Migrating from GPT Image 2.5 to 2.6: new parameters, new formats, and doc differences to verify honestly
OpenAI launched Images 2.6 on November 4, 2026, with API models gpt-image-2.6-flare and gpt-image-2.6-sunburst. The Changelog summary: adjustable complexity, SVG generation, animated GIF generation, and markup-guided editing for supported GPT Image 2.5 and 2.6. The product announcement claims about 50% faster generation on average, Flare about six seconds at low quality, and Sunburst emphasizing detail and edit consistency at medium/high quality. Migration is not merely changing a model string—it means knowing what is new in 2.6, what was partly available in 2.5, and whether guide-body examples match the model-page Endpoints table.
👉 Try the GPT Image 2.6 API on PoYo: per-image pricing and online playground.
Capability deltas versus 2.5
Per the image generation guide’s “GPT Image 2.6” section, Flare and Sunburst both support via the Image API: complexity control, markup-guided edits, SVG output, and animated GIF output. Most code examples in the guide body still demonstrate common generation and edit flows with GPT Image 2.5 model IDs and state that “the following existing examples use GPT Image 2.5”—do not treat sample gpt-image-2.5-* IDs as 2.6 defaults.
| Dimension | 2.6 doc highlights |
|---|---|
complexity | low/medium/high; defaults to high when omitted; independent of quality; no separate charge |
| GIF | generations only; frame_count, frame_duration_ms; no streaming |
| SVG | generations only; n=1; no streaming; not valid edits input/output |
| markup | transparent overlay on edits; cannot combine with mask or prompt_rewriter; supported on 2.5/2.6 |
| Two tiers | Flare for fast iteration; Sunburst for detail and edit consistency |
The image prompting guide’s migration checklist: save baseline prompts and samples; prior 2.5 Flare → test 2.6 Flare first; prior 2.5 Sunburst → test 2.6 Sunburst first; on first comparison freeze prompts, references, size, quality, and output format, leaving complexity at the model default; change only quality or complexity one at a time; roll out by workflow in batches and track latency, failures, retries, and accepted per-image cost. Edit requests should omit legacy input_fidelity; GPT Image 2.6 does not support reasoning.summary.
Responses API: read the guide examples and model pages separately
The image generation guide still explains multi-turn conversational image tools via the Responses API, with 2.5 identifiers in sample image-tool model fields. By contrast, the GPT‑Image‑2.6 Flare and Sunburst model pages Endpoints tables both mark v1/responses as Not supported, while marking v1/images/generations and v1/images/edits as Supported. The 2.6 section also stresses that complexity, markup, SVG, and GIF are provided “via the Image API.”
Production migrations should therefore default to wiring 2.6 into Image API generations and edits—and not assume a 2.5-era Responses tool config can simply swap in a 2.6 ID. If the business strongly depends on multi-turn conversational editing, re-verify against the model-page Endpoints table as authoritative rather than extrapolating from guide paragraphs that still exemplify 2.5 Responses. That is a real documentation difference, not a verbal dispute over “who mistyped”—integrators should gate production config on the model page.
Migration acceptance suggestions
Split acceptance into three layers. Capability: on business corpora, verify whether GIF/SVG/markup/complexity truly shorten iteration or reduce mistaken edits. Quality: compare 2.5 and 2.6 counterpart tiers on the same prompts for text, identity, product geometry, and alpha channels. Integration: confirm organization verification, safety_identifier, error handling (do not blindly retry moderation blocks), and that configs contain no unsupported Responses bindings. On pricing, 2.6 tier unit prices verified here match—do not assume “the faster tier is cheaper.”
In practice, watch parameter-habit changes. If a 2.5 workflow never set complexity explicitly, continuing to omit it after moving to 2.6 lands on default high, which may lengthen edit latency and add unwanted changes; official advice for most edits is instead to set low explicitly. GIF and SVG exist only on generations—edit pipelines that assume “any format can round-trip” will hit a wall on 2.6. Markup covers supported 2.5 and 2.6 but is mutually exclusive with mask and prompt_rewriter; migration scripts need field mutual-exclusion checks. The prompting guide also reminds: omit input_fidelity on edit requests; 2.6 does not support reasoning.summary. These are common pits that compile yet fail or drift at runtime.
A/B tests should freeze size, quality, and output format, move only model ID or complexity at a time, and save intermediate samples. If the business uses ChatGPT product-side Images 2.6 and the API together, account separately: product name ≠ model ID, and quota/verification do not auto-sync. Scale with small-traffic validation of failure rate and accepted per-image cost before expanding default routing; keep a 2.5 rollback window until latency and quality curves stabilize on business corpora.
Bottom line: the stateable 2.6 deltas versus 2.5 are speed and stronger two-tier positioning, plus the control surface of complexity, GIF, SVG, and markup. Migration success depends on same-condition comparisons per the prompting guide, and on facing the documentation fact that “the guide still shows 2.5 Responses while model pages mark 2.6 Responses unsupported”—then treating the model page as production authority.