Google's Gemini 4 Carbon raises model security and agent standards

Small teams should expect extra upfront work: re-evaluate model choices for security, add cost tracking, and build platform guardrails so agents don’t drift into compliance or budget risk.

Google launches Gemini 4 Carbon

Google released Gemini 4 Carbon, a new frontier model Google says delivers gains in coding, agentic work and cybersecurity and which it cites as topping the Artificial Analysis Intelligence Index (v5.0). The company framed Carbon as an advance on prior Gemini releases and emphasised improvements on security-related tasks. "Significant gains" was the phrase used in reporting to describe those improvements.

Why it matters: If you run production agents or security-sensitive tooling, you now have a new model option that Google markets as safer for those use cases; expect to re-test threat models, update model-selection policies and validate Carbon’s behaviour before replacing existing models.

MIT Technology Review: enterprise context limits AI value

MIT Technology Review Insights reports that while most organisations use AI, only 16% say they extract transformative value; 40% report moderate gains, 21% meaningful gains and 23% minimal or no returns. The research also finds many companies still piloting AI (38%) or using co-pilot style assistants (35%), with 12% reporting agents in production.

Why it matters: If your team is deploying tools without mapping how they interact with each other, you’re likely leaving ROI on the table; prioritise integration, governance and outcome metrics rather than adding more point solutions.

AWS details platform needs with Bedrock AgentCore

AWS published guidance for building agent platforms on Amazon Bedrock AgentCore, focusing on platform-level controls: approved models and tools, short-lived workload identity, isolated runtimes, automated evaluation, rollback, telemetry and cost attribution. The post frames the work as a Platform product that must prevent control drift as many Workstream teams build agents.

Why it matters: Operators who expect to scale agents across teams need to budget for platform engineering, telemetry and AgentOps rather than treating models as the main cost or blocker; without those controls, ownership and audits become hard.

Anthropic reveals tools to show AI spending

Anthropic launched tools intended to show customers what their spending on models actually buys, an effort reported alongside the company’s planned IPO preparations. The public write‑ups present the tools as a way to make model usage and spend more visible to customers.

Why it matters: If you buy model access by consumption, you should plan to demand comparable visibility from your vendors; these tools let purchasing teams tie model use to business outcomes and reduce bill surprises.

OpenAI and ServiceNow expand workflow integration

OpenAI and ServiceNow expanded their partnership to integrate OpenAI models and agents into ServiceNow and to surface ServiceNow workflows inside ChatGPT and Codex. The announcement presents tighter two‑way integration between ServiceNow automation and OpenAI’s model stack.

Why it matters: For organisations already using ServiceNow, this reduces integration work for automating incident, HR or procurement workflows, but expect to update governance and licensing agreements to cover agentised workflows.

Legora builds an agentic OS for legal work with OpenAI

Legora is using OpenAI models, ChatGPT and Codex to create an agentic operating system tailored to legal workflows, increasing automation of contract analysis and routine legal tasks. The write‑up positions Legora as a workflow-specific agent platform built on OpenAI technology.

Why it matters: Legal teams looking to reduce routine review time can pilot workflow‑specific agents, but must attach firm controls for data privacy, versioning and attorney oversight before deploying them on live matters.

What we don't know

• Exact pricing, quotas or licensing terms for Gemini 4 Carbon and whether Google will position it differently from other Gemini tiers.
• The MIT poll’s sample size, industry mix and methodology that produced the percentages reported.
• Which models and third‑party tools Bedrock AgentCore will formally approve and how AWS will expose cost attribution across accounts.
• The exact metrics Anthropic’s spending tools surface, and how they handle customer data and privacy.
• Contractual and deployment details for the expanded OpenAI–ServiceNow integration, including enterprise licensing and data residency.
• The data sources, training scope and compliance controls Legora uses when analysing live contracts.

What to do next

1. Run a two‑week inventory: list active models, agents, owners, telemetry endpoints and cost centres so you can spot duplication and assign accountability (aligns with AWS and MIT findings).
2. Ask your vendors for concrete cost and privacy reports: request the same visibility Anthropic announced and demand sample invoices or usage dashboards before scaling consumption.
3. If you rely on agents for regulated work, schedule a controlled pilot that includes security testing, rollback procedures and compliance sign‑off before full deployment (apply Gemini 4 Carbon to a non‑critical task first).