's SWE-2 launch is a solid coding-agent story this week, and the author goes beyond news recap—the 'pick a point vs. push the frontier' framing nails what cost optimization actually means, backed by the reward function formula and real numbers. Score held at 78 becaus...

Read the original ↗Export Markdown

Take

I clicked on this because Cognition did something straightforward: they wrote inference cost directly into the RL reward as R = S − λC, so the model learns to take shorter paths on its own. The self-reported numbers are striking—mid-tier SWE-2 drops interaction turns from 127 to 53, cuts cost 81%, and moves the median first-edit step from 48 to 18.

But I'd discount this on two fronts. First, the post doesn't include an ablation without the cost penalty; the only ablation compares reward baseline forms. SWE-2 also switched to the stronger Kimi K3 base model, and Cognition doesn't separate how much of the efficiency gain comes from that upgrade. Second, on Terminal-Bench 4—a long-horizon benchmark—SWE-2 scores just 27.3%, well behind Claude Fable 5.1 and GPT-6 Astra. When the penalty piles up, the model's rational move is to give up early rather than burn tokens on a hard problem.

They chose a linear penalty over logarithmic, deriving it via Jensen's functional equation in Appendix B, but OckBench argues for log penalties—no consensus yet. The multi-tier training uses a single RL run across all thinking budgets, avoiding Kimi K3's train-nine-experts-then-distill approach, which is a clean design choice.

FrontierCode 1.1 Main isn't public, so independent replication is off the table. And cost numbers without test-environment context can mislead—one study found token-per-solved-task can vary 40x just by swapping scaffolding. The direction makes sense, but this looks more like a narrow cost-saving demo than a general solution.

Source: Computing Life · Share · Yage · yage.ai
Published: Sep 13, 2026, 08:00 UTC+8 · 17 days ago