GPT Image 2.5 مقابل GPT Image 2: نفس سعر التوكن، تعديلات أدق، غير مثبت

When OpenAI shipped GPT Image 2.5 on September 8 — under the consumer name ChatGPT Images 2.5 and on the API as two separate identifiers, gpt-image-2.5-flare and gpt-image-2.5-sunburst — it priced the new line on exactly the same token rate card as GPT Image 2, the model it is built to succeed. That makes this the rare image-model upgrade that costs nothing on paper: the same $8 per million image-input tokens and $30 per million image-output tokens, while OpenAI claims sharper details, more precise multi-turn editing, and up to 50% lower latency than GPT Image 2. The tension is that every one of those claims is vendor-reported, and on the independent leaderboards it is still GPT Image 2 — out since April 21, 2026 — that holds the visible crowns. This head-to-head is about whether a same-price promise you cannot yet verify is worth acting on today, or whether the model with the receipts stays the better production default.

What actually shipped: two models, not one

OpenAI replaced one API model with two siblings, and the split is the first thing to understand about this upgrade. GPT-Image-2.5 Flare is the fast default — OpenAI positions it for high-volume generation and says it delivers higher quality than GPT Image 2 at up to 50% lower latency. GPT-Image-2.5 Sunburst is the precision model, deliberately slower per generation and aimed at editing work "where editing precision matters most": campaign creative, product imagery, anything an edit has to survive scrutiny. On the consumer side the same release brought ChatGPT Images 2.5 to ChatGPT, ChatGPT Work and Codex users on day one, with new product features — a Sketch tool, layout templates, on-image comments, prompt sharing — that sit in the chat product rather than the API.

Two API details matter for anyone comparing against GPT Image 2. First, the 2.5 models add xhigh and max to the quality ladder — GPT Image 2 tops out at high — and default to auto, so the documented parameter surface is strictly wider than the previous generation's. Second, transparent backgrounds are now a first-class parameter: both 2.5 identifiers accept background set to transparent, opaque or auto, with PNG or WebP output. Transparent output was a sore spot on GPT Image 2, which did not ship it and only gained a preview flag later, so day-one transparency support is one of the more concrete deltas in this release. OpenAI also notes the 2.5 rollout is additive — GPT Image 2 stays listed in the API — which matters for the decision below, because nothing forces a migration.

The head-to-head, dimension by dimension

• Release — GPT Image 2, April 21, 2026; GPT Image 2.5, September 8, 2026.

• API lineup — one identifier, gpt-image-2, versus two: gpt-image-2.5-flare (fast, high-volume) and gpt-image-2.5-sunburst (precision editing).

• Token pricing — identical published rate card: $8 per million image-input tokens, $30 per million image-output tokens, with text input billed separately on both.

• Quality tiers — GPT Image 2 supports low, medium and high; the 2.5 pair add xhigh and max and default to auto.

• Editing behavior — GPT Image 2 tends to re-render: edits can rebuild the image and drift from what you wanted kept. OpenAI's pitch for 2.5 is targeted edits that change only the requested region and stay consistent across successive turns.

• Transparent backgrounds — GPT Image 2 gained them only as a later preview flag; GPT Image 2.5 documents them as a standard background parameter from launch.

• Independent status — GPT Image 2 tops Artificial Analysis's text-to-image arena and sits inside its top two for image editing; neither 2.5 model appears on any independent leaderboard yet.

The price story: same rate card, unknown per-image cost

The identical token rates are the single most decision-relevant fact in this matchup, and they need one caveat. A rate card is not a price per image — per-image cost is the rate multiplied by the tokens a model actually burns, and OpenAI has not published token consumption or a cost calculator for the 2.5 snapshots. For GPT Image 2 the per-image figures are known because independent trackers have measured them: roughly $0.21 for a high-quality 1024×1024 image, about $0.05 at medium and around $0.006 at low, per Artificial Analysis. For GPT Image 2.5 the honest answer is that nobody outside OpenAI knows yet whether Flare and Sunburst land above or below those numbers at a given quality setting. A model that shares the rate card but burns more tokens per image is not actually a same-price upgrade, so per-image cost is the first thing to verify before committing volume.

The pass-through point is where a routing layer earns its keep in this specific comparison. GPT Image 2 is on OrcaRouter today at the provider list price — the same $8/$30 per million tokens, with no markup — so the rate card you see is OpenAI's own, and the day OpenAI changes it the change is live on our side without a renegotiation. Because the 2.5 identifiers share that rate card, the cost question is purely about token efficiency, not about which vendor you call. And for trialing an unproven model, a router is the safe way to do it: keep one OpenAI-compatible endpoint, point a slice of traffic at gpt-image-2.5-flare when it appears on an upstream list you already call, and leave automatic failover to GPT Image 2 so a bad batch cannot take down the whole generation queue.

Latency: where the two new models diverge from each other

The headline latency figure — up to 50% lower than GPT Image 2 — belongs to Flare and is OpenAI's own measurement. Sunburst is the counterweight: OpenAI says it deliberately trades speed for precision and takes longer per generation. That means the real comparison is not simply "2.5 is faster than 2"; it is that OpenAI has split the old single model into a fast lane and a careful lane, and the fast lane carries the latency promise while the careful lane carries the editing promise. Early launch partners have reported large practical speedups on Flare — the integrations OpenAI chose to highlight described faster pipelines and better-behaved edits — but those are vendor-picked early customers, not an independent benchmark. Treat the 50% figure as directional until someone with no commercial interest measures time-to-image on both models at the same quality setting.

Editing is the real battleground — and it is unverified

If you read why OpenAI says it built 2.5, the whole release is a response to GPT Image 2's most-reviewed weakness: edits that re-render rather than surgically preserve. The recurring complaints about GPT Image 2 were that a small requested change could rebuild the whole image, that a person or product drifted across a multi-turn edit, and that text or backgrounds degraded on the second and third pass. The 2.5 launch materials attack exactly those failure modes — subject preservation from a reference photo, targeted edits that leave the rest of the frame intact, and multi-turn consistency where earlier changes survive later ones. Early integrations echo the claim: the editing-focused partner OpenAI showcased says the model "understands what not to change." None of that has been reproduced by an independent evaluation as of this writing, so the editing gap between GPT Image 2 and GPT Image 2.5 is, for now, a promise backed by demos rather than a measured result.

The benchmark gap and what to watch

This is the one place where the matchup is lopsided in GPT Image 2's favor. GPT Image 2 is a known quantity on the independent boards: it holds the top spot on Artificial Analysis's text-to-image arena and ranks inside the top two on its image-editing arena as of early September 2026 — positions that drift as new models enter, but they are real, blind human-preference scores. GPT Image 2.5 appears on no independent leaderboard yet. There is one wrinkle worth watching: anonymous OpenAI-provenance checkpoints ran in the image arenas through August under the codenames mona-lisa-1 and luna-lisa-alpha, and observers tied them to the 2.5 line before launch. If those accounts hold, the September 8 models may already have a head start in arena Elo that will surface once the boards add them under their real names.

The things to watch, concretely, are three. First, per-image token consumption on Flare and Sunburst at each quality tier — that decides whether the same-price claim survives contact with a real bill. Second, an independent latency or throughput measurement of Flare against GPT Image 2, since the 50% figure is the upgrade's biggest practical lure. Third, where Sunburst lands on an image-editing leaderboard once it is added — that is the result that would tell you whether OpenAI actually closed the editing gap it is claiming against not just GPT Image 2 but the dedicated editors ranked above it.

من الذي يجب أن يختار ماذا

• New project, high-volume generation, quality matters — start on GPT-Image-2.5 Flare. It shares GPT Image 2's rate card, adds quality tiers, and on paper is faster; run your first batches in parallel with GPT Image 2 and compare on your own images before committing the pipeline.

• Production already on GPT Image 2, nothing broken — there is no urgency. The rollout is additive, GPT Image 2 stays listed, and you would be trading a model with independent leaderboard results for one whose claims are still vendor-reported. Migrate when independent scores or your own evals justify it, not on launch day.

• Edits that have to survive scrutiny — GPT-Image-2.5 Sunburst is the interesting model, but it is also the one with the least public evidence. If a deadline forces the choice, test it on a representative edit set first; otherwise wait for per-image cost and an editing-leaderboard appearance.

• Budget-sensitive batch work at low or medium quality — stay on GPT Image 2 until the 2.5 models' token consumption at those tiers is documented. At $0.006-to-$0.05 per image the incumbent is cheap and known; a same-rate-card successor could still surprise you on the way up.

The bottom line is a verdict with a date on it. GPT Image 2.5 is the better bet on paper — same published price, a wider quality ladder, day-one transparency, and an editing story aimed squarely at GPT Image 2's known weaknesses — but every advantage is OpenAI's word, while GPT Image 2 is the one with independent receipts. For a one-line decision: trial GPT Image 2.5 now if you can measure it on your own images, and keep GPT Image 2 as the production default until the leaderboards and the token meter give the new line a verdict of their own.

For the models OrcaRouter does route today — browse OrcaRouter's model catalog — including openai/gpt-image-2 — the price you see is the provider list price, passed through with no markup.