But surely there IS a best model for general use?

There's a best on average, which changes every six months and is currently a coin-flip between GPT-5 and Claude 4.7. The catch is that "best on average" loses to "right for this task" on almost every individual task. Optimizing for the average leaves value on the table on the specifics.

Isn't task-by-task model selection too much overhead?

It is, if you have to switch tools to switch models. It isn't, if your tool surfaces the choice as a one-click pick. That's the whole reason multi-model platforms with per-message model pickers exist.

Will this stop mattering as models converge?

Models are diverging more, not less. As capabilities specialize (multimodal, agentic, reasoning, fast-and-cheap), the differences between flagship models grow rather than shrink. The case for picking the right one becomes stronger over time.

Why "the right model" matters more than "the best model"

"Which AI is best?" is the most-asked question in this category, and it has the wrong shape. The answer changes every six months, depends entirely on the task, and matters less than people think — because the model on top of the leaderboard is rarely the one you should be using for the thing in front of you right now. This essay is the case for replacing "best" with "right" — and what changes about your relationship to AI tools when you do.

Why "best" has the wrong shape

"Best" implies a total ordering. A is better than B which is better than C. If A is best, you use A and you've optimized.

Models aren't shaped that way. Claude 4.7 is better than GPT-5 at long-document analysis and structured output. GPT-5 is better than Claude 4.7 at marketing copy and agentic loops. Gemini 2.5 Pro is better than both at vision-text and Google Workspace integration. There is no single A that dominates B and C across all categories — the rankings shuffle by task.

So "use the best one" gives you a model that's good-but-not-great for whatever specific thing you're doing. You're optimizing for the average performance and getting average performance.

The "right" frame

"Right" doesn't imply a total ordering. It implies a match between a task and a tool. The right model for this question is the one whose strengths align with what this question rewards.

Two questions get you there in practice:

What kind of task is this? (Long-doc, code, vision, structured output, agentic, marketing, etc.)
Which model is strongest on that task?

You don't have to memorize the answers — we wrote a decision tree and a lookup table. The point is to be in the habit of asking the questions, not to pretend the answer is one model for life.

What changes when you adopt "right"

You stop having brand loyalty to a model. Which means you stop missing better answers because you didn't bother to try.

You start noticing the gap between models on specific tasks. Which means you develop calibrated intuition about when each one shines.

You stop arguing about "which AI is best" in the abstract, because the question feels meaningless. Which is good — it WAS meaningless.

You set up your tooling so the switch is cheap. A single multi-model platform with a per-message model picker. One shared instruction set. Branching so you can keep two model answers alive to compare.

The cost

You give up the simplicity of "I'm a Claude person" / "I'm a ChatGPT person". The identity is real — people build communities around their preferred tool, swap tips, develop a sense of expertise that's specific to one product. Switching frequently dilutes that identity.

For most readers, the trade is worth it. Better answers on every task, in exchange for one less in-group identifier.

What this means for tool builders

Tools that lock you into one model (Claude.ai, ChatGPT, Gemini app) are betting that you'll keep using their model out of loyalty even when it's wrong for your task. Multi-model tools (oran.chat, Poe, TypingMind, OpenRouter — see our comparison) are betting that the right-model-for-the-task posture wins over time.

We obviously believe the second bet wins, because that's the product we built. But you don't have to take our word for it — try the right-not-best posture for a week with whatever tool you currently use that supports model switching. Notice when the switch helps. Notice what stays the same.

Connecting back

The right-model posture is the practical complement to the thinking is yours, the models do the typing. When the thinking is yours, the choice of "right model" is part of the thinking. The tool's job is to make the choice easy to act on once you've made it.

More on model choice and AI agency in Essays. Or try oran.chat free — built around the per-message model picker this essay argues for.