A UT Austin study ran nearly 35,000 coding-agent runs on SWE-bench Verified and Terminal-Bench while varying how context is compressed, when it triggers, and how much is removed. On Terminal-Bench with Qwen, policies using about a third of the tokens can take 20% to 80% longer than keeping full context.
DAIR.AI describes ActiveSaddler, a paper from Microsoft and colleagues that changes which scenarios are used to optimize an agent harness. With the same optimizer, test Pass@1 improves by 4.4 points on GAIA2 and 7.5 points on Terminal-Bench 2.0 versus a fixed scenario order.
Command Code says Claude Sonnet 5.5 is available across GOAT, Pro, Max, and its API. It says the model runs 30% faster, costs 30% less, and beats Opus 5.5 on Terminal-Bench 4.0.
Cline says Sonnet 5.5 beats Opus 5.5 on Terminal-Bench 4.0 while costing half as much. Cline also says it uses fewer tokens than earlier models, cuts cost by up to 30% per task, and generates output 30% faster than its predecessor.
A paper proposes SkillGym, which turns human-written skills into 2,756 training environments and fine-tunes on 8,364 successful trajectories. It reports that Qwen3.5-35B-A3B in Claude Code then gains 19.10 points on Terminal-Bench 2.1 and reaches 51.47% on skill-assisted SkillsBench v1.1.
StepFun has released the Step Code v0.1.0 command-line tool as open source under the MIT License, including one-command static site publishing with StepPage. StepFun reports 80.9% on Terminal-Bench 2.1 and 73.3% on Multi-Frame, its 150-task long-horizon benchmark.
