This turns Claude into its own judge. Instead of trusting a stranger's benchmark, you give two different Claude models the identical task, then a third chat scores them side by side so you can see which model actually wins for the kind of work you do.
As of July 2026, here is the real lineup straight from Anthropic's model docs. Pick one "smart" model and one "fast" model to compare.
Open the Claude app (or claude.ai), start a new chat, and use the model picker to select your smart model. Start a second new chat and pick your fast model. Keep them separate so nothing bleeds between them. On mobile the model picker sits at the top of the chat, and you can switch models per chat.
Use ONE real task you actually care about, not a toy. Paste the identical prompt into both chats. Same words, same context, same attachments. If you change even one line, the test is contaminated.
Rewrite this into a high-converting landing page hero section for [what you sell], for [who it's for]. Give me a headline, a one-line subhead, and one CTA button label. Keep it specific and benefit-led, no fluff, no em dashes.
Swap the landing page for whatever your real work is: a tricky client email, a bit of code, a caption, a cold DM, a SQL query. The point is to test the model on YOUR job, not a generic demo.
This is the part people skip. Open a third chat, paste both answers in, and let Claude judge them against your actual goal.
Here are two outputs for the same task from two different models. My goal is [what a great result looks like: converts cold traffic / sounds like me / passes to a junior dev with no edits]. Grade each one out of 10 on that goal, tell me exactly where each one wins and loses, and pick the one you'd actually ship. Be blunt, don't flatter either.
Because both outputs came from the same task, the grading is apples to apples. You will often find the "worse" model wins on your specific job, which is the whole point.
For a fast, fun gut-check, ask each model to build the same little thing and just eyeball which looks better.
Build a single-file HTML page with a realistic animated lava lamp: blobs that slowly rise, merge, split, and sink with soft physics and a warm glow. No libraries, just HTML, CSS and JS. Make it look genuinely believable.
Run it in both models, open both previews, and compare the physics. Animation and UI quality is one of the fastest ways to feel the gap between a smart model and a fast one.
This is the real trade-off the reel flags. The most capable models cost more per token (Opus 4.8 is $5/$25 vs Sonnet 5 at $3/$15, and Fable 5 is $10/$50), and on the Max plan there is even a separate weekly usage cap for Sonnet vs an all-models cap, because heavier models burn through your allowance faster.
Consistency check: Run my task 3 times on the same model. Show me all 3 outputs and tell me how much they vary. I want to know if this model is reliable for me or a coin flip.
Cost vs quality verdict: Here are outputs from Opus 4.8 and Sonnet 5 for the same task. The smart one costs more usage. Tell me honestly, is the quality jump worth the extra cost for THIS kind of task, or is the cheaper model good enough?
Build my test rubric: I keep testing new models on my work. Write me a reusable scoring rubric for [my main use case] with 5 criteria and a 0-10 scale, so every future model gets judged the exact same way.