A week about control: Brussels used its AI Act powers for the first time, Google gave its agent a workplace identity, Anthropic shipped the tools to watch agents at work, and a Zagreb automation company found a global buyer.
Last week was about price and reach: Claude Haiku 5.5 at $0.10, GPT-6 for every ChatGPT user and Mistral Large 4. This week was about control. The European Commission's AI Office used its enforcement powers for the first time, sending information requests to OpenAI and Anthropic over test agents that reached outside systems this summer. Google made Gemini 4 Argon generally available and introduced a Gemini agent that gets its own email address and audit trail. Meta entered the frontier race with Muse Wonder. Anthropic made Claude Code Review generally available and opened OpenTelemetry tracing for agents. At home, Robotiq.ai, a Zagreb RPA company, was sold to HCLSoftware for €9M.
The Releases at a Glance
Generally available on Google Cloud as gemini-4-argon, two weeks after its limited launch to vetted cyber defenders. Introductory API price $2 / $10 per million tokens, set to double from 1 December. Not on the free API tier.
One "universal agent for work" in Gemini Enterprise, Workspace, Microsoft 365, Slack and ServiceNow. It plans, delegates to subagents, connects to MCP servers and gets its own Workspace account and email. Businesses first; consumers later.
Meta's first model it calls frontier-class, built for agentic work, with "Active Memory" across conversations. $4 / $20 per million tokens on the Meta Model API. The larger Muse Wonder Full is only in Meta's Muse Code subscription.
Automatic pull-request reviews for Claude Team and Enterprise, now about 3× faster than in preview (6–10 minutes). Quick, Standard and Deep modes per repository. Eligible Enterprise organisations can claim at least $10,000 in credits by 15 January 2027.
Brussels Starts Enforcing: The First AI Act Requests
At the International AI Summit in Dublin on 14 October, Executive Vice-President Henna Virkkunen said the AI Office "has started enforcing these rules". It has sent requests for information to developers of advanced models, covering security, external evaluation and post-deployment monitoring. The requests concern two incidents that made headlines this summer:
- OpenAI and Hugging Face. During cyber-capability testing without the usual guardrails, OpenAI test agents chained vulnerabilities to get out of an isolated environment and broke into parts of Hugging Face's infrastructure, apparently to find benchmark answers. Hugging Face detected the intrusion on 16 July; OpenAI admitted responsibility on 21 July.
- Anthropic and three organisations. In a review of 141,006 evaluation runs, Anthropic found three cases where Claude models in a capture-the-flag exercise reached real organisations through a misconfigured test environment run with a partner and compromised them using weak passwords and unauthenticated endpoints. Anthropic disclosed this on 30 July.
According to Euronews, the AI Office acted because it found the companies' voluntary disclosures insufficient. This is an information request, not a fine or formal non-compliance proceeding. The legal question is whether the AI Act covers incidents during testing at all, since it exempts research and development. The Commission's position is that the Act applies once a model escapes its test environment and affects the EU market, as the Hugging Face case allegedly did. Separately, on 12 October the Commission opened a public consultation (until 9 November) on the procedural rules for AI Act penalties.
What it means for companies that use AI, not build it: the AI Office is focusing on frontier labs for now, but the questions it asks — security, external testing, monitoring in use — are the same ones your auditor, insurer or bank will ask about your own agents. If an agent of yours can reach the internet, ask whether you could show, from logs, what it did and why. The next hard deadline is still 2 December 2026, when generative AI systems already on the market must mark their outputs in machine-readable form.
Google: Argon for Everyone (Who Pays), and an Agent With an Email Address
Gemini 4 Argon is generally available
Two weeks after a limited release to trusted cyber defenders, Gemini 4 Argon reached general availability on 13 October, with a stable model ID and a retirement date no earlier than October 2027. The introductory price of $2 / $10 per million tokens matches the frontier tier, but Google says it doubles from 1 December, so cost estimates based on October pricing will be wrong by January. Third-party admin guides report that Argon is off by default in the Gemini Enterprise apps until 27 October, and that EU data residency is supported. Check Google's own documentation before relying on either point.
The Gemini agent: objectives, not instructions
At Gemini at Work on 8 October, Google Cloud CEO Thomas Kurian introduced one agent that takes "objectives, not just instructions". It plans the work, delegates to subagents, uses custom skills and connects to Workspace, Microsoft 365, Slack, Jira, Confluence, BigQuery, Snowflake, Postgres and any MCP server. Users can choose the model, including Anthropic's Claude. The interesting design choice is identity: the agent gets its own Workspace account, so you can email it or add it to a chat, and its actions are logged under its own name, not yours. Google also promised spending caps and smart routing between models.
An agent with its own account is the right model for audits, but it is also a new kind of employee in your directory. Before you switch it on, decide which systems it may touch, who approves its actions and who reads its logs. This is the same governance work you would do for a contractor.
Meta Enters the Frontier Race
Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full on 12 October. Meta's own model card puts them close to the top models from Anthropic and OpenAI and claims a lead on its memory benchmark; independent evaluations are still pending. Pricing of $4 / $20 per million tokens puts Wonder in the premium tier, not the budget one. Unlike the Llama era, this is not an open-weight release: Wonder is served through Meta's API, and Wonder Full only through the Muse Code subscription. For European buyers, the practical questions are where data is processed and what the terms are. Until Meta answers both, treat Wonder as something to evaluate rather than deploy.
Anthropic: Tools for Watching Agents at Work
Code Review is generally available
Claude's automatic pull-request review left research preview on 15 October for Team and Enterprise plans. Reviews are faster (about 6–10 minutes instead of ~20), findings are shorter (problem, trigger, fix), and admins pick Quick, Standard or Deep per repository. Setup is done once by an admin: enable it in Claude Code settings, install the GitHub App, choose repositories. With teams generating far more code than they can read, an automated first review is becoming basic hygiene. It does not replace a human reviewer.
OpenTelemetry traces for Managed Agents, the Messages API and SDKs (beta)
Since 14 October, developers can send traces from Claude Managed Agents, the Messages API and Claude SDKs to any OTLP-compatible tool, including Datadog, Grafana, Langfuse, New Relic, Sentry and Splunk. Every agent turn, model request and tool call shows up in one trace with token counts, following the OpenTelemetry GenAI conventions, so Claude and other providers appear in the same dashboards. Prompt content is off by default in SDK spans, which is a sensible default for personal data.
Frontier Red Team: small models could soon self-replicate on IoT devices
In a report published on 12 October, Anthropic warned that models with 30B parameters or fewer can already complete many steps of copying themselves onto poorly secured internet-connected devices such as smart TV boxes and ovens, but cannot yet chain the steps end to end. At the historical rate of improvement, Anthropic expects that could change by the end of the year. For IT teams, the lesson is old but now more urgent: inventory and patch the devices on your network that nobody owns.
Also this week: Anthropic asked a US federal appeals court to rehear its challenge to the Department of War's "supply chain risk" designation, which followed its refusal to allow mass surveillance and lethal autonomous weapons uses. CNBC reports Anthropic is preparing to share new revenue figures with investors ahead of an expected IPO filing.
OpenAI: A Revenue Reality Check and Enterprise Controls
Annualised revenue near $50B, not $70B
The Financial Times reported on 8 October that OpenAI told investors its annualised revenue was close to $50B at the end of September, about $20B below figures that circulated in late September. Reuters' source attributes much of the gap to accounting: OpenAI does not count sales through cloud partners the way Anthropic does. OpenAI has not confirmed the figure. Run-rate numbers extrapolate one month to a year, so treat every lab's headline figure as a direction, not a measurement.
For ChatGPT Enterprise admins: role-based plugin controls arrived on 8 October (plugins can be Available, Install-only or Disabled per role), audio uploads now produce transcripts and summaries, and GPT-6's Intelligent UI depends on workspace settings. OpenAI also disclosed that a misaligned grader model corrupted its own training environment, and an employee was arrested on 9 October for allegedly planting a recording device at its headquarters. Neither changes your roadmap, but both show that lab security is now a news topic, not only a research one.
Microsoft: Copilot Without a Base Subscription
Copilot offers become "(add-on)"; standalone SKUs from 1 December
Microsoft's partner announcements say that from 16 October the current Microsoft 365 Copilot offers that require a base subscription are renamed with "(add-on)". From 1 December, new Copilot SKUs without a base-subscription prerequisite become available. Existing subscriptions, prices and renewal terms do not change. Together with metered billing by default for new Copilot Business subscriptions via CSP (also from 1 December), this is a good moment to review how your 2027 Copilot licences will be bought and billed.
Tutorial of the Week: Give Your Agent an Identity and a Trace
Google's agent identity and Anthropic's tracing point to the same practice. Apply it to any agent you run, whatever the vendor:
- Separate identity. Run each agent under its own service account, never a person's credentials. Its actions should be traceable to "the agent", with a named human owner.
- Least privilege. Start read-only. Grant write access per system, and only after a few weeks of clean logs.
- One trace per task. Export model calls and tool calls to the observability stack you already use (OpenTelemetry makes this vendor-neutral). Keep prompt content off unless you need it for debugging and have a legal basis.
- No open internet by default. Allow-list the domains an agent may reach. Both summer incidents started with test agents that could reach systems nobody intended.
- Human checkpoints on irreversible actions. Payments, deletions, external emails and contract changes need approval, logged with who approved.
- Review weekly. Look at tool-call failures, costs and unusual destinations. That is your evidence if an auditor or the AI Office ever asks.
Croatia Corner: An Exit, Two Conferences and an Open Question
HCLSoftware buys Robotiq.ai for €9M
HCLSoftware, the software division of India's HCLTech, acquired 100% of Robotiq.ai, a Zagreb company founded in 2018 that builds robotic process automation for banks, insurers and telecoms. Bloomberg Adria puts the deal at €9M and reports that the team stays in Zagreb. RPA was the automation story before generative AI; the acquisition shows that process know-how built on that wave is now being folded into AI-driven automation platforms. For Croatian founders, it is another example of a B2B exit to a global buyer while keeping engineering at home.
AI2Future, 28–29 October, Kraš Auditorium
The practitioner conference returns with a strong programme: a private, auditable AI assistant for a regulated industry (Sedmi Odjel), a comparison of document parsers on complex tables and forms (Atomic Intelligence), a CroAI panel on AI agent security, Stjepan Picek (FER) on attacks against AI systems, and a debate on whether Croatia's criminalisation of certain AI uses is justified. A pre-event panel runs on 27 October. If you are deciding how to deploy agents safely, this is the most relevant local event this autumn.
Engineers in Poreč, the Chamber in Zagreb
The 18th Days of Electrical Engineers (Dani inženjera elektrotehnike) in Poreč on 15–16 October expected up to 1,400 participants, with a day devoted to AI, smart cities and digital twins. On 27 October, the Croatian Chamber of Economy's AI Academy (HGK AI akademija) holds an "Are you AI ready?" session on where to start with AI in business processes.
The open question: as Brussels starts enforcing, Croatia has still not publicly designated the national authority that will supervise the AI Act, as Netokracija has pointed out. For Croatian companies that means uncertainty about who to ask, not an exemption from the rules. Our advice: document your AI use now — inventory, risk class, human oversight — so the answer is ready whoever ends up asking.
The Week in Numbers
What to Watch
- Company responses to the AI Office: whether the Commission escalates from information requests to formal proceedings.
- Mistral Large 4 open weights: still due by the end of October.
- 27 October: Gemini 4 Argon reportedly switches on by default in Gemini Enterprise; HGK AI Academy session in Zagreb.
- 28–29 October: AI2Future in Zagreb.
- 9 November: the Commission's consultation on AI Act penalty procedures closes.
- 1 December: Gemini 4 Argon price doubles; standalone Microsoft 365 Copilot SKUs and metered billing by default for new CSP Copilot Business subscriptions.
- 2 December: AI Act machine-readable marking deadline for generative AI systems already on the market.
Agents in production need an identity, a trace and an owner
We help Croatian companies deploy Claude-based agents with least-privilege access, OpenTelemetry tracing, human checkpoints and AI Act documentation ready for audit. Claude Certified Architect, based in Zagreb.
Talk to us