Control Reasoning Effort, from ChatGPT to the API

Dialing in Reflection for Every Task

Lessons 1 through 3 gave you the map of settings and OpenAI’s guidelines. This lesson moves to the controls: how to adjust effort as you work, manage its cost, and automate it within an application. The principle remains unchanged: reasoning level is a trade-off, not a quality slider. It simultaneously affects four variables: the number of reasoning tokens, response time, depth of analysis, and quota consumption or cost.

In ChatGPT: A Setting Per Message

In Chat, the Thinking slider for GPT-6 adjusts per message: you can stay on Instant to rephrase, switch to High for analysis, then dial it back down. Higher levels draw from quotas specific to your plan; when you hit a limit, ChatGPT may continue with another available model and, where possible, indicates when the quota resets. In Work and Codex, the power slider plays the same role, ranging from Faster to Smarter, with Max and Ultra held in reserve. Best practice in both cases: start with the default setting and only increase it if the result isn’t sufficient.

In the API: Five Levers

For developers, five parameters in the Responses API put you in control of reasoning:

• reasoning.effort sets the level, from none to max depending on the model: GPT-6 Astra and GPT-6.1 Sol start at low, while GPT-6.1 Terra and GPT-6 Luna accept none.
• reasoning.mode, standard by default or pro, makes the model work harder without changing the effort level; this extra work is billed at the model’s standard rate but consumes more tokens.
• max_output_tokens caps the total number of generated tokens, including reasoning. Set it too low, and you get incomplete responses; OpenAI recommends reserving at least 25,000 tokens when you start experimenting.
• reasoning.summary requests a summary of the reasoning process, since raw reasoning is never exposed.
• configuration_update, an element inserted between two responses, changes the effort level mid-conversation on compatible GPT-6 models without touching the initial request settings: the prompt prefix remains intact, and the prompt cache continues to serve.

Here is what that last element looks like, inserted before the user’s next message:

{
  "type": "configuration_update",
  "reasoning": { "effort": "high" }
}

A conversation started at low thus shifts to high for the next response and those following, until another update occurs.

The Impact on Costs

Reasoning tokens are billed as output tokens, and their count grows quickly with the level: from a few hundred to tens of thousands for a difficult problem. Model choice matters just as much: per output token, GPT-6 Luna costs twenty times less than GPT-6.1 Sol, and GPT-6 Astra costs five times more. In ChatGPT, cost translates into quotas: the number of demanding requests possible per period.

Three Optimization Strategies

The first is escalation: do not start at the top. Begin with the default level, medium in the API, or lower for a trivial task; if the response is incomplete or incorrect, move up one notch; reserve xhigh, max, pro mode, Max, and Ultra for problems that justify them. In an application, configuration_update makes this escalation possible within a single conversation.

The second is task-based triage, which you should make a habit: rephrasing and classification at the lowest level, drafting and simple analyses at low, planning and judgment at medium, debugging and difficult problems at high, long-haul research at xhigh, and extreme cases at max.

The third is to pair the right model with the right level. GPT-6.1 Sol at high or above, possibly in pro mode, handles demanding problems; GPT-6 Astra remains for the most difficult; GPT-6.1 Terra handles writing, image, and application orchestration; GPT-6 Luna, at low or medium, absorbs volume at lower cost.

Your Turn

Find the point of diminishing returns. Ask the following question three times, at three levels (Instant, Medium, then High in Chat, or low, medium, then high in the API), and compare the depth of analysis and, in the API, the reasoning tokens consumed.

“A company has 5 projects. Project A: €100k, ROI 3.2x, medium risk. Project B: €250k, ROI 1.8x, low risk. Project C: €50k, ROI 5.1x, high risk. Project D: €150k, ROI 2.4x, medium risk. Project E: €75k, ROI 4.0x, high risk. The total budget is €400k. Which projects should be funded to maximize risk-adjusted ROI? Explicitly state the adjustment rule used.”

Identify the level at which going higher adds little value: that is your setting for this type of question.

Official sources: Reasoning models (OpenAI API: effort, pro mode, max_output_tokens, summaries, configuration_update, billing) · Models (accepted efforts and pricing per model) · GPT-6 and other models in ChatGPT (levels, quotas, and fallback in Chat)

Key Takeaways

• Effort varies tokens, latency, depth, and cost together: it is a trade-off
• In ChatGPT, the level adjusts per message, within plan quotas
• In the API: reasoning.effort, reasoning.mode, max_output_tokens, reasoning.summary, and configuration_update
• Escalation, task triage, and the right model-level pairing are the three levers for savings
• Changing effort mid-conversation with configuration_update preserves the prompt cache