The last three months have been busy for AI. Anthropic and Google have released new models, OpenAI and Anthropic have announced the retirement of older ones, and a clear pattern runs through it all: new models are built to work for longer without checking in, some are getting cheaper, and the models underneath existing agents are being retired faster than many businesses expect.
If your business runs AI agents, or is about to, those changes affect how you build, budget and manage them. Here’s what changed, and what to do about it.
What’s changed
Agents can now run for hours. Anthropic released Claude Opus 5.5 on 22 September 2026. One customer reported it stayed on task for over 18 hours during an overnight engineering run. Anthropic also says the model is much less likely than recent models to take hard-to-reverse actions, and more resistant to prompt injection. Those are the vendor’s own results. It’s cheaper too: US$4 per million input tokens and US$20 per million output tokens, which Anthropic says is 20% less than Opus 5.
Models are being trained to pause. Google released Gemini 4 Argon on 13 October 2026, and says it trained the model to pause for human confirmation before taking irreversible steps. Its launch price of US$2 per million input tokens and US$10 per million output tokens is introductory, and doubles after 30 November 2026.
Older models are being switched off. On 14 October 2026, OpenAI announced that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini will be removed from its API on 14 January 2027. Anthropic announced on 30 September 2026 that Claude Sonnet 4.5 is scheduled to retire on 30 November 2026. Any agent calling those models through OpenAI’s or Anthropic’s own API stops working on those dates unless it’s moved. If you use them through Azure, AWS or Google Cloud, check that platform’s schedule, which can differ.
Agents are becoming identities. Since May 2026, Microsoft has given every new Copilot Studio agent its own Microsoft Entra Agent ID automatically, and admins can apply Conditional Access to the connectors it uses. For New Zealand businesses running on Microsoft 365, that means agents can be governed much like staff accounts.
What it means for how you run agents
Longer runs need designed checkpoints. A model that can work for 18 hours can take a lot of actions before anyone looks. A model that’s trained to pause is welcome, but it isn’t a control. The approvals that matter, before money moves, records change or customers are contacted, need to be enforced by the workflow around the agent. We set out how in our article on agent flows and open-ended agents.
Every agent now has an expiry date. The model under an agent is a dependency, like any other software, and dependencies get retired. Moving to a newer model usually improves results, but it can also change behaviour: answers get longer or shorter, tools get called differently, and edge cases that used to work may not. That makes re-testing a regular maintenance cost, not a one-off.
Falling prices change the maths. Jobs that were too expensive to automate a year ago may now be worth it. But introductory prices end, and agents that run for hours can use far more than a quick chat. Re-run the numbers for each agent, and set spending limits so a runaway task can’t run up a surprise bill.
Agents need managing like staff accounts. An agent with its own identity can be given least-privilege access, covered by Conditional Access, reviewed and switched off. That only helps if someone does it. Our article on MCP, permissions and roles covers how to scope what an agent can reach.
It’s worth revisiting what didn’t work. If you tried an agent six months ago and it lost track halfway through, the same job may now be within reach. Test it on your own work rather than relying on vendor claims, but don’t assume the old answer still holds.
What managers should do differently
The biggest change isn’t technical. It’s ownership. Each agent should have a named owner who knows which model it runs on, what it can access, what it costs and how it’s performing. When a vendor announces a retirement or a better model, that person decides when to test and switch.
Treat model changes like software updates. Run the new model alongside the old one on real examples, compare the results, then switch. Keep track of a few simple measures for each agent: cost per task, error rate, and how often a person has to step in. Those numbers tell you whether an upgrade actually helped.
Risks and costs to keep in view
- Bigger mistakes. The longer an agent runs unsupervised, the more it can get wrong before anyone notices. Prompt injection, where hidden instructions in a document or email steer the agent, matters more as autonomy grows. See our explainer on prompt injection.
- Vendor claims aren’t your results. Benchmarks and customer quotes describe someone else’s tasks. Only your own test cases tell you how a model handles your work.
- Lock-in. Build agents so the model can be swapped without rebuilding everything, because you will need to swap it.
- Hidden running costs. Long agent runs, retries and large documents add up. Watch usage for the first few weeks of any change.
Try this in the next 30 days
- List every agent and AI automation you run, with the model and version each one uses.
- Check each model’s retirement date on the vendor’s deprecations page, and diary anything due in the next six months.
- Build a small test set for each agent: 20 to 30 real examples with the result you’d expect.
- Run the test set on a newer model and compare quality, speed and cost before switching.
- Set spending limits on every AI account and agent.
- Review each agent’s identity and permissions, and remove anything it doesn’t need.
- Mark the steps where a person must approve, and confirm the workflow enforces them, not just the agent’s instructions.
Where to start
Start with the inventory. Most businesses are surprised by how many AI tools and automations they already run, and which models sit underneath them. If you’d like help testing, upgrading or tightening the agents you have, see our AI and agents work or get in touch.