Google DeepMind's Gemini 4 Carbon tops AI benchmark with strong security focus

Google DeepMind unveils Gemini 4 Carbon

Google DeepMind calls Gemini 4 Carbon its strongest model, with state‑of‑the‑art cybersecurity and long‑horizon capabilities¹. It follows the Gemini 4 Argon release two weeks earlier and will initially be provided to a limited set of cyber‑defense partners through the Fairwind Program^1,2.

On the SEC‑bench Pro vulnerability‑finding benchmark it scored 88.2%, while it reached 54.8% on ExploitGym and 71.0% on CWE‑bench for patching tasks. Internally the model reduced misalignment and fabrication rates in live coding tests and can run for days on complex, multi‑step engineering problems. Pricing mirrors Argon at $2 per million input tokens, $10 per million output tokens, with cached input tokens discounted by 95%¹.

Google is engaging with the U.S. government’s voluntary pre‑release testing to refine guardrails before broader release to developers and consumers. Google calls this a deliberate approach to rollout¹.

“our strongest model with state-of-the-art cybersecurity and long horizon capabilities”

Chrome adds Gmail‑sourced autofill

Google announced that Chrome will now use its Gemini‑powered Personal Intelligence to pull information from users' Gmail when filling online forms. Users can enable or disable the enhanced autofill in Chrome settings^3,4.

The updated autofill can suggest data such as flight confirmation numbers, membership IDs, tracking numbers, order numbers, addresses, phone numbers, and credit‑card details by scanning relevant emails^3,4. It builds on earlier enhancements that already drew names, passport numbers, email addresses, and physical addresses from Google Accounts and Google Wallet⁴. All suggested data remains encrypted and users retain control over what the feature can access^3,4.

TechCrunch notes the update could give users another reason to stay in Google’s AI ecosystem³.

OpenAI models added to ServiceNow and Oracle workflows

The partnerships add ChatGPT, Codex and other OpenAI models to ServiceNow products and to Oracle applications accessed through ChatGPT^5,6.

ServiceNow will embed OpenAI models and agents across its workflow suite, enabling users to query incidents, assess change risk and update records directly inside ChatGPT. The integration uses ServiceNow’s Action Fabric to enforce existing business rules and permissions, so actions respect each organization’s configuration. A preview also lets on‑call engineers view alerts and assign work within the chat interface⁵. Oracle’s tie‑up brings its applications and data into ChatGPT and powers new Oracle services with OpenAI models; Oracle already reports 130,000 active ChatGPT users and more than 95,000 active Codex users. Oracle’s recruiting team reports that tasks that once required two to four days now take 15–20 minutes with ChatGPT Work⁶.

Anthropic adds reporting tools for enterprise AI spend

Anthropic announced a suite of reporting tools on Wednesday to let enterprise managers see how much token spend costs and where workflows break down. The features let teams view production, cost and recommendations for improvement. Anthropic also introduced pooled group budgets to help control token usage⁷.

The new tools draw on internal analytics and employee surveys to break down a team’s work over a chosen date range. Reports show what was produced, the monetary cost and points where processes failed, and they do not use the data for model training by default. Anthropic’s move follows strong financial results, including operating profit in Q2 and 900% net revenue retention from Q3 2025 to Q3 2026⁷.

A Jellyfish report covering 276,000 of its customers found Claude Code and other tools built on Anthropic models behind nearly a third of proposed code changes, with spending on them up 10‑fold this year to about 2% of total engineering cost⁷.

Cursor releases Agent SDK

Cursor’s new Agent SDK lets teams create agents that react to events such as pull‑request submissions, compliance checks, or incident investigations. Teams can deploy these agents to Cursor infrastructure with a single command, treating them like any other software project⁸.

The Agent SDK provides templates and integrates with GitHub, GitLab, and Bitbucket, enabling agents to assess change risk, route reviews, and automatically approve compliant updates. A Cursor team built a knowledge‑base reviewer called kb‑reviewer in about a day; it now runs on every PR, flags contradictions, cites source notes, and prompts engineers to resolve issues. The platform also supports handing tasks to specialized sub‑agents, parallel execution, built‑in evals, and event replay for reliable workflow testing⁸.