Google Launches Gemini 4 Carbon for Cyber Defenders as Anthropic’s Claude Chains Browser Exploits and Alphabet Hits a Half-Trillion Run Rate

Google Launches Gemini 4 Carbon, a Frontier Model Released First to Cyber Defenders

Google has a new top model, and most customers cannot use it yet. Google’s announcement of Gemini 4 Carbon, written by Google DeepMind SVP Koray Kavukcuoglu and published Wednesday, calls Carbon the company’s strongest model, with state-of-the-art cybersecurity and long-horizon capabilities. Google says Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59, and leads three security benchmarks: 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching. It builds on Gemini 4 Argon, which Google says launched two weeks ago. On internal benchmarks, Google reports more than twice Argon’s success rate at generating secure code, and says the Chrome team found Carbon performed more than 10 times better than another lab’s large frontier model at producing high-quality patches. Google also reports that Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0%, reduced unsupported claims by about half compared with Argon, and resisted every evaluated attack on Gray Swan’s indirect prompt injection benchmark.

Access is the catch. Starting this week, Carbon goes to a subset of cyber defenders, with more trusted partners added gradually through Google’s Fairwind Program, and Google says it is taking part in the U.S. government’s voluntary process for pre-release model testing. Developers, businesses and consumers will get it later, after more work on guardrails. Introductory pricing matches Argon at $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper; a footnote says the price rises to $4 and $20 after the introductory period. The Verge’s report on Carbon’s cybersecurity focus notes that the index places Carbon ahead of Claude Opus 5.5 and GPT-6 Astra, and that Google says Carbon also beat GPT-6 Astra on SEC-bench Pro.

“Internally, our teams have seen Carbon run for days to persist through dead ends to tackle complex, multi-step engineering tasks like root-causing and patching subtle bugs across large codebases and low-level systems.” — Google

Anthropic’s Red Team Shows an Unreleased Claude Model Chaining Flaws Into Working Browser and Kernel Exploits

The same week Google pitched Carbon to defenders, Anthropic showed what these capabilities look like in practice. Anthropic’s Frontier Red Team report on discovering and exploiting vulnerabilities with Claude, published October 27, describes five attack chains built by an unreleased internal research model after initial prompting. In Chrome, Claude combined a V8 heap corruption bug, a V8 sandbox escape, a CacheStorage race condition and a use-after-free in FedCM into a one-click exploit that ran code outside the browser sandbox in a default-configuration Chrome browser on macOS. Anthropic says 62 CVEs it reported to Google have been assigned and patched as of Chrome 156, all rated critical or high, including 13 sandbox escapes. The other chains bypassed Safari’s same-origin policy, pulled photos off a locked Android phone over USB, crashed Windows servers that answer HTTP/3 requests, and ran commands as root on Linux servers with a multipath TCP listener; the Linux proof of concept took 48 minutes to leak a kernel pointer and five more to reach a shell. Every issue was reported to vendors first under coordinated disclosure.

The Hacker News’ analysis of the Anthropic exploit chains adds useful caution. It says Google’s release notes credit an Anthropic researcher assisted by Claude for specific Chrome fixes, though not for all 62, and that Apple credits Anthropic researchers with Claude for the Safari fix. It also notes that Anthropic does not break down how much of each chain a person did, and that the public record for the Windows bug names no finder. Anthropic’s own advice is to scan code at pull-request time, and it has released a Claude Code security plugin for that purpose.

“Frontier models have the ability to discover and combine multiple vulnerabilities into a working exploit. These skills have historically been the domain of few specialist teams with expertise in these targets.” — Anthropic Frontier Red Team

Alphabet Reaches a Half-Trillion-Dollar Run Rate, but Capex and Cash Flow Weigh on Shares

Google’s parent reported its biggest AI-driven quarter yet and still disappointed investors. Alphabet’s third-quarter 2026 earnings release filed with the SEC shows revenue up 25% to $127.6 billion, Google Cloud revenue up 103% to $30.7 billion, and Search & other up 12%, with year-over-year search query growth at its highest in five years. Operating margin was 31.8%, and net income fell to $26.2 billion, or $2.12 per share, from $35.0 billion a year earlier. Sundar Pichai’s prepared remarks on the Q3 2026 earnings call add that Google’s first-party model APIs now process nearly 32 billion tokens per minute, 40% more than last quarter and more than four times a year ago, and that the Gemini app has reached one billion users.

CNBC’s report on Alphabet’s third-quarter results says shares fell more than 10% after the company missed analysts’ revenue and earnings estimates and reported negative free cash flow of $9.07 billion. Third-quarter capital spending was $51.36 billion, below the $55.09 billion analysts expected, and Alphabet held its full-year capex guidance at $195 billion to $205 billion. Microsoft reported the same day: Microsoft’s fiscal first-quarter results announcement shows revenue up 19% to $92.3 billion, Azure revenue up 48% and commercial remaining performance obligation up 73% to $677 billion. CNBC’s analysis of Microsoft and Google capex shortfalls reports that Microsoft cut its capex forecast for the year to $170 billion from $175 billion, and that CFO Amy Hood said delays at third-party GPU providers will push some capacity into the first half of 2027.

“We continue to expect our capex to increase significantly in 2027, and will provide further details on our Q4 call.” — Anat Ashkenazi, Alphabet CFO, via CNBC

OpenAI Gives ChatGPT Pro and Business Subscribers Up to $500 a Month in API Credits

OpenAI is using its consumer subscriptions to pull more people onto its developer platform. OpenAI’s developer guide to API credits for ChatGPT subscribers says eligible Pro and Business subscriptions can now claim free monthly API credits by linking to an API organization they own. Pro subscribers get credits equal to their tier: $100, $200 or $500 a month. Business workspaces get $20 per paid Standard seat and $100 per paid Premium seat, pooled and capped at $500 a month. The credits cover paid API Platform usage including the Responses, Agents, Decisions, Realtime and Images APIs and built-in tools, but not Codex usage through API keys, and no API payment method is required to use them. OpenAI’s Help Center article on the API credits benefit fills in the rules: the paid subscription must be at least seven days old, Plus, Enterprise, Edu and free trials do not qualify, annually prepaid Business workspaces get up to $6,000 upfront, the person activating must own both the ChatGPT subscription and the API organization, and the benefit may change or be removed in the future.

“Unused credits expire at the end of the billing cycle and don’t roll over.” — OpenAI

Anthropic Adds a Gallery of Ready-Made Artifacts to Claude for Legal

Anthropic is moving from model supplier to workflow provider for lawyers. Anthropic’s announcement of artifacts for legal work introduces a gallery of interactive tools inside Claude for Legal, which launched in May. The set includes a cite check that flags mismatched quotes and wrong page numbers, a jurisdiction survey matrix, a matter board, a playbook deviation board for contracts, a regulatory horizon scan and a tabular contract review, each with links back to sources. Partner-built artifacts include Harvey Intake, Goodwin’s AI Governance Benchmark, Foley & Lardner’s Obligations Checker and PwC’s Subpoena Response Tool, with more coming from Legora and Orrick. Anthropic says it has added more than 60 legal connectors since launch, including Bloomberg Law, Filevine and Legora, and that the gallery includes nine access-to-justice artifacts, five built by nonprofit and public interest groups it supported with Claude credits. It is also starting a Legal Ambassadors Program. Law.com’s Legaltech News report on the Claude for Legal artifacts describes the artifacts as reusable digital products that sit alongside the chat window.

“With these artifacts, a legal professional brings their own documents and sources, Claude completes the first pass, and a lawyer applies their judgment to the results.” — Anthropic

NVIDIA Releases Nemotron 3.5 Transcribe and Mistral Opens a Responses API Preview

Two developer releases target the plumbing of voice and agent applications. NVIDIA’s research page for Nemotron 3.5 Transcribe, published October 28, introduces an open-weight speech-augmented language model that handles verbatim, contextual, speaker-attributed and rich transcription, speech translation and language identification in 40 languages, on recordings up to an hour long. NVIDIA says it posts a mean word error rate of 4.71% on the English Open ASR Leaderboard, the best among open-weight models, cuts named-entity errors six-fold when given a list of expected terms, and can transcribe up to 2,238 seconds of audio per second on a single RTX PRO 6000 Blackwell GPU using its NVFP4 checkpoint and a draft model. The Nemotron 3.5 Transcribe model card on Hugging Face lists 34 billion total parameters with 4.3 billion active, BF16, FP8 and NVFP4 releases, and the OpenMDW-1.1 license.

Meanwhile, Mistral’s documentation for its Responses API says the API is in public preview and lets developers build with Mistral models using the OpenAI Responses format. The preview supports text, image and file inputs, JSON Schema structured output, streaming, client-executed function calling and reasoning effort settings, and the examples use Mistral Large 4. The API is stateless, so applications keep the conversation history and send it back with each request.

“There is no single correct transcript of a recording: what counts as correct depends on who reads it and for what purpose.” — NVIDIA

Meta Funds Five Universities to Test Muse Models on Science and Medicine

Meta is courting academic researchers with models and money. Meta’s newsroom post on its partnerships with five US universities says it is granting $2 million across Caltech, The Ohio State University College of Medicine, Louisiana State University, Indiana University and UTHealth Houston, plus access to its latest models and direct collaboration with Meta researchers. Projects include an LSU computer vision system to detect dislodged tracheostomy tubes in infants. Meta AI Research’s detailed write-up of the university projects says researchers will use Muse models including the open-weight Muse Spark Open and Muse Glimmer. At Caltech, the work includes MathDB Frontier, a benchmark built on a database of more than 86,900 research-level open problems, and TorchLean, a Lean 4 framework for agents that build formally checkable artifacts.

“The universities will independently conduct the research and retain control of all patient and clinical data; that data will not be shared with, processed by, or otherwise made available to Meta.” — Meta AI Research

Report: Cognition Weighs a New Round at $70 Billion to $80 Billion

Valuations in AI coding keep climbing, at least according to unnamed sources. Business Insider’s report on Cognition’s possible new funding round says the maker of the Devin coding agent has received offers valuing it at $70 billion to $80 billion and is seriously considering a round before the end of the year, weeks after announcing funding at a $48 billion valuation. The report says the talks are early and valuation targets could change.