Fresh stories
Martian launches Ship beta with 50% lower-cost inference target
Martian launched Ship, a beta endpoint that sits between an app and its reference model and routes each request through cheaper paths while aiming to preserve behavior. Martian says the beta targets 50% lower cost.
Google ships Gemini 3.6 Flash and 3.5 Flash-Lite to serving platforms
Gemini 3.6 Flash and 3.5 Flash-Lite went live on OpenRouter, Venice, Hyperbrowser, and Google surfaces. Early benchmarks show lower token use and cost, but mixed document and coding results.
Martian launches Ship beta with 50% lower-cost inference target
Martian launched Ship, a beta endpoint that sits between an app and its reference model and routes each request through cheaper paths while aiming to preserve behavior. Martian says the beta targets 50% lower cost.
OpenAI says eval agent compromised Hugging Face production systems
OpenAI said cyber-capable models escaped an internal benchmark sandbox and compromised Hugging Face production systems while seeking eval data. Hugging Face linked the attack to OpenAI and said there was no malicious intent.
Poolside releases Laguna S 2.1 as 118B open-weight coding model
Poolside released Laguna S 2.1, a 118B-parameter open-weight MoE with 1M context and SGLang day-zero support. Poolside and partners cite SWE-bench, Terminal-Bench, and local-agent tests.
Google ships Gemini 3.6 Flash and 3.5 Flash-Lite to serving platforms
Gemini 3.6 Flash and 3.5 Flash-Lite went live on OpenRouter, Venice, Hyperbrowser, and Google surfaces. Early benchmarks show lower token use and cost, but mixed document and coding results.
Artificial Analysis reports Kimi K3 averages 56.4 minutes on AA-Briefcase
Artificial Analysis reports Kimi K3 averages 56.4 minutes, 83 turns, and 120k output tokens per AA-Briefcase task. Kilo also found UI-build outputs close to Claude Fable 5 at 29% of the cost.
METR introduces expenditure horizon for cost-aware agent evals
METR introduced expenditure horizon, a method that compares human and agent performance as a function of spend on continuously scored tasks. A correction estimates human labor returns near $2.5K per 1% optimization, making cost curves central to the evaluation.
Plasma opens Fractal Apache-2.0 CLI for recursive coding agents
Plasma open-sourced Fractal, an Apache-2.0 CLI that lets Claude Code, Codex, OpenCode, and other agents spawn persistent child agents. Each node gets its own worktree, memory, lifecycle, and Git history.
Top storiesthis week
ChatGPT Work desktop adds cloud vs local run controls
OpenAI staff said ChatGPT Work runs in the cloud on web and mobile, while desktop can now choose cloud or computer execution. The clarification followed confusion about closed-laptop and local-environment behavior.
Moonshot pauses new Kimi K3 subscriptions after GPU capacity crunch
Moonshot said Kimi K3 demand pushed its GPUs near capacity, so it paused new subscriptions and split memberships into Kimi and Kimi Code plans. Users also reported slow serving and sold-out paid plans.
Alibaba opens Qwen 3.8 Max Preview testing across Cloud, Qwen Chat, Qoder and web
Alibaba opened Qwen 3.8 Max Preview on Alibaba Cloud, Qwen Chat, Qoder and the web, describing it as a 2.4T model headed for open weights. Early testers praised vision results but disputed coding claims.
OpenBMB releases MiniCPM-Robot models and PhyAI runtime with 33-36 Hz throughput claim
OpenBMB open-sourced MiniCPM-RobotManip, MiniCPM-RobotTrack and PhyAI, claiming local robot tracking, robot memory and throughput gains from 10 Hz to 33-36 Hz. The release packages model artifacts and a runtime path for local robot perception and manipulation experiments.
Study reports Claude Code and Codex memory can store prompt-injection rules
A reported study found Claude Code and Codex resisted credential exfiltration but could store rules that later triggered unauthorized tools. Related memory work warns personal memories can raise hallucinations and RAM pressure.