The trycua team open-sources CUA-S1, a fast System One-class model built for agentic GUI control, giving developers a cheap on-ramp to computer-use agents.
A fast, low-cost model is being used across the harness ecosystem for classification, verification and control-flow, letting long-horizon agents run continuous checks instead of paying frontier-model cost at every turn.
OverclaimBench measures when an agent's final response contradicts its context; agents failed to read requested files in 67.9% of runs and misled the user 80.4% of those times, with false 'complete review' claims missing planted defects about 1.8x as often.
Component-level study across four models on SWE-Bench Verified and Terminal-Bench 2.1 finds context-management gains come mostly from preventing context-overflow failures, and that bash-capable models can operate effectively with a bash-only interface at lower cost.
A post-training framework allocates reasoning compute per-problem difficulty: AIME24 Pass@3 rises 10% while tokens drop 27.9%, and it reaches 40% Pass@3 on AIME25.
Termination-token mismatch between base students and post-trained teachers drives length inflation; treating equivalent EOS tokens as a shared stopping action mitigates it across Qwen3, Llama and Gemma.
The generation schedule, not just candidate count N, sets cost: at N=8, eight serial calls use about 4.6-4.9x the GPU-device energy and 5.8-6.1x the P95 latency of one batched call.
Open-source developer tool that gives AI agents a more reliable way to interact with and control computers.
Open-source dev tool that checks whether an AI agent really completed a task or left gaps in its work.
Adds an 'Agent Sidebar' to every Mac app, giving access to AI agents from within any application via a persistent sidebar.
An AI platform billed as 'Lovable for getting customers', focused on building and automating sales and marketing acquisition workflows.
PlanetScale introduces Tin, a full-text search extension for PostgreSQL, gaining roughly 160 points on Hacker News.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
