Skip to main content
EN · FR Start for free
Menu
Lesson 1 of 12 · OpenAI Reasoning: Reflection and Effort Levels, from ChatGPT to API

Which model and reasoning level for which use case

Lesson 1 of 12 Updated November 5, 2026 Intermediate
On this page
  1. Reasoning is no longer a separate model category
  2. In Chat: GPT-6 and the Thinking slider
  3. In Work and Codex: four models and a power slider
  4. In the API: four models and an effort parameter
  5. If you are coming from o3 or o4-mini
  6. How to choose in practice
  7. Your turn
  8. Key takeaways

Reasoning is no longer a separate model category

Not long ago, OpenAI separated conversational models from “reasoning” models like o3, o3-pro, or o4-mini. That era is over: o3 left ChatGPT on August 26, 2026, following a transition announced on May 28, and its API versions will be retired on December 11, 2026. Today, every model can respond quickly or think deeply, and you control that intensity. Choosing the right model and the right reasoning level is therefore the first skill this course teaches.

This setting looks different depending on where you work. This lesson maps out the three surfaces: Chat, Work and Codex, and the API.

In Chat: GPT-6 and the Thinking slider

In ChatGPT, the model menu offers GPT-6 and a Thinking slider: Instant for fast responses, Medium for standard reasoning, High for deep reasoning, Extra High for the most advanced level of GPT-6 Sol, and Pro, powered by GPT-6 Astra, for difficult tasks and long-form work. Plus accounts access Medium and High; Pro, Business, and Enterprise accounts access all levels; Free and Go accounts get a Luna model and a Think button for harder questions. The course ChatGPT, Getting Started (Lesson 5) details this menu.

In Work and Codex: four models and a power slider

ChatGPT Work and Codex offer four models: Astra, the most capable; GPT-6.1 Sol, close to Astra at a lower cost; GPT-6.1 Terra, fast and efficient for writing, image understanding, and app control; and Luna, for targeted, repetitive tasks. A power slider ranges from Faster to Smarter, with two extreme settings completing the lineup: Max, which gives the model more time on a single task, and Ultra, which distributes a complex task across sub-agents working in parallel.

Three surfaces, one principle Three columns. Chat: GPT-6, a slider from Instant to Extra High, then Pro with GPT-6 Astra. Work and Codex: Astra, GPT-6.1 Sol, GPT-6.1 Terra, and Luna, a slider from Faster to Smarter, then Max and Ultra. API: the reasoning.effort parameter, from none to max depending on the model, then reasoning.mode pro. ADJUSTING REASONING Three surfaces, one principle Chat Model GPT-6 Slider Instant to Extra High Beyond Pro, with GPT-6 Astra Work and Codex Models Astra, Sol, Terra, Luna Slider Faster to Smarter Beyond Max and Ultra API Parameter reasoning.effort Values none to max depending on the model Beyond reasoning.mode pro Everywhere: start with the default setting, increase effort before changing models.
Based on the OpenAI Help Center, the ChatGPT Models documentation page, and the API Models page (November 2026).

In the API: four models and an effort parameter

For developers, the Responses API exposes the same families, with a reasoning.effort parameter that sets the reasoning intensity. Accepted values depend on the model:

ModelIdentifierAccepted EffortPrice per million tokens (input / output)
GPT-6 Astragpt-6-astralow to max$10 / $50
GPT-6.1 Solgpt-6.1-sollow to max$2 / $10
GPT-6.1 Terragpt-6.1-terranone to max$1 / $5
GPT-6 Lunagpt-6-lunanone to max$0.10 / $0.50

All four share a 1.05 million token context window and support up to 128,000 output tokens. Sending none to GPT-6 Astra returns a 400 error; GPT-6.1 Sol rejects none and minimal, defaulting to medium effort. A second, independent setting is also available: reasoning.mode, which defaults to standard or can be set to pro for difficult tasks that tolerate higher latency and token usage. There is no “ultra” value in the API: Ultra is a mode specific to ChatGPT Work and Codex.

If you are coming from o3 or o4-mini

The official migration path is simple. In the API, OpenAI recommends replacing o3 with gpt-5.6-sol, and o3-pro with gpt-5.6-sol with reasoning.mode set to pro; these GPT-5.6 models remain supported. For new projects, start with the GPT-6 family instead. In ChatGPT, there is nothing to migrate: the GPT-6 thinking slider does what the old reasoning models did, without requiring a model switch.

How to choose in practice

The golden rule comes down to one sentence: start with the default setting, and increase reasoning effort before switching models. OpenAI recommends this everywhere: in the API, medium is the balanced starting point for most use cases; in Work, start with the setting proposed for your account and slide toward Smarter when a task demands more planning or analysis. Only move to a higher-tier model when the current model, even at high effort, truly hits a ceiling.

Specifically, a fast level (Instant, or none to low in the API) covers formulation tasks: drafting an email, summarizing, translating, or classifying. Move to a medium or high level as soon as the task becomes technical (code, calculations, data analysis), depends on multiple intermediate steps, or requires the model to verify its own work. Reserve the highest levels—Extra High, Pro, xhigh, or max—for difficult problems where quality significantly outweighs latency and cost. Luna, at the other extreme, is the right choice when a simple task repeats thousands of times.

Four pitfalls come up repeatedly. Running everything at the maximum level is slow and expensive. Ignoring Luna means paying full price for bulk processing. Switching models before increasing effort skips a step that is often sufficient. And triggering Ultra out of curiosity costs a lot for nothing: most tasks need neither Max nor Ultra.

Your turn

Two settings, one question. Ask the same question at two reasoning levels, either in ChatGPT (Instant then High, or Think on Free and Go) or in the API (low then high): “Calculate the net present value of a €50,000 project that generates €12,000 in annual revenue for 7 years, with a discount rate of 5%. Detail the calculation.” Compare the response time, the detail of the steps, and the result, then verify the result with a spreadsheet.

Official sources: GPT-6 and other models in ChatGPT (OpenAI Help Center: Thinking slider, plans) · Models (ChatGPT documentation: Work and Codex, Max and Ultra) · Models (API: identifiers, accepted effort, pricing, context) · Reasoning models (effort and pro mode) · Deprecations (o3 and o3-pro, replacements) · ChatGPT release notes (o3 removal from ChatGPT)

Key takeaways

  • Reasoning is a setting, not a separate model; o3 left ChatGPT on August 26, 2026, and leaves the API on December 11, 2026
  • Chat: GPT-6 and the Instant to Extra High slider, then Pro; Work and Codex: four models, Faster to Smarter, plus Max and Ultra
  • API: reasoning.effort from none to max depending on the model, and reasoning.mode set to standard or pro; no “ultra”
  • Start with the default setting and increase effort before changing models
  • Reserve maximum levels, Max and Ultra, for problems that resist everything else
Next lesson →How Reasoning Models Work