Wednesday, 28 October 2026

Google announced Gemini 4 Carbon on Wednesday and calls it its strongest model, aimed at cybersecurity and long, multi-step tasks [1][4]. It is not on general release. Starting this week, the company says, a subset of cyber defenders gets access, followed by more trusted partners through its Fairwind Program; developers, businesses and consumers get it later, with no date given [1]. Fairwind is Google's vetted early-access scheme for governments, critical infrastructure operators and core technology platforms, with background checks, phishing-resistant MFA and access limited to security teams [3].

The price is public before the model is. Carbon launches at the same introductory price as Gemini 4 Argon, $2 per million input tokens and $10 per million output tokens, with cached input 95% off [1]. After the introductory period, Google says, the price becomes $4 per million input tokens and $20 per million output tokens [1]. When that period ends is not public.

The cyber scores are Google's own, measured without guardrails: 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching [1]. One outside ranking backs the general claim: Artificial Analysis lists Gemini 4 Carbon (High) first on its Intelligence Index at 59, one point ahead of Claude Opus 5.5 at 58 [2]. Google also says Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0% and resisted every attack in Gray Swan's indirect prompt-injection benchmark, where Argon let 0.7% through [1].

Cursor has made its Agent SDK generally available, the framework it says it uses to build its own managed agents, including Bugbot, Security Reviewer and Rollouts [1]. An agent is a project folder: prompts are written in Markdown, tools are TypeScript functions, and the evals that test it sit alongside, so changes go through the same version control and code review as the rest of your software [1]. You start with `npx @cursor/agent-sdk init ./my-bot`, optionally from a template for a security reviewer, an owners-policy approver, a PR autofixer, or a Linear or Jira triager [2].

Agents wake on pull request, comment and CI events from GitHub, GitLab and Bitbucket, on Slack mentions and DMs, on webhooks from any provider and on a cron schedule [2]. Each conversation keeps its own session and a record of every message and tool call, and a follow-up on the same pull request or Slack thread resumes that session [2]. An eval sends a fixed message and checks which tools ran and what the reply looked like; `agent-sdk github replay` turns a real pull request into webhook payloads so a bot can be tested without waiting for a push [2].

Running `agent-sdk deploy` builds the agent, gives it a URL and starts it on Cursor's infrastructure, with secrets and integrations managed from a dashboard [1]. You can also run it yourself: each agent is a single Node process that can be packaged in a container [1], and the docs cover self-hosting for teams that need their own network or process controls [3]. The company says one internal team built a reviewer that checks every pull request against its knowledge base in about a day [1]. What hosted agents cost is not public in the launch posts.

Anthropic has acquired Mithril Medicine, a one-person company developing treatments for two rare forms of muscular dystrophy [1]. Its founder, Cody Tranbarger, used Claude to organise contracts, logistics and experiments with 11 external labs, and Mithril took one candidate from an idea to promising results in mice in less than 18 months, the company says [1]. If the candidates hold up, Anthropic expects a drug could start human trials as soon as late 2028 [1].

The acquisition came with a progress report on six partnerships [1]. At Generate Biomedicines, six scientists used Claude Code to redesign part of a CAR T-cell therapy; one design was about 70% shorter and, in lab tests, let immune cells make roughly 10 times as many receptors [1]. BigHat Biosciences made and tested about 1,010 Claude-proposed variants of two antibody fragments in about 10 weeks [1]. At Enveda, across 1,061 pages of manufacturing documents, Claude caught all 11 issues human reviewers had found, flagged seven they had missed, and cut review time by 89% [1]. Bristol Myers Squibb is using Claude to assess clinical-trial plans; Divergence Labs and The Column Group are the other partners [1].

Anthropic says the work is mostly preclinical: for the most part the drugs are not yet being tested in humans [1]. Its S-1 filing is more guarded than its blog, according to Wired. The filing says life-sciences work takes the company outside its historical experience as a software and research organisation, and that any drug candidates face lengthy, costly and uncertain development and "may never generate revenue" [2].

Anthropic has signed a new OneGov agreement with the General Services Administration that gives government buyers a 40% discount on Claude Enterprise, Claude for Government and the Claude API [1]. For the first time the discount applies on every route: buying direct now, and through cloud providers in the coming months [1]. It also extends past federal agencies to state, local, tribal and territorial governments [1]. There is no minimum spend commitment, and agencies pay for the usage they consume [1]. Nextgov reports that the discount covers token-based usage, took effect on Tuesday, and runs under a 27-month agreement [2].

It arrives as a flat-fee deal ends. Under the earlier OneGov agreement, which expires on Oct. 31, federal agencies could get Claude for Enterprise and Government for $1 per user, according to Nextgov [2]. Anthropic estimates the new discounts will save government $1 billion [1]. GSA says the OneGov programme has produced more than $1.8 billion in savings, roughly $1.6 billion of it from AI agreements, Nextgov reports [2].

Anthropic says about 500 US federal, state and local government organisations now use Claude, and that weekly Claude for Government active users are up about eightfold since February [1]. On Amazon Bedrock and Google Cloud's Vertex AI, Claude is authorised for FedRAMP High; Claude Enterprise is in process for FedRAMP Moderate through the FedRAMP 20x programme [1]. Agencies can buy at the OneGov discount through GSA's Multiple Award Schedule today, the company says [1].

Google has launched a preview of its Connected Apps Platform, a single place for developers to build apps that work inside Search and the Gemini app [1]. Until now, building a Connected App meant working closely with Google, according to 9to5Google [3]. Developers connect a cloud-hosted MCP server, or Android MCP through the AppFunctions APIs, then register, test and submit the app in a new Google Connected Console [1]. Before review they choose whether it goes live in Search, the Gemini app or both; once it is live they can track traffic, latency and error rates and keep control of verified logos and authenticated accounts [1].

Access is limited. TechCrunch reports that the preview is open to US developers who sign up for a waitlist, and that Google has not said when it will be broadly available [2]. Google says it is opening access in phases and will add developers, markets and Google products over the coming months [1].

The consumer side is growing at the same time. Google says it has launched more than 50 Connected Apps across Search and Gemini; integrations with Adobe, Experian, Localiza and Peloton are live in Gemini and will reach Search in the coming weeks, and Calm is a new partner on both [1]. 9to5Google lists Dropbox, Instacart, OpenTable and Ticketmaster among the apps already available in Gemini [3]. TechCrunch frames the platform as a step toward people using services through their AI assistant instead of opening each app, and compares it to the way SEO let businesses show up in search rankings [2].

Chrome's enhanced autofill can now suggest information pulled from Gmail while you fill in online forms, Google says [1]. It runs on Gemini with Personal Intelligence and can surface order numbers, tracking information and flight bookings [1]. The Verge adds hotel reservations to the list, and notes that last year's update already let enhanced autofill draw on services like Google Wallet [3]. TechCrunch lists membership IDs among the details it can suggest, and describes the feature as an expansion of Personal Intelligence, which lets people connect Google apps to Gemini [2].

The rollout starts in the US on desktop, Android and iOS, and only for Gemini in Chrome users who have Personal Intelligence turned on [1]. Google says it will soon extend enhanced autofill to information in the images stored in your Google Photos library [1]. Sensitive data is encrypted, the company says, and you manage whether enhanced autofill is on and what it can access in Chrome's settings [1]. The Verge notes you can also turn it off completely [3].

What the announcement leaves out matters for anyone who handles other people's data. Google's post does not say how long Gmail content used for a suggestion is kept, or whether it is used for anything beyond the suggestion; that is not public [1]. TechCrunch reads the change as one more reason for users to stay in Chrome and in Google's AI ecosystem [2]. If you use one browser profile for work and personal mail, check which Google account Personal Intelligence is connected to before you rely on it.

Nvidia has released Nemotron 3.5 Transcribe, an open-weight speech model that does transcription, speaker attribution, translation and questions about the audio in one checkpoint [1][3]. It transcribes 40 languages and can translate 31 languages to and from English, and it handles recordings up to an hour long with up to eight speakers, according to TechCrunch [2]. You steer the output with plain-language prompts, and prompts can include product terms and formatting rules for fields such as healthcare, financial services and customer service [3]. Nvidia also released Granary v2, the open speech dataset and data pipeline behind its training [1][2].

Nvidia's blog puts the model at about 4.3 billion active parameters [3], and its Hugging Face post says about 4 billion of its 30 billion parameters are active per token [1]. TechCrunch reports the 30 billion figure and says it can run on a single chip, from the Jetson AGX Thor robotics controller and DGX Spark to Blackwell data-centre chips [2]. Nvidia says quantized versions are coming soon for local use on RTX Spark and DGX Spark systems, deployable through NeMo-Speech.cpp or vLLM [3]. Its accuracy and leaderboard claims come from Nvidia's own posts [1][3].

There are limits worth knowing before you swap it in. Transcribe works on recorded audio, not livestreams, according to TechCrunch [2]. Nvidia says language support differs by task and that audio question answering is trained in English, and it tells users to spot-check names, speaker changes and answers against the recording [1]. Because the weights are open, you choose where the audio is processed, from a desktop to the cloud [3].