The headline is not 92.8.
Cognition reports that SWE-2 scores 92.8% on Terminal-Bench 2.1 and 27.3% on Terminal-Bench 4. It also reports 50.0% on FrontierCode 1.1 Main, one point below Fable 5.1 at 64% lower cost. Those are vendor-published measurements, not independent guarantees for your repository.[1]
The more actionable number is behavioral: Cognition says SWE-2 medium used 53 mean steps on its FrontierCode runs, versus 80 at high and 98 at max. Medium reportedly reached its first real edit in a median 18 steps, compared with 48 for SWE-1.7.[1] The release is selling a model, but the engineering idea survives the sales pitch: deeper is a budget with a shape.
“Max” is not a quality setting. It is a purchase order for more search.
The effort knob is spreading.
OpenAI’s current reasoning documentation exposes model-dependent effort values from none through max. It describes lower effort as favoring speed and token use, while higher effort spends more work on complex reasoning; it explicitly recommends evaluating whether the extra latency and cost earn a benefit.[2]
Cognition says SWE-2 trains multiple effort levels in one reinforcement-learning run with a cost penalty tuned to each point on the base model’s Pareto curve. It says SWE-2 is post-trained from Kimi K3.[1] Moonshot’s Kimi K3 repository identifies K3 as a 2.8-trillion-parameter mixture-of-experts model with 104 billion active parameters and publishes weights under its own Kimi K3 License.[3]
Different vendors expose different gears and train them differently. Do not pretend the labels are portable units. “High” is not horsepower. It is a product-specific request to spend more inference on the problem.
Benchmarks are the dyno, not the street.
The Hacker News discussion immediately centered on the gap between SWE-2’s Terminal-Bench 2.1 and Terminal-Bench 4 scores. Commenters dispute whether that gap indicates benchmark saturation, contamination, different difficulty, or weak generalization.[4] The argument is useful precisely because the post’s clean table cannot settle it.
Run your own mixed garage route: small edits, ambiguous bugs, ugly migrations, tool failures, stale tests, and tasks where the correct move is to stop. Record success, review burden, first-edit latency, total cost, and rollback pain. A model that wins a public terminal benchmark can still make your monorepo miserable.
Wire the shift points into policy.
- Start from consequence. Low-risk, reversible work can begin in a cheaper gear. Security, migrations, money, and destructive operations start higher.
- Add ambiguity. Unknown ownership, unclear acceptance criteria, or weak tests justify more exploration—or a human question, not infinite thought.
- Name the shift-up trigger. Escalate on failed verification, contradictory evidence, repeated edits, or contact with a protected boundary.
- Name the shift-down trigger. Once the plan is proven and the remaining work is mechanical, stop paying for philosophical suspension.
- Keep proof independent. More reasoning does not waive tests, review, provenance, or rollback. It only changes how much work the model spends before handing you the claim.