Featuring:

While Workflows First and Trendwise AI have been quiet for a couple months, the AI industry has not.

Anthropic released Claude Opus 5 last Friday, and the coverage went where you would expect: benchmark wins, near-Fable intelligence at half the price, and a new default model for Claude Max.

One number underneath those headlines matters just as much. Opus 5 costs $5 per million input tokens and $25 per million output tokens, identical to Opus 4.8. Same price as its predecessor, different behavior.

All impressive. The detail that caught my attention was buried in Anthropic’s prompting guide.

If your prompt tells Claude to double-check its answer, run a final verification step, or use another agent to review its work, Anthropic now recommends removing those instructions.

Opus 5 already verifies and corrects its work. The old instructions can stack on top of behavior you are already paying for, adding time and tokens without improving the result.

That reads like a prompt-engineering footnote. It is really an operating problem.

A prompt library can become technical debt.

For the past few years, teams have treated a well-engineered prompt as a durable asset. Someone found phrasing that worked, saved it, shared it, and eventually wired it into a process that other people now depend on.

The model underneath that process has since changed, often more than once.

Instructions that once improved reliability may now duplicate behavior the model performs on its own. The prompt still runs. It still looks reassuringly sophisticated. The workflow simply becomes slower and more expensive, quietly, with nothing in the tooling announcing that it is time for maintenance.

The easy version of this advice is “delete the redundant lines.” That is only half the job.

Opus 5’s default responses run longer than earlier Opus models. It can expand the scope of a task on its own, and it delegates to subagents more readily. On a small task, that can multiply cost and completion time for work that never needed a team.

Some prompts need instructions removed. Others need a tighter boundary added.

Anthropic offers a particularly useful example from code review. A prompt that says “only report high-severity issues” or “be conservative” may now be followed so literally that the model reports less. Anthropic recommends asking for all findings, then filtering for severity in a separate pass.

That is worth sitting with. An instruction that improved an older workflow can degrade the new one. Technical debt does not merely accumulate. Sometimes it inverts.

Here is the pattern, super-simplified:

Before

After

The second version removes verification the model already performs, separates discovery from filtering, and adds the length boundary Opus 5 may need.

It is also a hypothesis. You still test it against your own work before turning it into a standard.

Affiliate disclosure: As an Amazon Associate I earn from qualifying purchases.

Microsoft published a useful companion story the day before Opus 5 launched.

Its AI team described specialized MAI models trained inside the product environments where they will actually work. One powers agentic coding in GitHub Copilot. Another started from the same checkpoint and was trained further inside an Excel reinforcement-learning environment.

In plain English, Microsoft took a coding model and taught it the tools, workflows and evaluation criteria of spreadsheet work.

Microsoft reports that its Excel model performs on par with GPT-5.6 on the most common tasks while costing less to run. Its MAI-Code-1-Flash model reportedly produces an approximately 10% higher code-acceptance rate than GPT-5.4 Mini and Claude Haiku 4.5 in VS Code while using about 10% fewer median tokens.

These are Microsoft’s own production results, just as the Opus 5 benchmarks come from Anthropic. The strategic direction matters more than the decimals.

The future AI stack will not place the largest general-purpose model in every slot. A smaller model trained for a specific tool, harness and evaluation set may produce a better result at lower cost.

That means model choice becomes part of workflow architecture rather than a permanent brand decision. You can replace one layer of the stack without rebuilding everything around it.

You still have to test the interfaces.

Your prompts, tools, permissions, handoffs and evaluation criteria were written around the behavior of the old model. A clean model swap can still create an unexpected workflow change.

Pick one prompt or AI-assisted process your team runs repeatedly.

Most Opus 4.8 prompts will continue to work. That is helpful for migration, although “it still runs” is a low standard for an operating process.

The prompt is only one small component. Treat it like one.

If your organization has a prompt library, when did anyone last audit it?

👉Vote today, see the results next edition!

📝 75% of this newsletter is written personally by our team.

🤖 The remaining 25% is supported by AI—for things like topic research and grammar checks.

🦾 Each piece of content takes a lot of effort from our team, so your feedback means a lot to us. Let us know what you think!

To make sure you never miss an update, please move this email from the Promotions or Spam tab to your Primary Inbox. This helps ensure you receive the latest AI news, tips, and tutorials—right where you’ll see them first.

Reply to this email with any specific feedback or interesting insights!

Thanks for reading!