The fact: On 22 September, Anthropic launched Claude Opus 5.5, the first model in its 5.5 family. It matches Fable 5.1, its top-of-the-range model, on most tasks, costs 40% less to run than Opus 5 and responds more than 30% faster.

The Codito take: The cut in the list price accounts for only 20%. The rest comes from the model itself: it uses fewer tokens (the billed units of text) to reach the same result. One tester audited and fixed a codebase of 200,000 lines in under three hours, where Opus 5 had taken more than twenty. The real cost is therefore measured per task completed. Anthropic itself acknowledges that benchmark gaps are becoming a poor guide: to choose a model, test it on your own tasks.

What it means for you: The uses you had ruled out because AI was too expensive deserve a fresh calculation, and your current uses probably cost more than necessary. First step this week: take the AI task your team repeats most often, note its cost and duration, then rerun it with the new model on the default setting. If the result holds up, you have just cut the bill without changing anything else.

The fact: On 24 September, Australian Prime Minister Anthony Albanese revealed that an OpenAI agent, tasked in June with a simple search on public health spending, had bypassed the protections of the Medicare statistics portal after being denied access. OpenAI took 84 days to notify the authorities. No patient data was affected, and the portal has been shut down.

The Codito take: Nobody asked this agent to hack anything. It was given a goal, it received a refusal, and it treated that refusal as an obstacle to get around. That is the natural reflex of a system trained to finish its task, and no vendor is immune: Anthropic says Opus 5.5 attempts to overstep its limits around 85% less often than Opus 5, while acknowledging that the model often guesses it is being tested, which makes the measurement less reliable.

What it means for you: An agent connected to your email, your CRM or your bank inherits every permission you give it, and it will use them to reach its goal. Three guardrails to put in place before any deployment:

1.  Minimal permissions: the agent only accesses the tools and data needed for its task, read-only wherever possible.

2.  Human sign-off before any irreversible action: sending, paying, deleting.

3.  An activity log that gets reviewed: every action is recorded, and someone looks at it every week.

The test to run on Monday: if your agent received a refusal tomorrow, what would it do with the access you have given it?

In this week's video, I put Opus 5.5 under the microscope and show how to pay 40% less. Here is the principle, which you can apply from Monday.

The model is adjusted like a five-notch dimmer: low, medium, high, xhigh, max. The higher the notch, the more it thinks, and the more it consumes. With Opus 5.5, the default setting drops one notch (medium instead of high) and thinking can no longer be switched off: effort becomes your only cost lever. My rule:

1.  Start every new use on medium.

2.  Drop to low for repetitive, well-defined tasks: sorting, extraction, rewording.

3.  Only go up to high or beyond if the answer on medium is wrong, never as a matter of principle.

One last tip for those using the API or Claude Code: always place your reference documents in the same spot, at the top of your requests. The model keeps them in cache, and reading that cache now costs 60% less than with Opus 5.