Models
Artificial Analysis puts it just ahead of Claude Opus 5.5. Only some cyber defenders get access for now, not developers, at $2 in and $10 out per 1M tokens.
Why it matters: Carbon is pitched for multi-day agent runs and finding and patching bugs, but Google is gating it behind a trusted program first, so you can plan around the price today and not the access.
Deep Dive · FridayModels
SpaceXSI says it's the voice model behind Starlink's support lines and Tesla's in-car assistant, and SpaceXSI sells Grok Voice 3.0 to anyone at $0.08 a minute.
Stories that change something.
Models
Why it matters: Carbon is pitched for multi-day agent runs and finding and patching bugs, but Google is gating it behind a trusted program first, so you can plan around the price today and not the access.
Tools
Why it matters: Dropping the separate seat changes the budget math: Copilot spend now tracks usage, so teams need credit limits and usage review before the agents fan out across repos.
Tools
Why it matters: If your product needs to show up inside Google's assistant, an MCP server is now the entry ticket, which makes a well scoped MCP layer a distribution channel and not just an integration.
Models
Why it matters: Teams stitching ASR, diarization and translation from three models can try one open checkpoint they can host and fine-tune on their own vocabulary, though it does not cover live audio.
Tools
Why it matters: Agents that guess URLs and curl them waste calls on 404s. A single logged search endpoint makes grounding easier to audit, but check each provider's rules on storing results first.
Tools
Why it matters: Browser assistants keep pulling private data into form fields, so anyone building checkout or sign-up flows should expect values sourced from email and validate them like any user input.
Tools
Why it matters: Agent spend is starting to show up on engineering budgets, and team level reports plus pooled caps give leads a way to justify it or rein it in before finance asks.
Hype we checked so you don't have to. How we judge
Agent config files make repos 8x more productive? GitHub's own stat is only a correlation
What we checked: The GitHub Blog Octoverse post compares average commits between new repos with and without an agent config file and calls it a sign of a different pace. It shows no control for team size, repo type or automated commits, so it cannot show the file causes anything.No one needs a mental model because Claude explains the code? One answer misread its repo
What we checked: InfoQ quotes the Anthropic post and documents a Claude-generated maintainer comment saying nothing is closed solely by inactivity, while the repo's issue-lifecycle.ts closes issues after 14 plus 14 days and more than 6,000 issues with repro steps were closed as inactive.AI-driven hackers hit South Korea's brokerages again? Investigators have not said that
What we checked: TechCrunch reports researchers tied the September wave to a suspected China-based hacker using the Artex AI penetration tool, but says officials are still investigating the root cause of the latest breaches and Mirae and Bithumb have not confirmed data theft.Nemotron 3.5 Super VL cuts inference cost 40x? That is a best case from partner fine-tunes
What we checked: The NVIDIA Developer blog attributes the 40x cost, 10x latency and 40% accuracy figures to post-trained models at Scale AI, Ramp and AI21, without saying what they were compared against, and the 5x throughput claim is against unnamed open models of similar size.Signal only. One email, weekday mornings.