- OpenAI deprecated
gpt-image-2on 10 November and removes it from the API on 10 February 2027. The recommended replacements,gpt-image-2.6-flareandgpt-image-2.6-sunburst, cost the same $30 per million output tokens. The quality names are where the swap goes wrong: on gpt-image-2,highwas the top of three steps. On 2.6 it is the middle of five. - I sent 19 image requests on 11 November at 1536x864. gpt-image-2 at
highbilled 4,312 output tokens. Flare and Sunburst athighbilled 1,080, and onlymaxcame back to 4,312. Going the other way,lowwent from 120 tokens to 504. Change the model string and nothing else, and every quality setting you had lands at a different price. - Two more things don't carry across. The 2.6 models have no Batch support and no Batch price, so the 50 percent discount on gpt-image-2 jobs is gone. And they ship with no dated snapshot: asking for one returned a 400. If you never set quality at all,
autochoselowon every one of my seven calls, including the script that makes this blog's heroes.
OpenAI's deprecation notice is three lines long. The model and its gpt-image-2-2026-04-21 snapshot go away on 10 February 2027, and the replacement column says gpt-image-2.6-sunburst or gpt-image-2.6-flare. Nothing about settings. gpt-image-2 launched in April, so it gets about ten months of life and three months of notice.
The pricing page makes the swap look free. All three models charge $8 per million image input tokens and $30 per million image output tokens. That part is true. What you pay per image depends on how many output tokens a quality setting produces, and that number moved.
The ladder got two more rungs
gpt-image-2 accepts low, medium, high and auto. The 2.6 models accept low, medium, high, xhigh, max and auto. Send xhigh to gpt-image-2 and you get a clean 400: "The model 'gpt-image-2' does not support quality 'xhigh'." So the old top step and the new middle step share a name.
I sent the same photoreal prompt at 1536x864, a 16:9 hero size, to every combination on 11 November. One call per cell, run in parallel, so treat the timings as rough. Flare and Sunburst billed within two tokens of each other at every level.
| Quality | gpt-image-2 tokens | 2.6 tokens | gpt-image-2 cost | 2.6 cost |
|---|---|---|---|---|
| low | 120 | 504 | $0.004 | $0.015 |
| medium | 1,078 | 924 | $0.032 | $0.028 |
| high | 4,312 | 1,080 | $0.129 | $0.032 |
| xhigh | 400 error | 1,917 | n/a | $0.058 |
| max | 400 error | 4,312 | n/a | $0.129 |
Read the high row first. A pipeline that ran gpt-image-2 at high and now runs Flare at high will see its image bill drop by three quarters. Nobody will complain about that in the first week. Whether the pictures got worse is a separate question, and on my prompt I can't say they did. Flare's high rendered film canisters with more convincing rims than gpt-image-2's high. But the budget the model spends on rendering is now a quarter of what it was, and that is a change you made by accident. If your quality bar was tuned on the old high, the matching spend on 2.6 is max.
The low row runs the other way. A thumbnail or draft job on gpt-image-2 at low cost about four tenths of a cent at this size. On 2.6, low costs 4.2 times as much. If you generate a lot of throwaway previews, that's the line that moves your invoice.
medium is the only setting that roughly holds: 1,078 tokens versus 924.
Flare and Sunburst bill the same and run at different speeds
Same tokens, same price, very different waits. At max, Flare returned in 18.7 seconds and Sunburst in 86.7. At high, 7.9 against 33.6. gpt-image-2 at high took 59.9 seconds for the same 4,312 tokens that Flare produced at max in under 19.
OpenAI positions Sunburst for detailed work and editing precision and Flare for fast iteration. For a text-to-image hero with no edit step, I couldn't find a reason in my output to pay Sunburst's latency. Your edit workflows may differ, so test those separately.
The 2.6 models also add a complexity field, which controls how much the model plans before rendering. It defaults to high. Setting it to low on Flare at high quality changed nothing I could measure: 1,080 tokens either way, 8.3 seconds against 7.9. The docs say the setting has no separate fee and that higher complexity can "introduce unwanted changes," which matters more for edits than for fresh images.
What doesn't migrate at all
Batch. The gpt-image-2 model page lists v1/batch as supported, and the pricing page has a Batch table that puts gpt-image-2 at $15 per million output tokens instead of $30. The same holds for the 2.5 Sunburst and Flare models. Both 2.6 model pages list Batch as "Not supported," and neither appears in the Batch price table. I couldn't confirm the rejection with a live job: the restricted key I test with lacks the files-write scope a batch upload needs. But if you run nightly image jobs through Batch, the recommended replacement removes the discount, and the closest model that keeps it is 2.5, the generation before.
Pinning. gpt-image-2 had a dated snapshot, gpt-image-2-2026-04-21. The 2.6 model pages list exactly one snapshot each, and it is the bare alias. I tried gpt-image-2.6-flare-2026-11-04 anyway and got "The model 'gpt-image-2.6-flare-2026-11-04' does not exist." So whatever OpenAI changes behind gpt-image-2.6-flare reaches you the day it changes. If you run visual regression tests on generated assets, note that you can't freeze the model under them.
auto picks low, and our own script never noticed
This is the part I found in my own repo. The script that generates hero images for this blog calls gpt-image-2 with a prompt and a size and nothing else. No quality. The default is auto, which the docs say selects "the best option based on the prompt."
On 11 November, auto returned low on all four gpt-image-2 calls I made without a quality setting, across two different prompts and two sizes. Flare did the same twice and Sunburst once. Each of those gpt-image-2 heroes billed 120 to 158 output tokens. The response tells you this, because the Images API echoes the resolved quality back next to usage. We just never printed it.
So the first migration step is not the model string. Log the quality and usage.output_tokens fields from a week of your current gpt-image-2 traffic. That tells you which rung you're really on, which is often not the one you think you asked for.
The migration checklist
- Log resolved
qualityand output tokens on your current gpt-image-2 calls. If you seelowwhere you expected better, fix that before anything else. - Map settings by spend, not by name: old
highis newmaxat 4,312 tokens for 1536x864. Then tryxhighat 1,917 andhighat 1,080 and decide with your eyes whether the cheaper rungs are good enough. On my prompt, they were. - Budget for
lowcosting about four times as much if you generate drafts or thumbnails in volume. - If you rely on Batch, either stay on
gpt-image-2.5-sunburstorgpt-image-2.5-flarefor those jobs or accept the full rate. - Default to Flare for plain generation unless a side-by-side on your edit tasks shows Sunburst earning its wait.
- Treat the 2.6 alias as unpinned and keep a few reference outputs to compare against when results drift.
The whole test cost about 73 cents: 24,040 image output tokens across both model families. That is less than six gpt-image-2 heroes at high, and it answered a question the deprecation notice didn't.


