Interleaved thinking lets the model reason, call a tool, then keep reasoning over the result inside the same turn, replacing plan then act with think act think.
Imagine you are writing a math homework problem that needs today's gas price. Without interleaving, you have to plan the whole solution first, then go look up the gas price, and only then start over with the number. With interleaving, you can write a bit of the solution, pause to look up the price, see the actual number, and continue writing using the real number. The second way is obviously better because you do not commit to guesses you could have just looked up. Interleaved thinking gives a reasoning model that same flexibility when it uses tools.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Interleaved thinking is one of the smaller-sounding reasoning-model features that turns out to matter a lot in practice. It removes an artificial structural constraint, the requirement that all thinking precede tool use, which had been forcing reasoning models into plan then act patterns that are demonstrably worse for live-data workflows. The question's correct answer is the precise statement of what interleaving changes; the three wrong answers each represent a real misunderstanding worth unpacking.
This explanation walks the plan then act failure mode in detail, the structural change interleaving introduces, why latency and billing are unchanged, the practical consequences for agent design, and the new failure modes interleaving creates that engineers should design against.
The plan then act failure mode
A reasoning model without interleaving has a strict order of operations per turn. The model produces a contiguous thinking block, the thinking block ends, the model emits its final response (which may include tool calls), and the turn closes. Tool results arrive in the next turn as part of the new input, and the model has to re-establish context before reasoning about them.
This structure creates a specific failure mode: the model is forced to plan its entire approach before it has access to any tool result. If the plan depends on a value the model could have fetched, the model has to either guess and proceed, or break the workflow into multiple turns. Both options are bad.
Guessing produces the classic estimation failures: the model commits to a price, an API shape, a date, or a count that turns out to be wrong, and the downstream reasoning is built on the wrong value. The model often does not catch this until much later, if at all.
Multi-turn splitting fragments the chain-of-thought. Each new turn starts fresh; the model has to re-read the prior thinking and re-establish its reasoning state from scratch. Context length grows quickly, coherence drops, and the workflow becomes harder to debug.
The defect is structural, not capability-based. The model is capable of reasoning about live data; it is just not allowed to do so in the middle of a thinking block. Interleaving removes the prohibition.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- Anthropic Claude Opus 4.7 and Sonnet 4.6 with the interleaved-thinking-2025-05-14 beta header expose this pattern explicitly.
- OpenAI o-series in the Responses API supports reasoning around tool calls in a single response, with similar structural benefits.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow does interleaved thinking interact with extended thinking budgets?
The thinking budget is consumed across all thinking blocks in the turn, not per block; interleaved patterns can run out of budget mid-derivation if not sized for the workflow's expected tool and think cycles.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Confusing interleaving with parallelism or with free thinking tokens; the change is about where deliberation happens, not about latency or cost.
60 second bullets to scan on the way to the call.
What plan then act looks like without interleaving
What think act think looks like with interleaving
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.