We spent this week running Claude's newest models against each other inside our own build environment, on real client-shaped tasks: a bug fix in a live codebase, a landing page brief, a short announcement video. The result was not surprising once we looked closely. The model picked mattered less than whether we told it what "done" actually meant. Managing AI agents, not shopping for the best model, is the actual skill for 2026.
What we tested this quarter
We ran the same prompt, the same codebase, and the same brief through two tiers from Anthropic's current model lineup, Opus 5 and Fable 5, inside Claude Code. Same task, same starting files, no extra hints for either model. This is close to how we scope every automation build for a client: one clear brief, one clear codebase, one clear pass or fail bar.
The pattern held across every task we ran. On a bug fix with a written expected behavior and a test to check against, one model passed every check and the other passed roughly half. On a task with no test to check against, just an open brief to "make it look good," both models produced usable but different work, and the difference came down to taste more than correctness. Cost and time swapped winners run to run too: sometimes the newer model finished faster and cheaper, sometimes the older flagship did. Anthropic's own pricing page lists a flat per-token rate for each tier, but that list price does not predict which task actually costs less once an agent needs several passes to finish the job, which is exactly what we saw run to run. Neither model won on every axis, and neither one deserved to be the default without a task-specific reason.
Where the model actually mattered
| Task type | What decided the winner | What we do about it |
|---|---|---|
| Bug fix with a test suite | Whether the model kept iterating until the test passed, not which model wrote the first draft | Always attach a pass/fail check before the agent starts |
| Open-ended design or copy brief | Taste and house style, which no model gets right without our input | Feed the agent our actual brand and past examples, not a blank prompt |
| Cost and time per task | Varied by task, no single model won consistently | Log cost per task so routing decisions are based on our own data, not a benchmark chart |
We also ran a fully open-ended visual task through both models with no test to grade against: build a diagram explaining a technical concept to a new hire with zero context on our business supplied. Both came back with a workable diagram, laid out differently, and one was noticeably more organized while the other packed in more detail. Neither was wrong. That is the tell that a task is a taste call, not a correctness call, and taste calls are exactly where feeding the agent our own brand voice, past examples, and house style matters more than which model drew the boxes.
Verification beats model choice
Anthropic's own engineering team has written about this directly. Their guide on building effective agents makes the case that the biggest gains come from giving an agent a clear way to check its own work and a stopping condition, not from swapping in a bigger model. That matched what we saw. The run with a test suite attached kept correcting itself until it passed. The run with no way to check its own output just stopped once it felt finished, right or wrong.
This is the same principle we build into every AI Operations Agent we ship. An agent that drafts a follow-up email is not useful on its own. An agent that drafts the email, checks it against a list of required fields, and flags anything missing before it goes out is a system we can actually trust unattended. The gap between those two is the entire gap between a demo and something that runs 24/7 without a human re-checking every output.
The real skill for 2026 is managing agents, not chatting with them
Picking a favorite model is a habit left over from when there was one chatbot window and one conversation. Running a business on AI agents in 2026 looks more like managing a small team: you assign a task, you define what a correct answer looks like before the work starts, and you check the output against that bar. The operators getting real hours back are the ones treating this as a management skill, not a subscription choice.
What managing an agent actually looks like day to day
- Write the brief down before the agent starts, the same way you would brief a new hire, not a search bar.
- Define the pass or fail check first: a test that runs, a checklist a human scans, a number that has to match.
- Run two models side by side on anything ambiguous or high stakes, and let the actual output decide, not a benchmark chart.
- Log the cost and time per task so you know which jobs are worth a bigger model and which ones a smaller one handles fine.
- Put a second check on the first pass, whether that is another agent or a person, before anything ships to a customer.
None of this requires picking a side between models. It requires building the habit of defining done before the work starts, which is exactly what most teams skip when they first start experimenting with agents.
How you instruct the agent matters as much as which model runs it. A one-line prompt with no context produces a generic answer no matter which model reads it. A brief that includes the actual document, the actual past example, and the actual constraint produces a usable answer from almost any model tier. We have started treating prompt and context quality as its own skill our team practices, separate from the model debate entirely, because a well-briefed smaller model regularly beats a poorly-briefed larger one on the tasks we run day to day.
What we are building against this quarter
Every system we ship for a client gets a verification step built in, not bolted on after something breaks. That is true whether it is a lead-response agent, a voice agent answering the phone, or a back-office workflow pulling data between tools. Our voice agent builds are checked against real call transcripts before they go live, not just a script read-through, because a voice agent that mishears a name is worse than no agent at all.
The economics back this up too. Once a system is live and verified, it runs 24/7 without needing a bigger model to cover for a missing check, and the client owns 100% of the code and files behind it, so there is no dependency on us to keep it running. Most of the businesses we start with are already leaking somewhere close to 10 hours a week in the manual version of a task we end up automating, and a working first system is usually live within days to weeks once the brief and the pass/fail bar are actually clear. If you want a fast read on where that leak is sitting in your own business before committing to a build, our revenue leak heatmap is a free way to see it mapped out.
We start most engagements with the AI Concierge assessment: a fixed scope, 999 euro engagement that maps the single most expensive manual task in your business, defines what a correct output looks like, and prices out the build. That fee is credited to the build if you move forward, so the diagnostic itself does not cost extra once you commit. You can see how this plays out end to end in our case studies, and more of our thinking on model routing and agent design is on the blog.
What to do this quarter
Stop asking which model is best. Pick the task you would hand to an agent this week, write down what a correct output looks like before you start, and only then decide which model gets the job. Run two models on anything you are not sure about and let the actual output settle it. The businesses seeing real hours back this year are not the ones with the best model subscription. They are the ones that stopped chatting with AI and started managing it like a team member with a real task, a real deadline, and a real way to check the work.