Fetching from the wire…
01
02
03
04
05
06
07
08
Research2026-08-20 · source-backed
2,910 programmatically verified tasks built from an ontology of 97 canonical UI components. Holding the harness fixed and changing only observation and action space, GPT-5 mini scores 83.1% with accessibility-tree observations and 48.9% with coordinate-only pixel control. Across seven models, even the fastest config takes 3.7x the matched human reference trajectory time. (arXiv 2608.18307) Anyone benchmarking computer-use agents without controlling for observation modality is measuring the wrong thing.
Each link below shares sources, entities, or timing with this story.
arXiv 2608.05108 skips the RL-trained attacker models that dominate red teaming and generalize poorly, instead accumulating a strategy library across a sequence of (dataset, target) pairs that transfers to unseen targets with no retraining. AgentDojo: 86.7% ASR against Gemini-...
arXiv 2608.24358 switched models mid-run on long coding tasks using cheap/expensive pairs from the Claude and GPT families. Full-trajectory escalation from weak to strong recovers under half the gap while costing a substantial premium, which the authors call the handoff tax. D...
GitHub expanded Copilot's Rubber Duck mode with something that caught my attention: cross-family review. Claude now critiques GPT-authored sessions. GPT-5.5 reviews Claude sessions. Two different model families, trained on different data with different failure modes, checking...
Terminal-Bench 2.1 results (entries dated June 17) put Codex CLI on GPT-5.5 first at 83.4%, Claude Code on Fable 5 second at 83.1%, and Claude Code on Opus 4.8 at 78.9%. The asterisk matters more than the ranking: Fable 5 and Mythos 5 have been export-suspended since June 12,...
In 30-day simulations where fifty shipper agents on GPT, Claude, and Gemini procured truckload capacity under real digital-freight rules, every model independently picked the same modal first-choice carrier on day one, drawing up to 76% of requests, with concentration rising s...
Accuracy drops 30–50% well before you hit the documented context limit. Not at the limit. Before it. Cross-model testing across GPT-4.1, the Claude 4 family, Gemini 2.5, and Qwen3 quantified what everyone shipping long-context features has felt and couldn't measure (Glasp). Th...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.