Fill in the knobs: OpenAI effort levels vs Claude extended thinking budgets
OpenAI's o-series exposes a reasoning effort dial (low or medium or high); Anthropic's Claude extended thinking mode takes an explicit thinking-token budget.
Imagine two tutors who can both think hard before answering. One tutor lets you say 'try a little' or 'try a lot' or 'try really hard' and figures out the rest. The other tutor asks you exactly how many minutes you want them to think before they speak. Both end up doing the same kind of careful work, but the way you tell them to do it is different. OpenAI built the first style into o1 and o3 with a parameter called effort. Anthropic built the second style into Claude with a mode called extended thinking, and the unit you give it is tokens, not minutes. Knowing which knob you are turning matters because the price and the latency depend on it.
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.
Reasoning model APIs in 2026 have converged on two control surfaces: a qualitative effort tier (OpenAI o-series) and an explicit token budget (Claude extended thinking). Gemini Thinking sits between with a hybrid. Knowing which knob you are turning and how it bills is foundational for any production reasoning workload, and confusion between the two patterns is one of the most common interview slip-ups.
This explanation walks each control surface in detail, the billing model behind each, the production patterns that emerge for each, and the cross-vendor abstractions that production code tends to settle on.
OpenAI reasoning_effort: qualitative tiers
The reasoning_effort parameter on o-series endpoints takes one of three values: low, medium, or high. The parameter is request-scoped, so each call can pick its own tier. The model maps the tier to an internal token budget you do not directly observe; what you see is the resulting reasoning_tokens count in the usage object after the response completes.
The tier abstraction is intentional. OpenAI's view is that the model is better positioned to decide exact token spend than the caller, so the API exposes the trade-off as a service-level decision rather than a numeric knob. In practice, low effort produces reasoning traces in the hundreds of tokens, medium in the low thousands, and high in the tens of thousands, though the exact ranges vary across model versions and even across problem difficulty within a tier.
Billing is straightforward: reasoning tokens are billed at the output-token rate. Your final invoice line for a request includes input tokens, reasoning tokens, and completion tokens, with the latter two charged at the same per-token rate. A high-effort request on a hard problem can spend most of its bill on reasoning tokens that the end user never sees.
The production pattern is per-class tier selection: build a router that picks low for easy traffic, medium for the bulk, high for the genuinely hard slice. Tier flipping is the main lever; you cannot fine-tune within tier.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
| Aspect | OpenAI reasoning_effort | Anthropic max_thinking_tokens |
|---|---|---|
| Type | Discrete enum | Continuous integer |
| Values | minimal / low / medium / high | 0 to ~64k tokens |
| User mental model | Tier name | Explicit token budget |
| Reported in usage | reasoning_tokens count | output_tokens (aggregated) |
| Best for | Product teams not thinking in tokens | Cost-engineering teams wanting hard caps |
Real products, models, and research that use this idea.
- OpenAI o3 in ChatGPT and via the API exposes reasoning_effort as a request parameter; production tools like Cursor pick the tier per query.
- Claude Opus 4.7 extended thinking is used in agent frameworks like Claude Code with budget_tokens commonly set between 4,000 and 32,000.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you calibrate Claude budget_tokens for a new task class from scratch?
Run a sweep: 2k, 8k, 16k, 32k budgets on a representative eval slice; chart accuracy versus budget; pick the budget at the elbow where additional tokens stop earning quality.
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.
Mixing the two vocabularies; the OpenAI parameter is a qualitative tier and the Claude knob is a numeric token cap, and the API shapes look different.
60 second bullets to scan on the way to the call.
The OpenAI o-series parameter name and its three values
The Claude extended thinking parameter name and unit
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.