Google's Gemini 4 Carbon tops the leaderboards, and its first users are only cyber defenders

Google announces Gemini 4 Carbon, #1 on the indexes but limited to cyber defenders: frontier access now comes in stages

Carbon scores 59 on version 5.0 of the Artificial Analysis Intelligence Index, first place, ahead of Claude Opus 5.5 and GPT-6 Astra. It entered LMArena at #1 with a 1540 rating from about 4,400 votes, and jumped 89 Elo to fourth on WebDev Arena. Its biggest leads are in security: 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for turning them into working exploits, and 71.0% on CWE-bench for patching them. Google ran all three without guardrails. It also says its own teams watched Carbon work for days on single engineering tasks, and that the Chrome team found its patches more than ten times better than those of "another lab's large frontier model."

The rollout is the news. Carbon goes to a subset of cyber defenders this week, then to trusted partners in Google's Fairwind Program, and only after that to developers and consumers. Google says it is taking part in the US government's voluntary pre-release testing. The price matches Argon's introductory $2 input and $10 output per million tokens, with cached input 95% off. Earlier this year, Washington temporarily limited the rollout of cyber-capable models from Anthropic and OpenAI, and last week's PHP2Shell scramble showed why. Cyber capability has become the thing that decides who gets a frontier model first. My bet is that every frontier launch from here comes with a defenders-first window, and that builders should plan around a gap of weeks or months between the announcement and an API key.

The safety claims are aggressive: successful cheating on Terminal-Bench 4.0 cut to 0.0% (with attempts at 0.4%), high-severity misalignment alerts in internal coding down almost 80%, and a 0.0% attack success rate on Gray Swan's indirect prompt-injection benchmark. Google measured all of these itself. A zero on an injection benchmark says as much about the benchmark's limits as about the model. Ars Technica noted that Carbon beats Argon by slim margins on every test. It could be a post-training upgrade of Argon rather than a new model, and the whole frontier sits within a few points of each other.

The Hacker News thread barely argued about the numbers. It was about access and naming. The top comment came from someone who said they weren't sure they could figure out how to prompt Argon, let alone Carbon, through Google's plans and apps. Others asked whether Carbon replaces Argon or sits beside it.

Given how quickly the field is moving, showing benchmarks without releasing the model is quite meaningless.

— solarpunk_enthu, Hacker News

Google opens Search and Gemini to any MCP server as Search growth slows to 12%: agents become the storefront

Alphabet's numbers explain the urgency. Revenue rose 25% to $127.6 billion, but Search grew 12%, down from 17% in the second quarter and 19% in the first, and the shares fell 8%. Google's answer is to put ads inside AI. Ads in AI Mode are going international for the first time, multi-turn ads are coming in the US, and Google says sponsored listings in AI Mode already get clicked about as often as organic ones. Sundar Pichai said queries grew at their fastest rate in five years, that AI Mode queries run three times longer than traditional searches, and that the Gemini app has reached a billion monthly users.

The same afternoon, Google opened a preview of its Connected Apps Platform. A developer hosts an MCP server in the cloud, or uses Android MCP through the AppFunctions APIs, registers it in a Google Connected Console, and chooses whether it appears in Search, the Gemini app or both. Once it's live, the developer gets traffic, latency and error dashboards. More than 50 apps are already live, including Adobe, Experian, Localiza and Peloton, and Calm is coming. The preview is limited to US developers on a waitlist. TechCrunch compared it to early SEO, and that's the right frame. An MCP server is turning into the way a business gets listed inside Google. Chrome's autofill now pulls flight numbers, order IDs and tracking details out of Gmail through Gemini, opt-in and US-only, which is the same move aimed at the browser: keep the agent inside Google's own surfaces.

The products those apps plug into are getting real use. Wired ran six always-on phone agents through a day of tasks: Meta's Muse, OpenAI's Dots, Instinct, Claude Duet, Gemini Spark and Grok Bot. The reviewer's verdict was that they now mostly work, which they didn't a year ago. Gemini Spark wrote the best morning news pitch, though that was one test of several. A new benchmark, Omni-InteractScope, shows how far there is to go when an assistant has to watch a live audio-video stream and act on its own. None of the 12 omni models tested responded on time to more than 40% of tool-call moments. Reminders the user never asked for peaked at 23% recall. Nine of the 12 completed none of the 26 multi-step guidance sequences. Proactive agents work in a text thread. Watching and acting in real time is still unsolved.

Microsoft books $10B from OpenAI in one quarter and $677B in backlog: the labs' compute bills are the clouds' revenue

Microsoft's tenth straight record quarter came to $92.3 billion, up 19%, with net income of $36.6 billion. For the first time, Microsoft reported Azure in dollars: $33.2 billion, up 48%, beating both analyst estimates and CFO Amy Hood's own 44–45% guidance. That puts Azure just ahead of Google Cloud's $30.7 billion, though Google Cloud is growing about twice as fast. Amazon reports today.

The OpenAI line is the one to remember. Commercial arrangements with OpenAI, including revenue sharing, brought Microsoft $10 billion in the quarter, nearly 11% of its revenue and a $40 billion annual pace. OpenAI's unpaid balance to Microsoft rose to $7.2 billion, and Microsoft's stake slipped from about 25% to 24%. Commercial remaining performance obligation, the revenue customers have already contracted for, rose 73% to $677 billion, and capex hit a record $45.2 billion.

Read that next to Anthropic's S-1, which commits about $111 billion to Google Cloud, $110 billion to AWS and $31 billion to Microsoft "regardless of actual usage." The labs' noncancelable compute commitments are showing up as the hyperscalers' backlog. Every cloud now books revenue from the same handful of labs that compete with its own models. It's a good business for the clouds while lab revenue keeps compounding. The risk sits wherever a lab's growth stalls first.

Anthropic ships enterprise spend reports as a teardown finds Claude Code's agent status line ate 16% of one team's bill

Anthropic's new reports let managers see, over any date range, what a team produced with Claude, what it cost and where workflows broke, along with recommendations and pooled group budgets. The timing follows the S-1 numbers TechCrunch cited: an operating profit in the second quarter and 900% net revenue retention. Jellyfish released data from 276,000 of its customers the same day. Tools built on Anthropic models account for nearly a third of proposed code changes. Spending on agentic coding tools has grown tenfold since January and now equals about 2% of total engineering cost. Codex has only started showing up in that data since OpenAI's Astra. Jellyfish's AI product lead put Anthropic's hold down to workflows teams have already standardized on.

That 2% is exactly what the reports are meant to defend. A gist from Flamerix AI showed one way it leaks. About every 30 seconds, for each background agent, Claude Code sends the agent's entire transcript to its model, usually Opus, and asks it to "describe your most recent action in 3-5 words." The answer only updates a status line in the agents panel. Nothing turns it off, and it has been on by default since version 2.1.232 in August. The author's OpenTelemetry export put these summary calls at 16.4% of Claude Code spend, slightly more than the main conversation's 15.7%. They also made up one in five API requests and ran at roughly 9,000 input tokens per output token. The gist cites four other independent measurements of 18–26%, and it ships an 80-line Bun preload that disables the timer.

An engineer on the Claude Code team replied in the thread. They said the summaries only run for background agents that are actively working, that a 16.4% share looked like a team running mostly background agents, and asked for the session ID. Others pointed out that the figures are Claude Code's own cost estimates, not invoices, and that resent transcripts should mostly hit the cache. On a subscription, the loss is rate-limit headroom, not dollars. One Pro user reported their usage limits stretching noticeably in the past two days, as if something had been fixed. My read is that this is a design choice nobody priced, not a scandal. It's also the clearest argument yet for the dashboards Anthropic just shipped: the harness is a cost center, and the vendor runs it.

have their own slight quirks of how they handle agent.md files, etc., and so that cost to switch is not zero.

— source

Time to start questioning the motivations of model providers shipping their own harnesses.

— franticgecko3, Hacker News

A developer compiles a RISC-V CPU into Qwen 3.5's weights: hybrid models' recurrent layers can hold exact memory

Qwenputer is a stock Qwen 3.5 4B with 11 extra layers set directly by a compiler, with no training. Three of every four of Qwen 3.5's attention layers use Gated DeltaNet, a linear RNN. Its author found that pinning the gate and the write strength to 1 and using one-hot keys turns each head's state into exact, addressable RAM that stays the same size however long the context gets. On top of that RAM sits a 32-bit RV32IM core. It passes the official riscv-tests, ran 18 million instructions with no errors, and runs C and MicroPython inside the forward pass. Qwen calls it through its normal tool-call format, but no external runtime is involved. On long-integer arithmetic, Qwen 3.5 4B goes from 48.6% to 100%, matching what it gets with a Python tool. Its ordinary text stays within noise of the base model. There is even an experimental filesystem in the recurrent state that persists across chats.

Hacker News split between delight and "why not just use a computer." One commenter called it a Rube Goldberg version of the 2014 Neural Turing Machine. The arithmetic is the least interesting part. The author's follow-ups argue that hybrid models could learn to hide reasoning in their RNN state, out of reach of chain-of-thought monitors, and that compiled weights could work as built-in monitors the model can't tamper with. As more frontier-adjacent models adopt linear-attention hybrids, the idea that a recurrent state can hold exact hidden information matters more for interpretability than for calculators.

In practical open-weight news, Nvidia released Nemotron 3.5 Transcribe. It's a 30-billion-parameter hybrid Mamba-Transformer mixture of experts with about 4 billion active, and it transcribes 40 languages, translates 31 to or from English, and labels up to eight speakers in recordings up to an hour long. It ranks first among open-weight models on Hugging Face's English Open ASR leaderboard at 5.33% WER, and ties for first on Artificial Analysis's speech index at 2.8% with a sixth of the active parameters of the model it ties. It handles recorded audio only, not live streams. The open Granary v2 dataset ships with it, and a medical fine-tune cut word error rate 45.8% against the best prompt alone. Voice is crowded this month, with Mistral's Voxtral 3, ElevenLabs Realtime and Salesforce's Launa-Omni all out.

Subtly wrong fine-tuning data misaligns all 12 tested models more than obviously wrong data

Researchers from the Second Look Fellowship fine-tuned 12 open-weight models on correct, obviously wrong and subtly wrong advice across eight domains, from health to automotive. They replicated OpenAI's finding that subtle errors cause more emergent misalignment, meaning broad bad behavior on unrelated questions after narrow training. Across 96 model-domain pairs, subtle errors added 10.6 points of misalignment over obvious ones (95% CI 9.1–12.0), and they did so in every one of the 12 models. Fabrication and concerning-behavior scores on Petri evals rose the same way. Which domain the data came from barely mattered, a 5-point spread against 38 points between models. Misalignment also seemed to rise with model size. There are caveats: GPT-4o did the grading, and learning rates differed across models. The practical point stands, though. The errors that do the most damage are the ones reviewers are least likely to catch, so anyone fine-tuning on synthetic or scraped data should screen for plausible-but-wrong examples, not just garbage.

Two other papers are worth a line. The Trinity, a recursive self-improvement framework that evolves agent harnesses inside generated training worlds, claims a 26-point mean gain on 18 unseen benchmarks. A reader on Hugging Face pointed out that each benchmark got five preparation trials turned into skills before scoring. SPOT makes diffusion language models faster without changing their output. A 0.62M-parameter drafter predicts the order in which positions get filled, which cuts model passes up to 6.8x on SDAR-8B.

The reported gain therefore includes adaptation to the target benchmark.

— alphabell-research, Hugging Face

xAI offers to buy 418 Southaven homes if owners drop data center claims: the buildout's local costs keep rising

An xAI subsidiary, MZX, is offering to buy homes near its Southaven, Mississippi data center and gas-turbine plant if sellers sign away their claims in a resident lawsuit over near-constant noise and vibration, according to The Guardian and the Mississippi Free Press. A realtor told one resident the company plans to buy 418 homes. Renters say they have been given 60 days to leave. This lands alongside an NAACP suit over unpermitted turbines, a 135-day construction pause adopted by Memphis last week, and growing opposition to data centers ahead of the midterms. The capex lines in this week's earnings reports have real neighbors.

Two smaller fights point the same way. Meta says its new business terms, which publishers read as permission to train on their sites, may now be delayed past November 20 indefinitely, after objections from senators in both parties. And an ACLU report circulating on Hacker News describes an FBI agent stopping a driver because AI-flagged plate-reader data showed "suspicious travel patterns." A federal jury in Texas recently sided with another driver stopped the same way.