GPT Image 2 Has 92 Days Left, and the Replacement Is Not a Drop-In

OpenAI's deprecations page gained a new section on November 10, and it is the kind of entry that creates work for a few thousand engineering teams.
The text is short: the gpt-image-2 model and the gpt-image-2-2026-04-21 snapshot are deprecated and will be removed from the API on February 10, 2027. The recommended replacements are gpt-image-2.6-sunburst and gpt-image-2.6-flare. The model page for gpt-image-2 now reads "Status: Deprecated".
Two model identifiers are on the list, not one. The moving alias and the pinned snapshot both die on the same date, which matters because a lot of code pins the snapshot precisely to avoid the alias moving underneath it. Pinning bought stability for a year and did not buy immunity.
The ninety-two days
From the November 10 announcement to the February 10 removal is 92 calendar days. That is not arbitrary. OpenAI's published notice tiers set a floor of at least six months for generally available models, at least three months for specialised variants and for image generation models, and roughly two weeks for previews. GPT Image 2 sits in the image tier, so 92 days clears the floor by two days.
Two days of margin is the entire cushion between "compliant notice" and "you missed it". If your team ships quarterly, this lands in the middle of the next cycle.

There is also a detail that any production team will notice: the notice gives a date, not a time, and not a time zone. Plan as if the API stops answering at the start of February 10 in whatever zone your on-call engineer is in, or finish in January and stop thinking about it.
What the swap looks like, and where it lies
The mechanical migration is genuinely small. Both 2.6 models support the same three endpoints (image generations, image edits, and Batch) so the HTTP surface does not move. The token rates are row-for-row identical to the outgoing model: $8 per million image input tokens, $30 per million output, $2 cached input, with a half-price Batch row.
So a developer could reasonably conclude that this is a string replacement. It is not, and the notices say so in the way that matters: the replacement names are different models, not a new version of the same one.
What actually changes when you move to 2.6:
GPT Image 2.6 arrived on November 4 with adjustable complexity tiers, SVG output, animated GIF generation, and markup-guided editing extended to the 2.5 and 2.6 families. Flare is the smaller, faster model positioned for iteration on everyday work. Sunburst is the base model aimed at detailed compositions and edit accuracy. OpenAI's own prompting guidance suggests starting with Flare on workloads where latency dominates, and Sunburst where object relationships and layout precision dominate.
Both of those words (complexity, markup) describe parameters that did not exist on the model you are leaving. The new models can be cheaper per usable output and more expensive per call at the same time, because a tunable budget means you can now overspend on a request that used to have one setting.
The three places this breaks
First, retries and fallbacks. Image pipelines accumulate model identifiers in places nobody audits: a queue worker's config, a fallback chain in the Responses API path, a cost estimator table, an allowlist, a test fixture, a Notion page that a vendor's account manager keeps current. Grepping for the string catches code. It does not catch the spreadsheet.
Second, output QA. A new model generation shifts output in small ways even when the API shape is stable, and the drift is rarely dramatic enough to trigger an alert. Nobody files a bug because a shade moved two points. If your pipeline generates covers for articles, product thumbnails, or marketing variants, the acceptance criteria are visual, not numerical. The only way to know whether Sunburst still renders the label on a bottle correctly is to run the same source image and prompt through both models and compare the label.
Third, billing. As of mid-November, the pricing page's Batch section had no gpt-image-2.6 row. The outgoing model has one. If your volume runs through Batch at half price, treat 2.6 Batch billing as unresolved until the page updates, and do not model your costs on an assumption.
A tangled calendar
The GPT Image family has been churning all year, and the November notice closes a chain rather than opening one.
gpt-image-2 launched on April 21, 2026, with a snapshot dated to launch day. On May 12, DALL·E 2 and 3 were hard-removed from the API and gpt-image-2 absorbed the migrations. On September 8, GPT Image 2.5 entered the official SDK types. On October 23, gpt-image-1 shut down under its own earlier notice, with 2.5 as the official replacement. On November 4, GPT Image 2.6 shipped. Six days later, GPT Image 2 got its end date.
And there is a second deadline that most write-ups of this story miss. In the same pass, OpenAI re-pointed the older June 2026 deprecation table: gpt-image-1-mini, gpt-image-1.5 and the chatgpt-image-latest alias all shut down on December 1, 2026, and their listed replacement changed from gpt-image-2 to gpt-image-2.6-sunburst.
Read that twice. The model that was supposed to be the safe landing spot for the December 1 shutdown is itself shutting down seventy-one days later. Anyone who migrated to gpt-image-2 in November to escape the December deadline has to migrate again in January.
What to do this week
Scan for both identifiers, alias and snapshot, across code, config, queues, tests and docs. The grep that catches one often misses the other, because a pin written as a string constant and a pin written in a YAML file do not look alike. Move new development straight to 2.6 and skip the intermediate step. Build a small comparison set (ten real production inputs, same prompts, same sizes, both 2.6 models) and judge on whether the output passes your acceptance check rather than on whether it looks better. Pin whatever you choose, and put February 10 somewhere a human will see it.
The alternative is to discover this in February, when the endpoint returns an error and the article covers stop generating on a Tuesday morning. There is no incident response that fixes a missing model id, only a rollback to a model that is also going away.
None of this is dramatic. It is the ordinary maintenance cost of building on a rented model, and the bill arrives on a schedule you do not control. The teams that treat model ids as dependencies rather than settings are the ones who will not notice February. Everyone else gets a Tuesday morning they will remember.
Related articles
The Jensen Discount: What Nvidia Loses When Its Founder Walks Out
Nvidia lost roughly $600 billion in market value when its founder stepped down. The earnings did not change. The multiple did, and that says where the value really lived.
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.