OpenAI released ChatGPT Images 2.5 on September 8, 2026, with two API models: GPT-Image-2.5 Flare for fast, everyday generation and GPT-Image-2.5 Sunburst for workflows where editing precision matters most. The launch claims up to 50% lower generation latency than Images 2.0, better reference fidelity, and more consistent multi-turn edits. Those are meaningful improvements, but they are not a migration plan.
The production question is narrower: when a user asks to change one detail, how much of the approved image moves unintentionally? That is edit drift. A migration test should measure it across your own products, faces, layouts, typography, and brand rules. The result should decide which requests go to Flare, which justify Sunburst, and which still need human review.
The Release Changes the Routing Decision
OpenAI says people create more than 3 billion images each week across ChatGPT Images and GPT-Image APIs. Its Images 2.5 announcement positions Flare as the default for most applications and Sunburst as the higher-precision option for detailed creative work with longer generation times.
The official image generation guide makes the split operational. Use the Image API for one-shot generation or editing, and the Responses API for conversational, multi-turn image workflows. Both models accept text and image inputs, support generation and editing, and expose quality controls from low through max plus auto.
Do not route only on the labels "fast" and "precise." Route on the failure cost of the edit. A background cleanup for a disposable social draft can start on Flare. A product hero whose packaging, logo-free geometry, proportions, and lighting must survive six revisions deserves a Sunburst test. If both pass the same acceptance bar, choose using measured latency and cost.
Define Edit Drift as a Testable Failure
Create a small gold set from real work rather than public demo prompts. Twenty to forty source images can be enough for the first decision if they cover the risky shapes of your workload. Include portraits, products, dense compositions, transparent assets, text-bearing layouts, and scenes with repeated objects. Remove any item you do not have the right to process.
For each source, define one requested change and several protected regions. A test record can look like this:
{
"case_id": "product-hero-07",
"source_version": "sha256:...",
"instruction": "Replace the gray background with warm white.",
"editable_region": "background-mask.png",
"protected_regions": ["product", "label", "shadow"],
"must_keep": ["object count", "label wording", "camera angle"],
"must_change": ["background color"]
}
Score every output on two separate axes. Instruction success asks whether the requested change happened. Preservation asks whether unrelated content stayed fixed. A model can pass one and fail the other. Combining them into a single aesthetic score hides the most expensive regressions.
Use automated signals as triage, not as the final judge. Pixel or perceptual difference outside the editable region can flag large changes. OCR can detect altered copy. Face or product embeddings can catch identity drift. Object detection can catch additions and removals. Then have a reviewer inspect every borderline case and a random sample of passes.
Run a Five-Step Drift Ladder
A single edit is too easy. Production sessions accumulate instructions, so test a controlled sequence against the same approved source:
- Change the background color.
- Replace one accessory or product detail.
- Adjust lighting without moving the camera.
- Change one line of visible copy.
- Export the final asset with the required background and format.
After each step, compare the output both with the previous turn and with the original. The previous-turn comparison detects a sudden regression. The original comparison shows slow degradation that compounds across the session. Keep the prompt sequence identical between models and repeat each case enough times to reveal variance.
For conversational editing, the Responses API can continue from an earlier image-generation call or use previous_response_id. The multi-turn documentation explains that the image call can remain in context while later instructions refine it. Treat that context as part of the test fixture: a prompt that passes only after an unrecorded manual correction is not reproducible.
Treat Masks as Guidance, Not a Lock
Masks are useful for defining intended change, but they are not hard pixel boundaries. OpenAI's mask documentation explicitly says GPT Image uses a mask as guidance and may not follow its exact shape with complete precision. That is exactly why protected-region measurement matters.
Prepare the source and mask in the same size and format, keep the required alpha channel, and store both with the test case. Measure spill beyond the mask, especially around hair, glass, reflections, shadows, and thin product edges. Those areas make a visually reasonable edit look like a brand or catalog error.
Transparent-background work needs its own cases. The API guide says transparency requires background: "transparent" with PNG or WebP output. Test edge halos on light and dark checkerboards; a transparent file can be technically valid and still be unusable in a design system.
Compare Realized Cost, Not Rate Cards Alone
The Flare and Sunburst model pages list the same token rates: $5 per million text input tokens, $8 per million image input tokens, and $30 per million image output tokens, with lower cached-input rates. The Flare model page and Sunburst model page also note that the older GPT-Image-2 calculator does not estimate GPT-Image-2.5 token consumption.
That means the rate card does not settle the routing question. Log the input and output tokens, quality setting, dimensions, format, compression, wall-clock latency, retries, and reviewer outcome for every case. The useful unit is not cost per generated image. It is cost per accepted asset.
A cheap attempt that needs three retries and a manual repair can be more expensive than a slower first-pass result. Conversely, sending every rough concept to the precision model wastes latency if Flare clears the same acceptance threshold. Build a routing table from your measured pass rates:
| Workload | Default route | Escalate when |
|---|---|---|
| Rough concepts | Flare | Composition repeatedly misses the brief |
| Social variants | Flare | Brand elements drift or copy changes |
| Product edits | Sunburst test | Protected-region drift exceeds threshold |
| Multi-turn campaign assets | Sunburst test | Consistency falls across the drift ladder |
| Transparent cutouts | Benchmark both | Edge quality fails on either backdrop |
Pin the Snapshot and Keep a Rollback Lane
Both model pages expose dated snapshots: gpt-image-2.5-flare-2026-09-08 and gpt-image-2.5-sunburst-2026-09-08. Use a dated ID during evaluation and the first production rollout. An undated alias is convenient, but it can make a later behavior change difficult to separate from a prompt, preprocessing, or product change.
Roll out by workload, not by percentage alone. Start with reversible concept generation, then low-risk edits, then identity-sensitive or revenue-bearing assets. Preserve the old route, source images, prompts, masks, and acceptance thresholds until the new path has survived a full release cycle.
Also verify provenance in your delivery pipeline. OpenAI's Images 2.5 system card says generated images include C2PA metadata and Google SynthID across ChatGPT, Codex, and the API. If your image optimizer strips metadata, document that behavior rather than assuming provenance survives every transformation.
Limitations and Tradeoffs
Automated drift metrics are imperfect. A low pixel difference can hide a changed digit, while a harmless lighting adjustment can produce a large score. Keep semantic assertions and human review in the loop for high-value assets.
The launch's latency improvement is an "up to" figure against Images 2.0, not a guarantee for every prompt, resolution, region, or quality level. Benchmark from the same infrastructure and record tail latency, not just the median.
A small gold set can still overfit your current catalog. Add every production failure as a permanent regression case, and refresh the suite when your product expands into new languages, visual styles, or output formats.
Finally, stronger edit fidelity can increase confidence without eliminating misuse or rights concerns. Keep consent, data handling, content policy, and human approval controls separate from visual quality tests.
The Bottom Line
GPT-Image-2.5 gives teams a useful speed-versus-precision choice, but the winning model is workload-specific. Before migrating, build a versioned edit-drift suite, score instruction success separately from preservation, test a multi-turn degradation ladder, and calculate cost per accepted asset. Start Flare on low-risk volume, earn Sunburst's place on precision-sensitive edits, and keep a dated snapshot plus rollback route until the evidence is stable.
