What a retiring model actually costs you to replace
Two models leave OpenAI's lineup on the same day this catalogue tracks: GPT-4 Turbo and GPT-3.5 Turbo both retire on 23 October 2026. OpenAI names a replacement for each. Neither replacement is a same-price, same-shape swap — the two moves push cost in opposite directions.
GPT-4 Turbo → GPT-5.6 Sol#
GPT-4 Turbo currently charges $10 per million input tokens and $30 per million output, inside a 128,000-token window. Its named successor, GPT-5.6 Sol, charges $5 input — half — for the same $30 output, and accepts up to 1,050,000 tokens, roughly eight times the room.
Read only the headline rate and this migration looks like a clean win: half the input cost, eight times the ceiling. Two things complicate that.
First, Sol carries a long-context tier: past 272,000 input tokens, the entire request is billed at double input and 1.5x output. A workload that used to fit inside GPT-4 Turbo's 128,000-token ceiling stays comfortably under that line, so most migrations from Turbo won't hit it — but a workload that specifically moved to Sol because of the bigger window, and now routinely sends 300,000+ tokens, is paying $10 per million input on those requests, not $5. The headline saving evaporates exactly for the use case the bigger window was supposed to enable.
Second, the two models use different tokenizers — GPT-4 Turbo counts with cl100k_base, Sol with o200k_base. A token count measured against one does not carry over to the other; the same text produces a different number on each. Any cost estimate carried forward from a Turbo-era measurement needs to be recounted, not rescaled.
GPT-3.5 Turbo → GPT-5.6 Terra#
The other retirement runs the opposite direction on price. GPT-3.5 Turbo is $0.50 input and $1.50 output today, inside a 16,385-token window — by far the smallest window in the current catalogue. Its replacement, GPT-5.6 Terra, is $2 input and $12 output: four times the input rate, eight times the output rate.
The window, though, goes from 16,385 tokens to 1,050,000 — a roughly 64x increase. For a workload that was actually constrained by GPT-3.5 Turbo's small ceiling — anything that had to truncate history or chunk a document to fit — the new window may be worth the higher per-token rate outright, since the alternative wasn't "cheaper," it was "didn't fit." For a workload that never came close to 16,385 tokens in the first place — a short-form chat reply, a single-sentence classification — the migration is a straightforward 4x-to-8x price increase for capacity it will never use.
The output ratio matters more here than on the Sol migration, because it moved further: 3-to-1 on GPT-3.5 Turbo versus 6-to-1 on Terra. A workload with long, verbose replies feels that shift on the output line more than on the input line.
What both moves have in common#
Neither replacement is priced or shaped to make the migration invisible. Both retiring models are among the cheaper or more tightly-bounded options in the catalogue; both replacements sit in a newer generation with a different tokenizer, a different context ceiling, and in Sol's case, a tier the old model never had at all.
What to do with this#
With roughly eight weeks left before the 23 October cutoff, the useful exercise isn't reading the two headline rates — it's re-running your actual prompts through the replacement's real tokenizer and, if they're large, checking which side of Sol's 272,000-token threshold they land on. A workload's real migration cost depends on its own token counts under the new tokenizer, not on the model-to-model rate comparison alone.