What Actually Changes in ChatGPT on 26 August 2026
OpenAI o3 leaves the ChatGPT model picker on 26 August 2026, at the end of a 90 day sunset period. The retirement covers ChatGPT only, not the API. Your prompts keep working, but from that date they run on a different model, chosen by OpenAI rather than by you.
Most people who kept o3 pinned in their picker have not noticed the notice. Almost nobody has checked whether the model OpenAI points them to is the one they should actually be using.
This is the second forced move inside ChatGPT this year. GPT-4.5 was retired on 27 June 2026 under the same kind of sunset schedule. The official DALL·E GPT is next, retiring on 30 August 2026.
The pattern matters more than any single retirement. Model choice used to be a setting you made once. It is now a maintenance item with a calendar entry, and if you treat it as permanent, a vendor will eventually make the decision for you.
Why OpenAI's Named Replacement Is Not Automatically Yours
OpenAI's published successor to o3 is GPT-5.6 Sol, its flagship tier. Sol is listed at US$5 per million input tokens and US$30 per million output. GPT-5.6 Terra, the balanced tier, sits at US$2 and US$12 after a price cut on 30 July 2026. For most work people did with o3, Terra is the honest starting point.
The reason is simple. o3 was a reasoning model used for an extremely wide spread of tasks, and the large majority of those tasks never needed frontier-tier compute. Summarising a messy transcript, restructuring a brief, checking a spreadsheet's logic: none of that is a Sol-class problem.
Terra lands at roughly 40 percent of Sol's token cost. If you route everything to Sol because a release note said so, you have accepted a price increase that nobody asked you to justify.
There is a timing trap here too. OpenAI cut Sol's API and credit pricing by more than 20 percent in August 2026, but only for three months. A budget built on a promotional rate is a budget that breaks in November.
If you only use ChatGPT through a subscription, you do not pay per token, and the picker now presents Instant, Thinking and Pro rather than model names. The cost question still reaches you indirectly, through which tier your plan gets and how much reasoning it is allowed to spend.
The 20-Minute Replacement Test You Should Run Before Wednesday
The reliable way to pick a replacement is to benchmark your own three most-used prompts against the cheaper candidate first, then escalate only where it measurably fails. This takes about twenty minutes and produces a defensible default instead of a vendor recommendation.
Open your ChatGPT history and find the three prompts you actually reuse. Not the impressive ones. The boring, repeated ones that carry your week.
For each, write a success test before you run anything: one sentence describing how you would know the output is wrong. Without that, you will judge on tone and pick whichever model sounds most confident.
Try this prompt:
--- You are helping me migrate off a retired AI model. I will paste one prompt I use regularly.
--- Step 1: Restate what this prompt is actually asking for, in one sentence.
--- Step 2: List the three failure modes most likely to appear in a weaker model's answer to it.
--- Step 3: Write a SUCCESS TEST: a checklist of 4 items I can verify in under 60 seconds to decide whether an output is acceptable.
--- Step 4: Tell me the cheapest class of model that can plausibly pass that test, and what specifically would force an upgrade to a stronger one.
--- Do not rewrite my prompt. Do not answer it. Only produce the analysis above.
--- MY PROMPT: [paste your prompt here]
Then run the original prompt on the cheaper candidate, score it against the checklist you just generated, and only move up a tier where a specific checklist item fails. Record the result somewhere you will find it again.
Where Each Destination Actually Fits
There is no single replacement for o3, because o3 was doing several different jobs. Match the destination to the job: a mid tier for routine analysis, a flagship for genuinely hard reasoning, a long-context model when the deliverable itself is large. These are the published facts that matter for that choice.
Published token pricing, as of 25 August 2026
--- GPT-5.6 Terra: US$2 input, US$12 output per million tokens, cut from US$2.50 and US$15 on 30 July 2026.
--- GPT-5.6 Sol: US$5 input, US$30 output. API and credit pricing cut by over 20 percent for three months from August 2026.
--- GPT-5.6 Luna: the fastest and cheapest tier, launched at US$1 input and US$6 output, also reduced in July 2026.
--- Claude Sonnet 5: US$2 input, US$10 output. Anthropic made this permanent on 10 August 2026, cancelling a planned rise to US$3 and US$15 on 1 September.
--- Claude Opus 5: available since 24 July 2026, with a 1 million token context window and up to 128K output tokens.
Which job goes where
--- Routine analysis, drafting, restructuring, data sanity checks: start on Terra or Claude Sonnet 5. Both sit at the US$2 input mark.
--- Genuinely hard multi-step reasoning where a wrong answer is expensive: Sol, and only after Terra has failed your checklist on that specific task.
--- Large documents where the deliverable is long: Claude Opus 5, where the context and output ceilings are roughly double most rivals'.
--- Fast, high-volume, low-stakes calls such as classification or tagging: Luna.
Reasoning effort is now a setting rather than a sentence you write. We covered how that changes prompt construction in this piece on prompting reasoning models, and it applies directly to whichever destination you land on.
The Gotchas That Quietly Cost People Money
Three things break model migrations: token accounting that changes underneath you, latency you did not budget for, and the assumption that a ChatGPT announcement also covers your integrations. Each has a specific fix.
A lower headline rate is not automatically a lower bill. Anthropic's newer tokenizer can raise billable token counts by up to 35 percent, which can eat the gap between a US$2 and a US$2.50 input rate. Measure on your own text, not on a comparison table.
ChatGPT retirement is not API retirement. The o3 snapshots o3-2025-04-16 and o3-pro-2025-06-10 carry API removal dates of 11 December 2026, set in a developer notice on 11 June 2026. If a Zapier, Make or n8n step names an o3 snapshot, it keeps running for now and breaks in December. Put that date in your calendar today.
Legacy access reports conflict. Some sources say o3-pro stays reachable for higher paid tiers through legacy model settings, others describe it as gone from the standard picker. Open your own settings and look, rather than trusting a summary written for a different plan.
More reasoning is not free. Higher reasoning levels add latency and cost, and on straightforward tasks they frequently add nothing else. Escalate on evidence, not on nerves.
One more honest limitation: the cost tables above are published list prices, not your invoice. Cached input, batch discounts and per-plan credit allowances all move the real number, and any comparison that ignores your own usage mix is decoration.
Make Model Choice a Maintained Setting, Not a Memory
Retirements will keep arriving, so the durable fix is a written default rather than a habit. A short file turns the next forced migration from a day of firefighting into a twenty minute re-run of a test you already wrote.
Keep four columns in a plain text note: task type, model, why, last tested date. Four lines is enough for most people. It is the "why" column that saves you, because in three months you will not remember it.
Keep your three golden prompts and their expected output shape next to that note. That is your benchmark set, and it is worth more than any leaderboard, because it is made of your actual work.
Where you can, avoid naming a specific model inside prompt text and templates. Model names in prompts are the reason a retirement turns into an editing job across dozens of saved documents.
Then re-run the set on a fixed cadence, quarterly is plenty, and immediately on any retirement notice. The cost of a switch is not the price difference. It is the afternoon you lose because nobody wrote down what "good" looked like.
The Takeaway Before Wednesday
o3 leaves ChatGPT on 26 August 2026. The named successor is the flagship tier, but the cheaper mid tier covers most of what you were doing, and the only way to know is to test your own three prompts against a checklist you wrote first.
Model churn is not going to slow down. What you can control is whether each retirement notice finds you with a documented default or with a picker you have not looked at since last year.
That is the part where a partner helps more than a leaderboard does. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team.
Want to Know Where Your AI Skills Actually Stand?
Knowing which model to switch to is one skill. Knowing which of your workflows still deserve a frontier model is another. Take UD's free AI IQ Test to benchmark your own AI level, and if you want the workflow built properly, we'll walk you through every step, from model selection to prompt templates to deployment.
