GPT-5.6's focus slider, Luna as the Free default, and the o3 and 5.4 retirements

OpenAI's August 4 ChatGPT update did three things. For Plus and Pro, GPT-5.6 Sol got higher factual reliability and a slider that sets how much reasoning goes into each answer. For Free, GPT-5.6 Luna became the default with a Think button for harder questions. And, in the lifecycle notes most people skip, two dates: o3 retires on August 26, and GPT-5.4 and 5.4-mini leave Codex for ChatGPT sign-in users on August 31. The slider is the visible change. The dates are the ones that will page someone.

Thinking budget as a UI primitive

A reasoning slider is a thinking budget exposed to the end user. That is a product decision with a clear message: OpenAI does not want to pick one effort level for every question, because effort costs latency and money and most questions do not need much of it. The Free tier gets the same idea in a cheaper form: Luna answers fast by default, and the Think button is an opt-in to spend more.

For anyone building on the API this matches what you already control with effort parameters, and the consumer UI is confirmation that per-task effort is the intended operating model, not a workaround. The GPT-5.6 family has always been a spread of models with different price and latency points; the slider just makes that spread visible to people who never read pricing pages.

Reasoning effort is a per-task parameter, and the vendor putting a slider in the consumer app tells you it should never be a global default in your pipeline.

Effort per task type

The way I set effort in production pipelines, and the slider makes the same shape obvious for chat:

• Extraction, classification, formatting: minimum effort. The answer is in the input; reasoning adds latency and nothing else.
• Summaries and rewrites: low. A little planning helps structure; more does not help facts.
• Code changes with tests: medium to high, scaling with the size of the diff. This is where Codex's model dial earns its place.
• Multi-step agent planning and debugging: high, but bounded, with a budget cap so a stuck loop cannot spend unlimited tokens thinking.
• Anything a user is waiting on synchronously: cap it regardless of task, because time-to-first-token is the thing they feel.

Measure two numbers per task type before you settle: quality at each effort level on your own eval set, and the cost and latency curve. The right setting is the lowest effort at which quality stops improving, and it is different for every row above.

The retirement dates

o3 retiring on August 26 affects anyone who still routes reasoning-heavy work to it via the API. GPT-5.4 and 5.4-mini disappearing from Codex on August 31 affects ChatGPT sign-in users of the CLI and IDE integrations; API-key users are not named in that note, so do not assume the same date applies to you without checking. The failure mode is the usual one: an alias that silently resolves to a different model, and an eval suite that starts drifting for no reason anyone can see in the code.

• Grep every config, environment file and prompt registry for o3, gpt-5.4 and 5.4-mini.
• Pin explicit model IDs; never depend on an alias the vendor can repoint.
• Run your eval set against the replacement model now, while the old one still answers, so you have a before and after.
• Re-tune effort on the replacement; a newer model at the same effort level often uses a different number of tokens.
• Put a calendar entry a week before each date, because the migration is trivial in July and an outage in September.

The honest gap

The August 4 notes describe the slider and the Think button for ChatGPT only. They do not say how slider positions map to API effort levels, whether Luna's Think mode corresponds to a specific API setting, or what the factual-reliability improvement in Sol was measured against. Until OpenAI publishes those mappings, treat the consumer controls as a signal about intent, not as a specification you can replicate, and keep your own effort calibration as the source of truth.