🛡️ Google (59), GitHub (>40%) and Microsoft ($10B) — cyber defense, agents and the real money

Google announces Gemini 4 Carbon and takes the #1 spot on the Artificial Analysis Intelligence Index v5.0 with 59, two weeks after Argon. On GitHub, agents already start more than 40% of pull requests in September: 110 million in a single month, six times March’s volume. Microsoft reveals that Azure billed $33.2B (+48%) and that commercial arrangements with OpenAI generated $10B in the quarter, 11% of its revenue.

On the same day the strongest cyber-defense model stays restricted to a select group, code is already mostly written by agents and AI money concentrates in a few cloud and revenue-share pipes.

🛡️ Google Gemini 4 Carbon: #1 (59) with cybersecurity focus

On 28 October Google DeepMind announced Gemini 4 Carbon, its strongest frontier model so far. Carbon scores 59 on the Artificial Analysis Intelligence Index v5.0, ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57). On cybersecurity it reports 88.2% on SEC-bench Pro (finding real vulnerabilities), 54.8% on ExploitGym and 71.0% on CWE-bench (patches). Launch pricing matches Argon: $2 per million input tokens and $10 output, with cached input at 95% off; later it rises to $4/$20.

Access starts this week only for a subset of cybersecurity defenders in the Fairwind Program. Google says it will expand to more trusted partners and then to developers and businesses “as soon as possible,” with no date. The model also participates in the U.S. government’s voluntary pre-release testing process. It is the second Gemini 4 model in two weeks.

59AA Index v5.0 #1

88.2%SEC-bench Pro

$2 / $10intro price (then $4/$20)

🤖 GitHub: agents start >40% of PRs (110M in September)

GitHub’s Octoverse 2026 report, presented on 28 October at Universe, confirms that in September more than 40% of pull requests began with an agent’s first draft, up from 17% in March. September alone accounted for 110 million, roughly six times March’s volume. Repositories created this year with an agent configuration file logged eight times as many commits as those without one.

Agent-started PRs touch multiple files in 87% of cases (79% for others) and merge faster: 90% within two weeks versus 86% for the rest. Among active Copilot users, 55% already use agentic features (35% a year ago). The cloud agent, the most hands-off mode, grew 230% year over year. In the year ending 30 September there were 17.9 billion pushes, more than 2024 and 2025 combined.

>40%agent-started PRs (Sept)

110Magent PRs in September

8×commits in agent repos

💰 Microsoft: Azure $33.2B (+48%) and $10B from OpenAI (11%)

Microsoft published on 28 October results for the quarter ended 30 September (fiscal 2027 Q1). Azure brought in $33.2B, up 48% year over year and above expectations. Total revenue was $92.3B (+19%). For the first time Microsoft breaks out Azure revenue in dollars. In the same quarter Google Cloud reported $30.7B, so Azure finished slightly ahead.

In its 10-Q Microsoft declares $10.0B in revenue from commercial arrangements with OpenAI, including revenue-sharing. That equals 11% of quarterly revenue. Accounts receivable from OpenAI rose to $7.2B (from $6.0B in June). Microsoft’s stake in OpenAI is around 24%. Azure already accounts for roughly 36% of Microsoft’s total revenue.

$33.2BAzure (+48%)

$10Bfrom OpenAI in the quarter

11%of Microsoft revenue

The connecting thread

Carbon shows that the frontier-model cycle has shortened to two weeks and that the clearest jump is in cyber-defense tasks most teams still cannot use. At the same time, on GitHub the coding workflow has already flipped: more than four in ten PRs start with an agent and merge faster. The money confirms concentration: Azure grows at 48% and 11% of Microsoft’s revenue comes from OpenAI. The model advances faster than access opens, agents rewrite the daily development work, and capital is funneled through a few cloud and revenue-share pipes.

My take: Carbon is the most interesting technical data point of the day and the most frustrating to use. 59 on the updated index and 88.2% at finding real vulnerabilities are numbers that matter, especially if the model is twice as good at generating secure code as Argon. The problem is access: Fairwind and U.S. government pre-testing mean most security and product teams will wait weeks or months while Google iterates on guardrails. Meanwhile the $2/$10 intro price is attractive, but the jump to $4/$20 after the introductory period is the one that has to be modeled. On GitHub the shift is already operational: if 40% of PRs start with an agent and 90% merge within two weeks, the bottleneck has moved from writing to reviewing and to paying for the compute. The real friction is not “does the agent work?” but how many review cycles, how many tokens and how much quality debt accumulate when volume multiplies by six in six months. Microsoft’s number closes the loop: $10B from OpenAI in one quarter (11% of revenue) and $7.2B in receivables show that revenue-share and Azure consumption are already material, not anecdotal. Anyone building on these models has to measure cost per completed task and have an exit plan when restricted access or post-intro pricing tightens. The cyber model is stronger, agents already write half the code, and the money is already concentrated. All three push toward fewer options and more dependence on whoever controls access and the invoice.

📌 Artículos relacionados

→ Google DeepMind (Genie 4), Oracle and OpenAI — interactive worlds and silicon à la carte → Google Cloud (GPT-6), Anthropic (Duet) and Microsoft (83%) — agents and models in production → Anthropic ($3.4B), Nvidia ($4.8B) and Apple — energy, risk and agents

✉️ Newsletter

Enjoying the content?

Subscribe to the newsletter and get each article in your inbox.

🤖 Este artículo está escrito con asistencia de IA y editado por un humano.