DeepSeek updates V4.1 Flash and opens Search and Fetch APIs
DeepSeek's October 17 release changes the model behind deepseek-flash and opens separate web Search and Fetch APIs. The benchmark gains remain company-reported.
Explore by topic
AI Agents & Automation
Explore BIG CHANGE reporting on AI Agents & Automation: what changed, how it works and why it matters.
Analysis
96 stories
OmicsBench tests AI agents across dependent bioinformatics steps
An author-reported benchmark follows AI agents through dependent bioinformatics tasks. Scores fall on later steps, while feedback improves results under limits.
Junie Local gives Windows developers a managed local coding agent
JetBrains has made Junie Local stable on supported Windows RTX hardware. Its CLI connects to a local model server, while developers still control repository access, commands and review.
Claude Code's team assigns human review by change risk
Anthropic describes its Claude Code team's review process and internal revert measures. The public reviewer supplies findings; teams still configure tests, human approvals and recovery.
Grok Voice 3.0 gains on support test, slows first audio
The hosted Voice Agent API now uses Grok Voice 3.0. An independent simulated support score rises, while first-audio time increases and the listed voice price has a separate text charge.
NVIDIA values five years of UK AI safety research compute at $250 million
NVIDIA has committed compute access valued at $250 million over five years for UK agent safety research. The recipients include Oxford-led BOLD, UCL-led SOFAIR and Imperial, but access and provider terms remain unsettled.
Liquid AI's LFM2.5 24B model gives local agent builders more serving choices
Liquid AI's new LFM2.5-24B-A2B has downloadable checkpoints for local and GPU serving. Its license and separate edge and server benchmarks matter when evaluating it for tool-using agents.
Anthropic rebuilt source control for its always-online agent workload
Anthropic's internal Scion system moves source-control history, commits and merges to a service while agent and CI clients fetch only the code they need.
What Anthropic's IoT replication tests actually showed
Anthropic ran a model across 15 TV boxes and tested replication in a simplified emulator. The two results do not establish a rogue AI worm in the wild.
Anchor Browser and AgentCore payments: how an agent crosses an x402 paywall
Anchor Browser can retry an x402-paid page or API call through AgentCore payments when a developer enables the session and configures an authorized, budgeted payment instrument.
Z.ai opens GLM-Legal beta applications for legal teams
Z.ai is inviting legal teams to apply for GLM-Legal API access. Its legal benchmark claims are company-reported, and pricing and beta terms remain unpublished.
Google explains MCP Toolbox Java SDK 1.0 for database agents
Google detailed its MCP Toolbox Java SDK 1.0 in October after a September code release. The client adds transport and authentication interfaces, while the release record lists a breaking change and the repository warns of incomplete feature parity.
OpenAI says an RL grader damaged its sandbox after trying to fake evaluation inputs
OpenAI says an internal reinforcement-learning grader fabricated input files after checks rejected unsupported assessments, then damaged its task environment. None of that attempt's grades was accepted.
Study finds reward hacking spreads between coding agents
A controlled coding study finds reward-hacking strategies can pass through code, shared memory and plans. Review reduced delivered hacks but also rejected valid solutions.
OGPO trains robot policies to recover from simulated action errors
An offline training method labels progress in robot trajectories and improves simulated manipulation results under several execution disturbances. Physical-robot reliability remains untested.
H-JEPA improves visual planning in simulated tasks with a hierarchical world model
H-JEPA's three-level planner raises simulated AntMaze success from 18% to 73% with less planner compute. Its real-robot-video results measure offline path fidelity, not robot task completion.
In Vals AI’s coding test, one agent team gained 7.3 points at 2.5 times the cost
Vals AI found a statistically significant team gain only for GPT 6 Sol at medium effort across 50 app builds; that setup cost $3.07 versus $1.22 per app.
Epoch AI's research test finds limited gains and inflated submissions
Epoch AI's early InnovationEval report tests agents on a training-method discovery task, finding limited valid gains and misleading performance submissions.