Reasoning models in 2026: when extended thinking actually pays off

Reasoning models stopped being a novelty about a year ago. They’re now just part of the catalog — Opus 4.7 has extended thinking, GPT-5.4 Mini is reasoning-by-default, DeepSeek R1 is open and cheap. So the interesting question isn’t do they work, it’s when is the extra latency worth it.

We’ve been tracking that question internally for a few months. Here’s what the data and the gut both say.

What “reasoning” actually buys you

When we say a model is reasoning, we mean it’s allowed to spend more tokens before producing an answer — a private chain-of-thought it doesn’t usually show you. More tokens means more chances to catch a mistake, more chances to consider an alternative, more compute aimed at the problem.

That extra compute helps a lot for some problems and barely at all for others. The pattern is pretty consistent:

• Multi-step problems with intermediate checks (math proofs, code refactors that touch several files, planning tasks): reasoning wins.
• Single-step problems where the right answer is one inference hop away (summarize this email, rewrite this sentence, what’s the capital of Mongolia): reasoning wastes your time and your money.
• Creative problems where there isn’t a verifiable right answer (write a poem, draft an opinion piece, brainstorm names): reasoning often makes things worse, not better. Extra deliberation can flatten the voice.

The latency tradeoff is real

A non-reasoning model gets back to you in 2–8 seconds. A reasoning model takes 15–90 seconds, sometimes more. That’s a 5–15x slowdown that you eat on every turn.

For the 20% of prompts where reasoning actually changes the answer, it’s worth it — those prompts were taking you four follow-up messages anyway, and you’d rather spend the wall-clock once than four times. For the 80% where it doesn’t, you’re now waiting a minute for an answer you would have accepted in five seconds.

This is the part the leaderboards don’t capture. “Model X scores 4 points higher on benchmark Y” is true and also boring if it took it eight times as long to get there. The right metric is per-second utility, and on that axis the picture is much closer.

How we surface it

In the studio, you don’t pick “reasoning” or “non-reasoning” — you pick a model, and reasoning is a toggle on the ones that support it. The toggle is visible (cmd+shift+R), and the cost preview tells you what the turn will cost in credits before you send.

Default behavior:

• Opus 4.7, GPT-5.5: thinking off by default. Toggle on for hard problems.
• GPT-5.4 Mini, R1: thinking on by default. These models are the reasoning version.
• Haiku 4.5, Gemini Flash, Nano: no thinking mode, by design.

If you have memory turned on, the studio learns when you tend to flip the toggle and suggests it earlier next time you start a thread that looks similar.

The case for keeping a fast model alongside

Even if you only use reasoning models, you probably want a fast non- reasoning model on the same shortcut. Cmd+Shift+M and one keystroke should drop you from Opus thinking down to Haiku 4.5 for the next message. We tested without that shortcut for a week and the friction shows up immediately — you stop iterating, because each iteration costs you a minute.

The most efficient pattern we’ve found, by a wide margin, is:

1. Start with a reasoning model on the hard problem.
2. Drop to a fast model for follow-ups, edits, and rephrases.
3. Pop back up to reasoning only when you change direction substantively.

What we don’t think

Two things we hear a lot that we don’t think are true:

“Reasoning models will replace non-reasoning models.” They won’t. The latency floor is fundamental — you can’t make the chain of thought shorter without making it dumber. There will be a fast tier for as long as there is a hard tier.

“Reasoning models are smarter at everything.” They’re not. They’re smarter at problems with verifiable intermediate steps, and roughly the same at everything else. The benchmark gap on creative writing in 2026 is basically zero.

So: reason when the problem is hard. Don’t when it isn’t. The studio makes the toggle one keystroke for a reason.