Hey there 👋 I went to sign up for the new number one model this morning. That was the plan: watch Gemini 4 Carbon take the top of the index, go find the button, kick the tires before I wrote a word. There is no button for me. Carbon went to cybersecurity experts first, then to a partner program, and the rest of us wait our turn. That is the whole day in one dead link. The crown moved to a model almost nobody can run, exactly one day after it sat on the cheapest, most-available model we have. So I did what you do when the frontier locks a door. I went looking for what shipped this week that I could put my hands on. Turns out plenty did. - Ron
The Big ThingGemini 4 Carbon is the new #1, and you probably can't run itGoogle DeepMind shipped Gemini 4 Carbon yesterday, and it went straight to the top of the Artificial Analysis Intelligence Index at 59, two weeks after Argon landed at 50. New crown, clean margin. Here is the catch the benchmark charts leave off: Carbon went to cybersecurity experts first, then to trusted partners through Google's Fairwind program, with broad access coming later. It is not generally available. If you are not on that list, the best model in the world this week is one you cannot call. Carbon is tuned for security work, which is the reason for the gate. It scored 88.2 percent on SEC-bench Pro, the highest of any model tested, and Google ran some of those exploit tests with the safety guardrails switched off to see what it could really do. Defense climbed too. Indirect prompt-injection attacks that landed 0.7 percent of the time against Argon now land 0.0 percent. The model is also going through voluntary pre-release testing with the US government before it opens up. On price, read the fine print. The intro rate matches Argon at $2 per million tokens in, $10 out, and runs through the end of 2026. After that the list price is $4 and $20, so today's number is a 50 percent launch discount with a clock on it. Cached input is 95 percent off. One figure cuts against the hype. Carbon's answer accuracy climbed from 48 to 58 percent, which is real progress. In the same jump its hallucination rate rose from 4 to 11 percent. It gets more of the hard questions right and invents more confidently on the ones it gets wrong. On coding it scores 59 on the agent index, ahead of Argon’s 50 but still behind Claude Opus 5.5 at 62. Why it matters: the model topping the index is one most teams cannot touch, so the play is to plan around the access tiers and the widening security-benchmark gap, not to rip out a working stack chasing a score you can't even run. Ship it? WATCH. Nothing to deploy today unless you are a cyber partner. Track it for when the gate opens, and keep an eye on that hallucination line before you trust it anywhere near production. Sources: The Decoder, Artificial Analysis, BigGo Finance.
Tour de Headlines💻 Liquid AI shipped the opposite kind of release the same week the crown got locked away. LFM3 is a family of six open-weight models, 2B class, all on one shared backbone, built to run on hardware you own. Browser agents, voice commands, image recognition, down to a roughly 1GB agent that fits on a phone. The weights are open, and it is free to use if your company is under $10M in revenue. That buys you an agent that never sends data off the machine and costs nothing per task, which is the control the gated frontier won't hand you. One flag before you build: the license carries EU-specific terms, so read those closely before anything touches production over there. Verdict: DEPLOY. VentureBeat 📊 Anthropic shipped the missing piece for agent governance yesterday, because cheap agents only help if you can see the bill. Three features went generally available for Claude Enterprise: Smart Reports, in-app Surveys, and Pooled Group Budgets. Smart Reports reads a team's transcripts across a date range and hands back a usage breakdown with list-price spend attached per session, plus the top uses and where people got stuck. Pooled Group Budgets set a shared monthly cap on top of each person's own limit, and a request stops at whichever one runs dry first. An owner can switch this on today, no build required. It is the governance layer that makes scaling agent spend something you can defend to finance. Verdict: DEPLOY. Unite.AI 🚨 Causal Security's agent, Cipher, took the delight slot with a story that also explains why Google is nervous. It went digging through EPICS Base, the open-source control toolkit running behind roughly 90 scientific and industrial facilities, and surfaced four previously unknown vulnerabilities. Total time: under four hours. The bugs are real, and the facilities that run EPICS are not toys. The signal is bigger than the bugs themselves. Autonomous vulnerability hunting has crossed from demo to useful, so assume the people you least want holding this capability are holding it too. Verdict: WATCH. The Register
Sponsor Your team's calls are full of signal. RapportScore reads it. RapportScore scores how your team communicates on real recorded calls and coaches them on it, so you can see where rapport holds and where it slips. Built for B2B teams who live on the phone. See your team’s score → |
Tool of the Day📈 TimestralForecasting inside a pipeline (demand, inventory, capacity, anomalies, finance) without calling a metered API. Mistral open-sourced four probabilistic time-series models ranging from 5M to 209M parameters, and the 209M, its largest, beats TimesFM-3 on the benchmarks Mistral cites. The whole thing runs on a CPU, so no GPU bill and nothing leaves your box. Not every job needs a 59-on-the-index brain. A lot of real forecasting work wants a small model that does one thing well and runs anywhere. Eval it against whatever forecaster you lean on now before you swap. It is new and open-weight, so confirm the exact license and that benchmark claim on Mistral's own model card first. Verdict: TRY. See the Timestral model card →
Worth a Click- Bloomberg BQuant Assistant went live this week, a conversational coding assistant built into BQuant Desktop and Enterprise to speed up quant data analysis and coding. Narrow by design (quant finance), but well-sourced and worth a look if you are already a Bloomberg shop.
- Google TPUs are coming to Oracle Cloud. Announced at Oracle AI World, Google's TPUs land on OCI starting in 2027 for agentic and long-context workloads. Nothing to deploy now, but a real tell of where the infra alliances are heading.
Two days, two crowns, pointed opposite ways. Yesterday's winner was Terra: cheap, fast, sitting in everyone's API. Today it's Gemini 4 Carbon: number one on the index, built to hunt and patch exploits, and handed to security experts first while the rest of us wait in line. The frontier is splitting into a workhorse you can afford and a specialist you can't get near. And the news you could act on ran the other way all week. Liquid AI put open-weight agents on a laptop for free, Mistral put a forecaster that beats the big names on a plain CPU, and Anthropic handed admins a per-session bill for what their agents burn. Want the reason Carbon is behind a gate? An agent named Cipher found four zero-days in reactor control software in an afternoon. Chase the index if you want. I'd deploy the edge. Go pull the forecaster, the on-device agent, or the spend report. One of those three is runnable before lunch, which is more than this week's number one can say. - Ron |
