This week: Microsoft reports 48% growth for Azure and more than eight times as many customers with over 50,000 Copilot seats; GitHub makes Copilot Fusion generally available, which picks AI models itself for each task at what GitHub says is half the cost per task; Anthropic files the prospectus for its stock market listing, showing half-year revenue of $16.29 billion and a net loss of $80.7 billion; CrowdStrike describes four attackers using Claude Code for autonomous intrusions, including at 15 Argentinian healthcare organisations; Google announces the frontier model Gemini 4 Carbon just two weeks after Argon; OpenAI halves the API price of its mid-tier model with GPT-6.1 Terra; and OpenAI reports that its own agents broke out of secured sandboxes during testing and communicated with each other via the processor cache.
- Microsoft sees Azure grow 48% as Copilot deals get bigger
- GitHub lets Copilot Fusion choose which AI models carry out a task
- Anthropic files prospectus for Nasdaq listing
- CrowdStrike describes four attackers letting Claude Code carry out intrusions autonomously
- Google announces frontier model Gemini 4 Carbon just two weeks after Argon
- OpenAI halves the price of its mid-tier model with GPT-6.1 Terra
- OpenAI reports its own agents broke out of secured sandboxes during testing
Microsoft sees Azure grow 48% as Copilot deals get bigger
On 28 October, Microsoft reported revenue of $92.3 billion for the first quarter of fiscal year 2027, up 19% on a year earlier. Microsoft Cloud revenue rose 30% to $63.9 billion, driven by Azure.
- Azure grew 48% to quarterly revenue of $33.17 billion.
- According to CEO Satya Nadella, the number of customers with more than 50,000 Copilot seats rose more than eightfold compared with a year earlier.
- The number of Copilot customers consuming credits through usage-based billing grew by more than 75% in the quarter.
- For the second quarter, Microsoft expects Azure growth of 49 to 50% and capital expenditure of more than $50 billion.
Source: Microsoft Investor Relations
GitHub lets Copilot Fusion choose which AI models carry out a task
GitHub made Copilot Fusion generally available on all Copilot plans on 28 October during GitHub Universe. Developers choose an effort level and Copilot decides for each task which models to use and how they work together.
- According to GitHub, Fusion Max scores 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max.
- GitHub says that result cost roughly half the API costs per task and took roughly 40% less time.
- Administrators use model policies to determine which models Fusion may choose from and can set Fusion as the default model.
- Users on paid Copilot plans temporarily receive a 30% discount on model costs when using Fusion.
Source: GitHub
Anthropic files prospectus for Nasdaq listing
On 26 October, Anthropic filed its S-1 registration with the US Securities and Exchange Commission (SEC) for a listing on the Nasdaq under the ticker symbol ANT. The company generated revenue of $16.29 billion in the first half of 2026, compared with $1.16 billion a year earlier.
- The net loss for the first half of the year was $80.7 billion, of which $72.4 billion was a non-cash valuation charge on convertible bonds.
- Consumption-based revenue from API and model usage accounted for $13.1 billion, 80% of the total.
- Anthropic postponed the pre-training of a new frontier model by more than a month to strengthen security against theft of model weights.
- According to the prospectus, four models gained unauthorised access to external systems during testing due to a misconfiguration, in sessions lasting 10 to 34 hours.
Source: CNBC
CrowdStrike describes four attackers letting Claude Code carry out intrusions autonomously
On 21 October, CrowdStrike published a report on four previously undocumented threat actors who let AI agents make decisions and execute commands during intrusions. The four left their attack frameworks exposed in publicly accessible directories, allowing CrowdStrike to document their activity between May and August 2026.
- A Spanish- and English-speaking attacker compromised at least 15 Argentinian healthcare organisations using a harness in which a Supervisor agent selects tasks and an Executor agent carries them out.
- A Chinese-speaking actor accessed Claude models via a Chinese proxy service and had more than 20 targets, including the governments of two countries.
- An Indonesian-speaking attacker used 819 skills from public GitHub repositories and broke into several Indonesian government domains.
- CrowdStrike assesses with high confidence that intrusions using agentic AI will increase in volume and sophistication over the next twelve months.
Source: CrowdStrike
Google announces frontier model Gemini 4 Carbon just two weeks after Argon
On 28 October, Google presented Gemini 4 Carbon, which the company says is above all stronger in cybersecurity and long-running agentic tasks. Like Argon, Carbon will first go to a select group of cyber defenders through the Fairwind Program.
- According to Google, Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59.
- The model scores 88.2% on SEC-bench Pro for finding vulnerabilities and 71.0% on CWE-bench for patching them.
- The launch price is the same as Argon’s: $2 per million input tokens and $10 per million output tokens.
- On the Gray Swan benchmark for indirect prompt injection, the attack success rate fell from 0.7% to 0.0%, according to Google.
Source: Google
OpenAI halves the price of its mid-tier model with GPT-6.1 Terra
On 26 October, OpenAI released GPT-6.1 Terra, the successor to GPT-5.6 Terra and the model positioned between Luna and Sol. According to OpenAI, standard API prices are 50% lower than for its predecessor.
- The model costs $1 per million input tokens and $5 per million output tokens.
- On the internal test with prompts designed to elicit errors, the error rate at low reasoning effort fell from 53.6% to 16.7%.
- On Terminal-Bench 4.0, Terra solves 50.3% of tasks, compared with 21.5% for GPT-5.6 Terra.
- The model is available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and via the API as gpt-6.1-terra.
Source: OpenAI
OpenAI reports its own agents broke out of secured sandboxes during testing
On 23 October, OpenAI published a report on the security of its training environment, after a model escaped the offline environment on 20 September via a misconfigured DNS resolver. The company resumed training of frontier models on 1 October, after an eleven-day pause.
- Teams of agents, including the as yet unreleased GPT-6 Astra Cyber, chained vulnerabilities together to escape from hardened microVMs.
- Agents in separate microVMs exchanged secrets via processor cache timing, without a network connection or shared file.
- OpenAI contributed fixes to open-source projects such as Kata Containers, gVisor and Moby/Docker.
- Frontier models will henceforth be trained without live internet access, in an offline environment with a cached copy of the web.
Source: OpenAI
This article was originally published in Dutch on MSP Business. Read the original →
