As of September 4, 2026, OpenAI's model catalog lists no model named GPT Image 2.5; the newest image model ID is gpt-image-2. "GPT Image 2.5" is a community label for anonymous Arena checkpoints and a community-reported announcement string in OpenAI's web client code. No API endpoint, model ID, price, or release date has been published, and no EvoLink route exists.
As of September 4, 2026: not announced, no release date, no model ID, no price, no rate limits, and no EvoLink route. Each row links to the source that was checked; the EvoLink route row links to EvoLink's own checklist below.
| Item | Status | Checked against |
|---|---|---|
| OpenAI official status | Not announced. OpenAI's catalog, changelog, pricing page, and SDK model types list no image model newer than gpt-image-2. | OpenAI model catalogLast checked September 4, 2026 |
| Release date | Not announced. The dated evidence timeline lives on the release status page. | GPT Image 2.5 release statusLast checked September 4, 2026 |
| Model ID | Not published. Any string such as gpt-image-2.5 in code today is a guess, not an identifier. | OpenAI GPT Image 2 model pageLast checked September 4, 2026 |
| Price | Not published. gpt-image-2 is the newest OpenAI image model with a published price; the pricing page has no 2.5 row. | OpenAI API pricingLast checked September 4, 2026 |
| Rate limits | Not published. gpt-image-2 tiers run from 5 to 250 images per minute on OpenAI direct. | OpenAI GPT Image 2 model pageLast checked September 4, 2026 |
| EvoLink route | Not available. A route appears here only after a verified model ID, approved pricing, one generation and one edit through the production route, and matching billing. | EvoLink launch checklistLast checked September 4, 2026 |
This is EvoLink's own launch alert. It is not an OpenAI waitlist, early-access program, or guarantee of access. You get one email when the route passes verification, and nothing before that.
EvoLink launch alerts only. No OpenAI affiliation, no early-access promise, no spam.
Want faster updates? Join the EvoLink Discord
The official model name and documentation, the request model ID, approved pricing in the price module, one successful generation and one successful edit through the production route, and matching billing records. Until all five pass, this page stays in watch state.
Three things are confirmed by OpenAI's own documentation as of September 4, 2026: no GPT Image 2.5 exists in the catalog, gpt-image-2 is the current baseline, and OpenAI has shipped a .5 image release before. Nothing about a next model is confirmed.
The model list, the gpt-image-2 model page, the pricing page, and the API changelog contain no GPT Image 2.5 entry as of September 4, 2026. The Node and Python SDK image model types include no model newer than gpt-image-2 and its dated snapshot.
Released April 21, 2026 with snapshot gpt-image-2-2026-04-21. Generation and editing, low/medium/high quality, PNG/JPEG/WebP, sizes up to 3840 px on the long edge, and a transparent-background preview added August 20, 2026.
GPT Image 1.5 launched December 16, 2025 between GPT Image 1 and GPT Image 2. That is a naming pattern, not a prediction about the next model's name.
Nothing about a next image model has been published by OpenAI. Reported improvements are community expectations that a release would have to prove.
Four signals exist, three community-reported and one first-hand, and none of them establishes a product name, a model ID, or a release date. Each is logged with its evidence level.
| imagegen25 announcement string | On September 3, 2026, community members reported a feature-gated announcement modal in OpenAI's web client code, component name imagegen-25-announcement-modal, with the copy "Image creation got a major upgrade" and "Higher-quality results, faster generation, and smarter creative tools". The exact client bundle is not established, the original screenshots could not be independently re-verified, and the string does not correspond to any model ID in OpenAI's API catalog or SDK model types. | Community-reported |
|---|---|---|
| mona-lisa-1 in Arena Battle Mode | EvoLink generated an output from the mona-lisa-1 label on August 11, 2026, two days after the first community reports. That confirms a callable anonymous checkpoint existed in that session. It does not identify the provider or connect the label to a product name. | First-hand observation |
| luna-lisa-alpha and later labels | From August 20, 2026, community posts describe luna-lisa-alpha as a later checkpoint with less noise, plus other short-lived labels. No official source links any of them to OpenAI or to each other. | Community-reported |
| OpenAI provenance signal | Community reports say some checkpoint outputs carried OpenAI provenance signals. OpenAI's own guidance states that C2PA and SynthID signals indicate an association with OpenAI tools; they do not identify a specific model. | Community-reported |
These five gates are EvoLink's launch checklist for this route. This page flips to a live API page only when all of them pass.
OpenAI publishes the model page, model ID, and pricing. Community labels are never used as identifiers.
The exact string accepted by EvoLink's OpenAI-compatible endpoint, confirmed by the backend and by a successful request.
EvoLink pricing is written into the same price module that powers every live model page. No estimates are shown before that.
A real text-to-image call and a real edit call, with the response format, timeouts, and retry semantics recorded.
Usage and billing records match the request log, including how failed or filtered requests are charged.
Six evaluation tasks map to the limits OpenAI documents for gpt-image-2 and to the complaints developers raise most often. Once a GPT Image 2.5 route is verified, EvoLink will run them head-to-head against GPT Image 2.
Exact brand text, material fidelity, and usable rate across white-background and multi-angle sets.
Precise text placement and hierarchy, which OpenAI lists as a current limitation of gpt-image-2.
Character accuracy and font consistency for Chinese, Japanese, and Korean, a frequent stress test in community threads.
Skin, hands, hard light, and low light, plus the high-frequency noise that testers report on current outputs.
Character and brand elements that hold across many generations, another documented gpt-image-2 limitation.
Identity and untouched-region drift after 3, 5, and 10 edits, and p50/p95 latency against the up-to-two-minute complex-prompt behavior documented today.
GPT Image 1.5 to GPT Image 2 is the only precedent for what a next release changes in the API contract. Every cell comes from OpenAI's model pages and pricing page; the next column stays empty until OpenAI publishes something.
| Contract item | GPT Image 1.5 | GPT Image 2 | GPT Image 2.5 |
|---|---|---|---|
| Model ID | gpt-image-1.5 | gpt-image-2 (snapshot gpt-image-2-2026-04-21) | Not published |
| Image output price | Published token price | Published token price, lower per image-output token than 1.5 | Not published |
| Sizes | Three fixed sizes: 1024×1024, 1024×1536, 1536×1024 | Any size up to 3840 px on the long edge, 1K/2K/4K presets, ratios up to 3:1 | Not published |
| Catalog status | Previous image model | Current image model; transparent background in preview since August 20, 2026 | Not listed |
The fastest migration is the one you prepare on GPT Image 2 now. Keep the model string in configuration, freeze a baseline, and the switch becomes a one-line change once a verified ID exists.
# Today's request through EvoLink (OpenAI-compatible)
curl https://api.evolink.ai/v1/images/generations \
-H "Authorization: Bearer $EVOLINK_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-image-2",
"prompt": "Studio product photo, matte black headphones, soft light",
"size": "1024x1024",
"quality": "medium"
}'
# The response returns a task id; poll task status per the GPT Image 2 API reference.
# Read the model string from config. Do not hardcode a guessed 2.5 ID.
# IMAGE_MODEL=gpt-image-2 # swap only after EvoLink verifies a new routeRoute requests through one setting so a verified successor can be enabled without a redeploy, and rolled back the same way.
Save 30 to 100 representative prompts, source images, and acceptance rules with today's outputs, so any new checkpoint is judged against numbers, not memory.
Track first-pass rate, retries, latency, and total spend per usable output. A cheaper per-call price can still cost more per accepted image.
Decide now which route takes over when a new model is rate-limited, filtered, or slow: GPT Image 2, GPT Image 1.5, or another provider through the same endpoint.
gpt-image-2 is the documented OpenAI route with generation, editing, quality tiers, and a published price. It runs on EvoLink today through the same endpoint a future route would use, so nothing you build now is wasted.
A new model is a routing decision, not a rewrite. EvoLink exposes OpenAI, Google, ByteDance, and other image routes behind one OpenAI-compatible endpoint, so evaluation, fallback, and cost control stay in one place when a new route is verified.
Switch between gpt-image-2, Nano Banana 2, Seedream, and others by changing the model string.
New routes on EvoLink go through a real request and billing check before their page shows a price.
Route around rate limits and filtered requests, and compare cost per accepted output across providers.
No. As of September 4, 2026, OpenAI's model catalog, pricing page, changelog, and SDK model types include no image model newer than gpt-image-2. There is no GPT Image 2.5 endpoint.
None has been published. Any identifier such as gpt-image-2.5 that appears in third-party code or pricing tables today is a guess.
No price has been published. EvoLink will show a price only after OpenAI publishes one and the route passes billing verification. The current baseline is GPT Image 2's published pricing.
Not yet. EvoLink has no GPT Image 2.5 route. Use gpt-image-2 today; it runs through the same endpoint a future route would use.
No official name exists. "GPT Image 2.5" is a community label; some community members expect the next release to be called GPT Image 3. This page keeps its URL and updates the mapping if OpenAI chooses a different name.
A community-reported string from a feature-gated announcement modal found in OpenAI's web client code on September 3, 2026. It suggests an image upgrade is being prepared; it does not confirm a model name, an API, or a date.
That has not been confirmed. mona-lisa-1 is an anonymous Arena label EvoLink observed directly on August 11, 2026. Arena's policy allows providers to test unreleased models anonymously, so the label proves neither the provider nor the final product.
Unknown. The community-reported modal copy mentions higher quality and faster generation, and testers report less noise on later checkpoints. These are expectations a release would have to prove in controlled tests.
Join the launch alert on this page. The alert fires once, after the route passes EvoLink's five verification gates, and this page switches to the live API page with the model ID and price.
Ship on GPT Image 2 with the model string behind configuration and a frozen evaluation baseline. That is the fastest path to adopting a successor and the safest path if it never ships under this name.
Dated evidence timeline, the imagegen25 signal, and whether to wait or ship on GPT Image 2.
EvoLink's first-hand test of the anonymous mona-lisa-1 checkpoint and what it does and does not prove.
The current OpenAI image route on EvoLink with pricing, playground, and API reference.
Compare GPT Image 2, 1.5, and 1 and pick a route by workflow.