Images 2.6 vs GPT Image 2.5: The Rate Card Didn't Move

OpenAI's Images 2.6 arrived on 4 November with the same price list as GPT Image 2.5: $5 per million text input tokens, $8 per million image input tokens and $30 per million image output tokens, with cached input at $1.25 and $2 respectively. That is the most useful fact in the upgrade. When a vendor ships a new generation and does not touch its rate card, it is telling you the improvement is in what the model can do and how fast it does it, not in the economics. So the question for anyone already on GPT Image 2.5 is narrow: do the new capabilities earn a migration, or is this a version you adopt on your next greenfield project and ignore otherwise?

That question has a cleaner answer than most model upgrades, because the two API models — GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst — keep the tier structure, the endpoints and the output contract of their predecessors, GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. Almost nothing breaks. That is also why the decision is easy to get wrong in both directions: doing nothing is defensible, and switching is cheap enough that not switching on latency grounds alone is hard to justify.

What is genuinely identical

The boring list first, because it is longer than the release notes suggest.

• Price — $5 / $8 / $30 per million tokens, unchanged, on both tiers.

• Endpoints — /v1/images/generations and /v1/images/edits, unchanged.

• The tier split — a speed model and a quality model, priced identically, unchanged since Images 2.5 introduced it in September.

• Modalities — text and image in, image out. No video, in either generation.

• The product surfaces — ChatGPT, Codex and the API. Images 2.5 already spanned all three.

If your integration is a thin wrapper around the images endpoint with a fixed quality setting, 2.5 to 2.6 is a model string and a regression run. There is no new authentication, no new response shape, no pricing migration.

What is genuinely new

Four parameters and one output format. All five are API-level, not just ChatGPT features.

• Complexity. Low, medium or high, separate from quality. It controls how much planning happens before drawing. For a flat product cut-out, low is correct and cheaper; for a scene with six constraints that must all hold, high buys planning room. This is a real dial that GPT Image 2.5 did not expose.

• Markup-guided editing. A transparent PNG overlay marks the region to change, drawn on the source image. It cannot be combined with a mask. OpenAI's changelog says this works on supported GPT Image 2.5 and 2.6 models — so part of the editing upgrade is available to the older generation too, which is worth checking before you migrate for it.

• SVG output. Vector export from the image call. New, and the one feature with no equivalent in 2.5.

• Animated GIF output. A loop from a still, with playback speed and per-frame editing in ChatGPT, and GIF as an output format on the API. Not accepted as input.

• Reference images in ChatGPT. Save a person, pet or character once and reuse it across turns without re-uploading.

The headline OpenAI led with is the GIF. The feature a production team will notice first is complexity, because it is the first time the model has let you trade planning for latency explicitly rather than implicitly through quality.

The speed claim, unpacked

OpenAI says generation is about 50% faster. The specifics behind that number are more interesting than the average: Flare at low quality is quoted at roughly six seconds per image, and Design Arena's comparison, as reported on launch, put Flare at about 2.6× faster than its predecessor and Sunburst at about 1.15×.

Those two multiples tell different stories. A 2.6× speed-up on the speed tier changes what is possible in an interface — six seconds is inside the window where a person keeps looking at the screen. A 1.15× speed-up on the quality tier is a rounding error for batch work and a mild convenience for interactive editing. If your workload is Sunburst-shaped, do not migrate for speed; migrate for the consistency claim, or don't migrate.

What 2.5 still has that 2.6 does not: a score

GPT Image 2.5 Sunburst is the top model on Artificial Analysis's text-to-image board at 1,224 Elo, with Flare second at 1,208; on the editing board Sunburst leads at 1,199 and Flare sits third at 1,183. Neither generation's list price appears as a bargain there — $210.70 per thousand images at the max setting is the most expensive row near the top.

Images 2.6 has no row. We checked both boards on 5 November, the day after launch, and GPT-Image-2.6 Flare and Sunburst are absent. The only ranking claim in circulation is OpenAI's own, that Sunburst took first place on Design Arena and displaced 2.5 Sunburst. That is plausible and specific, but it is a vendor-cited arena position without a published Elo we could verify.

So the honest state of play is: the model you can measure is the old one, and the model you are being asked to adopt is unmeasured. That is normal for a one-day-old release and it is not a reason to avoid it — but it does mean the migration should be gated on your own outputs, not on a leaderboard.

What an edit costs on either generation

Because the token rates are identical, a worked example carries over unchanged. Take a typical edit: 900 text input tokens, a 1,200-token image input and a 4,000-token image output. That is $0.0045 for text, $0.0096 for the image input and $0.12 for the output — $0.1341 per edit, on 2.5 or 2.6.

Now apply the only saving that matters. If complexity lets you drop from a high-planning call to a low-planning one on the 80% of edits that are simple, the output token count falls and the per-edit cost falls with it. If Sunburst's consistency claim means one fewer re-roll in five, the effective cost per usable asset drops by a fifth. Both are plausible; the first is directly controllable today, the second is a claim you have to test.

The migration nobody talks about: which version you can actually call

GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst are on OrcaRouter today, at OpenAI's list price with 0% markup, on the same key as the rest of our 200-plus model catalogue, with automatic failover across providers. GPT-Image-2.6 is not in our catalogue at the time of writing. That asymmetry is the practical decision for a lot of teams this week: the upgrade is real, but the version with a routable, failover-protected, single-key path is still 2.5.

If you are a single-provider shop, this is irrelevant and you should test 2.6 directly. If you are running image generation as a dependency rather than a feature — a marketplace that renders product shots, an app that generates avatars — the calculus is different. A model that is one day old has no uptime history, no rate-limit behaviour under your load, and no independent quality score. Running it behind a router that can fall back to the previous generation is how you adopt it without betting the pipeline on it. That is the argument for keeping both keys live rather than for waiting.

There is also the price-change case. When OpenAI does move the rate card — and image pricing has moved twice in eighteen months — a pass-through router reflects it the same day. A direct integration reflects it whenever the billing logic is updated.

Who should upgrade, and who should sit still

Upgrade now if: your product renders images in an interface a person watches (the Flare speed-up is the largest single change in the release), you need vector output, or you are building motion/GIF features and do not want a second vendor. The API surface is compatible and the spend is unchanged.

Test before upgrading if: Sunburst's edit consistency is the reason you are interested. That is the claim with the least evidence and the highest value, and a twenty-image regression set will tell you more than any launch post.

Stay on 2.5 if: your workload is batch, quality-tier and cost-sensitive. The rate card did not move, the quality tier got 1.15× faster, and the old model has a published score at the top of the board. There is no financial case to migrate this week, only a capability case, and if you do not need SVG, GIF or complexity, the capability case does not exist.

The thing to watch is the first independent 2.6 Sunburst number. If it clears 1,224, "stay on 2.5" becomes a decision you have to justify. If it lands near it, this was a feature release on a model that was already at the top, and the rate card is telling you exactly that.