In this guide · 8 sections
Short answer
Reasoning models spend hidden 'thinking' tokens before answering, billed as output tokens. They help on multi-step tasks with rules, such as quotes, analysis and code, and waste money on simple lookups and FAQ replies. Start at the lowest setting that passes your own test.
What a reasoning model actually does
A standard language model writes its answer word by word straight away. A reasoning model first produces internal reasoning: breaking the request down, trying approaches, checking its work. OpenAI's reasoning guide describes these as reasoning tokens used to plan, use tools, consider alternatives and recover from ambiguity before the visible reply.
You usually do not see this thinking. OpenAI says reasoning tokens are not visible through the API but still take up context window space and are billed as output tokens. Google's Gemini thinking guide says that when thinking is on, response pricing is the sum of output and thinking tokens. Anthropic reports thinking tokens as part of billed output in its usage data. If you are new to tokens, our plain-English LLM explainer covers the basics.
How each vendor controls thinking
All three major vendors now let you dial reasoning up or down per request. The names differ, but the idea is the same.
| Vendor | Setting | Values | Default on current models |
|---|---|---|---|
| OpenAI | reasoning effort (plus standard or pro mode) | Model-dependent: none, minimal, low, medium, high, xhigh, max | Medium for GPT-6.1 Sol, GPT-6 Sol and GPT-6 Luna |
| Anthropic | effort, plus adaptive or manual thinking | low to max, with xhigh on some models | Medium on Claude Opus 5.5 and Haiku 5.5; high on most others |
| thinking level | minimal, low, medium, high (varies by model) | Medium on Gemini 4 Argon and 3.8 Flash |
Anthropic's effort documentation notes that effort affects all output tokens, with or without thinking, and lower effort also means fewer tool calls. OpenAI says GPT-6 Astra does not accept 'none', so the cheapest setting is model-specific. Check the default before you assume a model is running in its cheapest mode.
What thinking does to cost and speed
Output tokens already cost about five times input tokens at the major vendors, as our API pricing comparison shows. Thinking adds more output. A simple illustration at a list price of US$10 per million output tokens: a 300-token reply costs US$0.003, but the same reply with 3,000 thinking tokens in front costs US$0.033, eleven times as much, and takes longer to start appearing.
Independent leaderboards show the same pattern. Artificial Analysis publishes the cost of running its test set for each model at each setting. We did not run these tests; the figures below are from its leaderboard on 21 October 2026.
| Model (setting) | Intelligence Index | Cost per index task |
|---|---|---|
| GPT-6 Luna (non-reasoning) | 18 | $0.01 |
| GPT-6 Luna (high) | 33 | $0.03 |
| Claude Haiku 5.5 (medium) | 34 | $0.05 |
| Claude Haiku 5.5 (max) | 43 | $0.21 |
| GPT-6.1 Sol (low) | 42 | $0.13 |
| GPT-6.1 Sol (high) | 50 | $0.32 |
Higher settings score better on hard tests, but the cost per task rises several-fold. Our explainer on AI benchmarks covers what these indexes do and don't measure.
When reasoning is worth it, and when it wastes money
Worth a higher setting
- Quotes that apply several pricing rules, discounts and exceptions
- Checking a contract or tender against a list of requirements
- Analysing a spreadsheet or reconciling figures
- Writing or reviewing code
- Multi-step agent tasks that call several tools
Keep it low or off
- Answering FAQs from your own website content
- Classifying emails or tagging tickets
- Extracting fields from a standard invoice
- Short rewrites, translations and summaries
- Live chat where customers expect a reply in seconds
A useful rule: if a careful junior employee could do the task without stopping to think, the model probably does not need to either. If they would need a pen and paper, reasoning tends to pay off.
Run your own 30-minute test
Published benchmarks are general; your tasks are specific. A short test on your own examples tells you which setting is good enough.
Collect 20 real examples
Recent customer questions, quote requests or documents, with personal data removed. Include a few awkward ones.
Write the correct answer for each
The right price, label or the points a good reply must contain.
Run three settings
The same model at low, medium and high reasoning. Record answers, total tokens and response time from the API usage data.
Score accuracy and cost
Count correct answers and multiply tokens by the vendor's price. Note the slowest response.
Pick the lowest setting that passes
Then route only the hard cases, such as complex quotes, to a higher setting.
Our guide to choosing an AI model per task by cost shows how to route different jobs to different models and settings.
What this means for a customer-facing chatbot
Website and WhatsApp assistants mostly answer questions from your own content. Speed matters more than deep reasoning, so a low setting on a mid-range or small model usually fits. Reasoning becomes useful when the assistant must apply rules, such as working out eligibility or a price from several inputs.
Do it yourself or get help?
If you call AI APIs yourself, set the reasoning level explicitly instead of relying on defaults, track thinking tokens in the usage data, and re-run your test set when vendors release or retire models.
Our AI Website Assistant (S$790) is trained on your website, FAQ and documents, captures leads, hands over to WhatsApp or email, and has guard-rails so it only answers about your business; AI usage fees are billed to you, usually a few dollars a month. IT Sorted is a Singapore-based team of developers with over 10 years of experience; you pay upfront, send the brief in writing and follow progress on your order page.
Frequently asked questions
It is a language model that produces internal reasoning steps before its final answer, planning the response, considering alternatives and checking its work. These hidden steps are measured in tokens and billed as output. Reasoning tends to improve results on complex, multi-step tasks but makes responses slower and more expensive.
Yes. OpenAI says reasoning tokens are billed as output tokens even though you don't see them. Google says Gemini response pricing is output plus thinking tokens. Anthropic reports thinking tokens within billed output. Because output is the priciest token type, thinking can multiply the cost of a reply.
Usually only at a low setting. Most customer questions are answered from your own content, where speed matters more than deep reasoning. Raise the setting only for requests that apply rules, such as complex quotes or eligibility checks, and test that the slower replies are actually more accurate.
Set it per request. OpenAI uses a reasoning effort parameter, Anthropic an effort setting with adaptive or manual thinking, and Google a thinking level. Defaults vary by model and several current models default to medium, so set the value explicitly rather than assuming the cheapest mode.
No. On simple tasks such as classification or short rewrites, extra thinking often adds cost and delay without improving results. Benchmarks show gains on hard problems, but your own test on 20 real examples is the reliable way to decide which setting your tasks need.