Today’s through-line is security-first frontier delivery, enterprise workflows inside chat interfaces, cloud giants’ financial ties to model labs, and open-weight speech foundations: Google ships Gemini 4 Carbon, OpenAI expands with ServiceNow, Microsoft discloses about $10 billion in a single quarter from OpenAI arrangements, Google pushes AI Mode ads into more markets, and NVIDIA open-sources Nemotron 3.5 Transcribe and Granary v2.

1. Google launches Gemini 4 Carbon: cybersecurity and long-horizon tasks advance together

On October 28, 2026, Google DeepMind announced Gemini 4 Carbon, calling it the next frontier model after Gemini 4 Argon roughly two weeks earlier, and the company’s strongest yet on cybersecurity and long-horizon capability. Per Google, Carbon leads the updated Artificial Analysis Intelligence Index v5.0 with a score of 59; VentureBeat, citing the same index, places Claude Opus 5.5 (58), GPT-6 Astra (57) and Argon (50) behind it. On cybersecurity evaluations (which Google says were run without guardrails), Carbon reached 88.2% on SEC-bench Pro (vulnerability discovery), 54.8% on ExploitGym (exploitability validation) and 71.0% on CWE-bench (patch repair). Internal benchmarks show more than a twofold rise in successful secure-code generation versus Argon; the Chrome team says high-quality patch performance exceeded another lab's frontier model by more than 10x; cloud-security firm Wiz likewise observed a lead on web penetration-testing benchmarks. For long-horizon engineering, Google says internal teams have seen Carbon run for days, push through dead ends, and complete root-cause analysis and fixes across large codebases; on WebDev Arena its Elo rose 89 points to 4th (1768).

On safety and alignment, Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0% (cheat attempts 0.4%); high-severity misalignment alerts fell nearly 80%; versus Argon, unsupported claims roughly halved, while over-refusal and under-serving fell about 26% and 29%; on Gray Swan’s indirect prompt-injection benchmark, attack success was 0.0% (Argon 0.7%). Release pacing is deliberately cautious: the company joined U.S. government voluntary pre-release testing; access opens this week to selected cyber defenders, then expands via the Fairwind Program to trusted partners, with a plan to reach developers, enterprises and consumer surfaces soon. Introductory pricing matches Argon: $2 per million input tokens and $10 per million output (95% discount on cached input); after the intro period, $4 input / $20 output.

Key takeaway: Carbon ties leaderboard leadership to a “defenders first, then broader release” path—cybersecurity skill is both the pitch and the rationale for delayed general access.

Sources: Google Blog; The Verge; VentureBeat

2. OpenAI and ServiceNow expand partnership: governed workflows enter ChatGPT

OpenAI and ServiceNow announced an expanded partnership on October 28, 2026, building on January’s relationship to embed more OpenAI models and agents in ServiceNow products and to bring ServiceNow workflows into ChatGPT and Codex. New ServiceNow integrations can carry business context and workflow actions directly into ChatGPT; customers connect their own instances and choose which capabilities to open, while ServiceNow Action Fabric keeps existing business rules and permissions. A ServiceNow MCP plugin is live on the OpenAI marketplace; enhanced capabilities are expected before end of 2026.

The partners also outlined three product lines: Codex inside AI Workflow Factory (expected 2027); joint security and risk offerings combining OpenAI cybersecurity capabilities with ServiceNow asset, identity and business context (early access expected January 2027, generally available in Q2); and OpenAI speech capabilities inside Super Intelligence Control Tower to describe AI use cases and complete evaluation fields (expected Q1 2027). ServiceNow plans ChatGPT Work and Codex for all employees; OpenAI also uses ServiceNow workflows in its own operations.

Key takeaway: Chat interfaces are absorbing ITSM/workflow permissions and approval structures—“can talk” has to connect to “can change tickets under the rules.”

Sources: OpenAI; Business Wire / ServiceNow

3. Microsoft FY27 Q1: $10 billion revenue from OpenAI commercial arrangements

On October 28, 2026, Microsoft reported FY27 first-quarter results for the period ended September 30, 2026: revenue $92.3 billion, up 19% year over year; operating income $45.9 billion, up 21%; GAAP net income $36.6 billion, up 32%. Microsoft Cloud revenue was $63.9 billion, up 30% (31% in constant currency); Azure revenue grew 48%; commercial remaining performance obligation rose to $677 billion, up 73%.

Per TechCrunch’s reading of regulatory disclosures, Microsoft recorded $10 billion that quarter from OpenAI commercial arrangements (including revenue share), about 11% of total quarterly revenue; for full FY2026, OpenAI-related arrangement revenue was $24.1 billion, and at the current quarterly pace annualized near $40 billion. As of end-September, receivables from OpenAI were $7.2 billion (versus $6.0 billion at end-June). Ownership stake edged down from about 25% in June to about 24% in September. Since October 2025 the exclusivity bind has loosened: OpenAI may work with other clouds and has committed to purchase an additional $250 billion of Azure services; under the April 2026 agreement, revenue-share terms run through 2030 with a cap.

Key takeaway: Model labs remain an explicit variable in cloud-giant earnings—partnerships, revenue share and receivables scale together; competition and dependence have not cancelled each other out.

Sources: Microsoft press release; TechCrunch

4. Google expands AI Mode ads to 10 new markets

On October 28, 2026, Google announced that ads in AI Mode expand to Australia, Canada, India, Kenya, Malaysia, New Zealand, Nigeria, Pakistan, the Philippines and Singapore, with further expansion planned in Q4 2026 including Brazil, Mexico and the United Kingdom. In the U.S., multiturn ads launch so follow-up query turns can also show ads. Ads will be clearly labeled “Sponsored.” Google says AI Mode queries average about three times the length of traditional search, and that click-through rates for sponsored and organic shopping listings are already similar. The move landed the same day as Alphabet’s third-quarter results; per TechCrunch, Alphabet’s quarterly revenue was about $127.6 billion, up roughly 25% year over year, while core search-ad growth slowed.

Key takeaway: Generative-search monetization is moving from pilots to multi-country rollout; ad formats are starting to follow multi-turn dialogue rather than stopping at a single results page.

Sources: Google Ads Blog; TechCrunch

5. NVIDIA open-sources Nemotron 3.5 Transcribe and Granary v2

On October 28, 2026, NVIDIA released open-weight Nemotron 3.5 Transcribe and published the Granary v2 multilingual speech data-curation pipeline and inventory. Transcribe packs transcription, speaker attribution, translation and audio Q&A into one checkpoint: it can transcribe 40 languages, translate between 31 languages and English, and do speaker attribution in 13 languages; it supports recordings up to about one hour and up to eight speakers, with word-level timestamps. Weights are on Hugging Face and ModelScope and deployable with vLLM; Baseten, DeepInfra, Simplismart and others offer hosted endpoints.

On the Hugging Face Open ASR Leaderboard, Transcribe leads open-weight English with an average word error rate of 5.33%, and ranks near the top in multilingual and long-audio comparisons. On Artificial Analysis’s speech-transcription board, with about 4B active parameters (a hybrid Mamba–Transformer MoE totaling about 30B), it scored 2.8% AA-WER, tied for first among open weights. For speaker attribution, it posted the lowest DER (18.26%) on David AI’s 21-language diarization set and ranked first among 14 systems on VoiceArena Diarization-Bench. Granary v2’s public inventory covers about 463,062 hours across 48 languages and 16 source corpora; a medical fine-tuning case cut WER on a held-out speaker set from 22.44% to 12.17% versus the best prompt baseline (about a 45.8% relative drop).

Key takeaway: Open speech models are moving from “can dictate” to a single promptable, attributable, domain-fine-tunable checkpoint—and the data pipeline itself is becoming a product.

Sources: Hugging Face Blog; NVIDIA Nemotron

---

One-line takeaway: Gemini 4 Carbon refreshes the leaderboard with cybersecurity and long-horizon skill yet ships limited first; OpenAI and ServiceNow push governed workflows into chat; Microsoft books about $10 billion in a quarter from OpenAI-related revenue; Google widens the AI Mode ads map; and NVIDIA fills another open-weight piece with speech transcription and a data pipeline.