Artificial Analysis benchmarked Cognition's Devin Fusion, which pairs a frontier model with the cheaper SWE-2 model as a sidekick, finding the GPT-6 Astra pairing costs 43 percent less and runs 31 per…
OpenAI opened its Agents API in public beta, giving developers managed sandboxes, long-running sessions and subagent orchestration that power Codex internally through a single API call. The release wa…
Anthropic is running internet-connected cybersecurity evaluations that let Claude models reach real systems instead of staying sandboxed, and has brought in independent auditor METR for an eight-week …
AI output is accelerating faster than human review can adapt, shifting the bottleneck from generation to verification, according to engineering teams and researchers. OpenAI claims an unreleased model…
OpenAI's Agents SDK v0.22.1 adds built-in guardrails that block MCP tool calls before they execute, plus configurable sandbox isolation, enabling agent builders to enforce safety policies without cust…
OpenAI's chief scientist said reasoning models have reached capability jumps that could sustain recursive self-improvement, urging stronger monitoring and coordination before the next jump. The remark…
OpenAI shipped GPT-6 Astra with finer prompt comprehension and marked gains in 3D-model and complex-scene rendering, giving developers calling its image API notably higher fidelity on structured 3D ou…
OpenAI launched GPT-6 Astra, which scored 99.9% on ARC-AGI-3 and became the first model to reach Critical cybersecurity capability under OpenAI's own safety framework, positioning it as the new defaul…
SemiAnalysis reports that published CVE descriptions can enable AI agents to construct working exploits against software beneath an agent sandbox, making unpatched host software part of the containmen…
Recraft shipped an MCP server that exposes its image and vector generation as tools any MCP client can call, including Claude and Cursor, making asset creation a scripted step in agent workflows. The …
Recraft's machine learning team attributes broken text rendering in AI-generated images to training captions that describe scenes without transcribing the words inside them, and says fixing this capti…
Relace reports that search consumed more than half the tokens across 1,200 coding agent traces, and a dedicated retrieval model achieved 0.71 Recall@k while a compaction pass cut one real trace by 55 …
Enterprise teams at the AI Engineer conference reported that the correction loop, not the agent loop, dominates system effort: Maersk's procedure corpus is about 20 times the size of its runtime and a…
Across talks and results from AWS, Figma, Anthropic, and Uber, teams report that verification, not model quality, is the primary constraint on agent adoption. AWS runs Lean proofs against Cedar's auth…
Four papers published today challenge the reliability of LLM-as-judge, showing that judge annotations can mark working code as failures, leak across rubric dimensions, and misalign with human creativi…
Research released today shows that agentic scaffolding, not the base model, drives most performance gains and benchmark scores are bands influenced by option order and prompt wording, while scaffoldin…
A state-of-the-art review of AI-assisted coding reports field experiments showing 26% more tasks per week, randomized trials showing a 19% slowdown, and team telemetry showing code review time up 441%…
Four launches today focus on verifying agent output rather than generating it: Conveyor requires every code diff to trace back to an approved intent, VinvAI evaluates fixes against hidden tests, Skill…
A wave of agent infrastructure releases on the same day signals that reasoning models are now a commodity, with the focus shifting to the layer underneath agents: observability, security, and context …
A new list ranks the most influential thought leaders in agentic engineering, AI, and vibecoding, with Anthropic's Claude Code, Microsoft's MarkItDown, and OpenAI's Codex CLI among the top tools. The …