VSModels
Back to blog

GPT Image 2.6 vs 2.5: What Changed and Should You Switch?

Same token rates, not the same bill. Low and medium 1024 images cost about 4.2× and 3.3× more on 2.6. What changed, what didn't, and when to stay on 2.5.

Nov 10, 2026VSModels

The quick answer. Move API work to GPT Image 2.6 if you need GIFs or SVGs, or if you already run high, xhigh, or max quality and want Flare's speed at the same output-token bill. Stay on 2.5 for low- and medium-quality volume, and for anything that depends on Batch. Do not read "same price" as "same cost per image," and do not declare 2.6 a quality upgrade in general: Artificial Analysis ranks 2.6 first at max quality, while an early Arena snapshot ranked 2.5 Sunburst ahead of 2.6 at Medium.

VSModels has not run its own generations for this comparison. Official pages were re-read on 10 November 2026 (UTC+8). Third-party timings and preference scores are cited with links; many are one-shot or few-prompt tests, and several come from platforms that sell these models. For GIF, SVG, markup, and reference-image how-to, see the companion GPT Image 2.6 guide.

Know which models you are comparing

Both generations ship as a pair:

  • GPT Image 2.5 — gpt-image-2.5-flare and gpt-image-2.5-sunburst, released September 2026.
  • GPT Image 2.6 — gpt-image-2.6-flare and gpt-image-2.6-sunburst, released 4 November 2026 in ChatGPT, Codex, and the API (launch post).

Flare is the fast everyday model. Sunburst is the detail and edit-consistency model. A "2.6 vs 2.5" claim that does not name the tier, the quality setting, and whether the call went through the API or ChatGPT is not comparing the same product. ChatGPT also does not show which model or quality produced an image, so consumer-app timings are not API benchmarks.

What changed, and what didn't

Changed, with documentation or consistent measurements behind it:

  • New outputs: animated GIF (generations and edits) and SVG (generations only). SVG cannot be used as an edit input.
  • A new complexity parameter (low / medium / high, default high, no separate fee). quality is unchanged as a separate control (low, medium, high, xhigh, max, auto).
  • ChatGPT can save named reference images and recall them later (OpenAI's community announcement; TechRadar documented an @Chester example). The API still has no dedicated character-lock parameter; edits still take up to 16 reference images.
  • Cost at low and medium quality rose, even though the per-token list price did not. Batch is listed for 2.5 and not for 2.6. Responses API is listed as not supported on the 2.6 model pages.
  • Speed split by tier: Flare is faster on API tests; Sunburst at default high complexity is often slower than 2.5 Sunburst.
  • Several testers describe 2.6 pictures as richer, especially people — and also more willing to add text or props the prompt did not ask for.

Did not change:

  • Per-token rates: text input $5 / million (cached $1.25), image input $8 (cached $2), image output $30, confirmed on 10 November 2026 from the pricing page and both 2.6 Flare and Sunburst model pages. 2.5 uses the same standard rates.
  • Markup-guided edits already work on supported 2.5 models. They are not a 2.6 exclusive (changelog; image generation guide).
  • On-image spelling, in the published tests, was already strong on 2.5. 2.6 does not open a new gap there.
  • Provenance watermarks: C2PA and SynthID, per the system card.

Cost: the same-price trap

This is the claim most launch coverage got wrong. The official guide states it directly: "Equal token rates don't mean equal cost per image" (calculating costs).

The on-page calculator, re-read 10 November 2026, defaults to GPT Image 2.6, quality low, 1024×1024, and shows 823 output tokens / $0.02469 at $30 per million image-output tokens. That estimate excludes prompt tokens, input-image tokens on edits, and 100 extra output tokens per streamed partial image.

Using the same official formula, image-output estimates for a 1024×1024 still are:

  • Low — 2.5: 196 tokens, $0.00588. 2.6: 823 tokens, $0.02469. About 4.2×.
  • Medium — 2.5: 439 tokens, $0.01317. 2.6: 1,449 tokens, $0.04347. About 3.3×.
  • High — both 1,756 tokens, $0.05268.
  • Xhigh — both 3,122 tokens, $0.09366.
  • Max — both 7,024 tokens, $0.21072.

The 4.2× and 3.3× figures are the multipliers the official calculator itself applies to the 2.5 estimate. High and above are not multiplied. Artificial Analysis lists $210.72 per 1,000 max-quality images, which is this max row exactly.

API bills reported by third parties land on almost the same numbers. CellCog — a studio that sells an image agent, then running 2.5 — measured 824 vs 196 tokens at low (one token above the calculator) and 1,758 vs the calculator's 1,756 at high, across 80 generations, with the low cell repeating on a rerun. OmniaKey and Hacker News users reported 1,449 vs 439 at medium, matching the calculator. Ignore older posts that still price a 2.6 low 1024 image at 2.5's roughly $0.006.

Then add Batch. The pricing page still gives 2.5 (and gpt-image-2) a half-price Batch row. 2.6 is absent. For bulk low- or medium-quality work, 2.5 Batch remains the cheaper official path, by a wider margin than the 4.2× / 3.3× still-image gap alone.

Speed: only Flare got faster

OpenAI's launch post says Images 2.6 "generates images 50% faster than before" and that Flare can finish in as little as six seconds at low quality. The baseline for "50%" is not defined.

Published API timings, all small and machine-specific:

  • Flare is faster. CellCog timed 2.6 Flare at 18.0 seconds against 45.1 for 2.5 Flare at max quality, and about twice as fast at every quality in that test. Kayaweb (61 self-funded API images, $2.05; three runs, clock includes network) timed Flare high at 10 seconds versus 16.8.
  • Sunburst, at the default complexity=high, is often slower. CellCog reported 2.6 Sunburst 17% slower on one prompt and 27% slower on another versus 2.5 Sunburst. Kayaweb timed Sunburst high at 37.5 seconds versus 24.5, and said setting complexity=low brought the time back in line with 2.5 without changing tokens. If your pipeline already uses Sunburst for edit precision, 2.6 is not automatically a latency win.
  • ChatGPT did not show a meaningful speed-up in PhotoWorkout's test: 2.6 Sunburst averaged 30.1 seconds against 31.9 for 2.5, about 6%, through a ChatGPT login rather than the API.

Quality: two boards, two answers

Do not merge these scores.

As of 10 November 2026, Artificial Analysis evaluates max quality. GPT Image 2.6 Sunburst leads text-to-image at 1259 Elo, then 2.6 Flare at 1228, then 2.5 Sunburst at 1227. On the editing board, 2.6 Sunburst is first at 1209, with 2.6 Flare at 1206 and 2.5 Sunburst at 1199. The gaps over 2.5 are modest, and they describe max-quality preference, not low-quality drafts.

Arena's text-to-image board, in a 6 November 2026 snapshot, listed 2.6 only at Medium and marked those rows preliminary (about 2,972 votes for 2.6 Sunburst). In that snapshot, 2.5 Sunburst led at 1429, with 2.6 Sunburst (Medium) at 1409, 2.5 Flare at 1402, and 2.6 Flare (Medium) at 1383. That snapshot was not re-opened for this article. It is enough to say the two public boards disagree, and that they are not measuring the same quality setting.

DesignArena snapshots in the same week conflicted with one another and are not used as a ranking here.

Same-prompt studio tests are smaller still, and several authors sell the endpoint:

  • PixelDojo (eight prompts, one take) described 2.6 as richer, especially on people, with text equally accurate and Sunburst slower.
  • BulkImagen (five text-to-image prompts and two reference tests, one take, plus GPT Image 2) said all three generations spelled every word and kept references, and that 2.6's difference was adding things they never asked for.
  • Fontly compared only the two 2.6 tiers, not 2.5, and found no misspelling across 75 images. That supports "spelling is solved," not "2.6 beats 2.5 at type."
  • OpenAI's system card reports offline (not production) safety figures: Flare's not-caught rate rose from 1.49% (46 of 3,077) on the 2.5 baseline to 2.14% (66 of 3,077) on 2.6, with tighter behavior called out on some sexual categories; Sunburst's rate moved from 1.92% to 1.72%. Those are vendor evaluations, not a user-facing quality score.

Text, extras, and consistency

On spelling, the published evidence is that both generations already get short English and multilingual copy right in ordinary posters, menus, and info sheets. Fontly's 75-image 2.6 typographic study, PhotoWorkout's 29-string camera sheet (2.6 and 2.5), and the seller-side PixelDojo and BulkImagen sets all point the same way. The remaining typography issues are layout — leading, hierarchy, "does this look like the named font" — not missed letters. Dense nine-panel comics are a thinner, mixed anecdotal set and should not be over-read.

The clearer behavioral change is unrequested content. Kayaweb saw unrequested text on 6 of 10 Sunburst French posters. BulkImagen and PixelDojo report extra slogans, props, or addresses. Higher complexity, which defaults to high, is exactly the setting the official guide says can introduce unwanted changes. If your 2.5 prompts were written to be followed literally, budget a pass to strip extras, or try complexity=low on edits before you rewrite the prompt library.

Character consistency is not a solved leap. There is still no lock parameter. Flixly, which sells these models, reported a slight 2.6 Sunburst edge over 2.5 in a small same-prompt set; other write-ups call the two close. For comparisons against Google's Nano Banana line, use the existing pages rather than a four-prompt seller test: Nano Banana 2.1 vs Pro, Nano Banana 2.1 vs 2, Nano Banana 2 vs Pro, and Nano Banana vs Midjourney.

New formats 2.5 does not have

These are the clean feature wins for 2.6, with early-quality caveats. Details and request fields are in the GPT Image 2.6 guide.

  • GIF. Official range is frame_count 2–16 (default 9), not 64; CellCog got HTTP 400 asking for 32 frames. Output works on generations and on edits with still references. Published tests support short, single-character, fixed-camera loops and are weaker on complex motion and real-photo animation (PhotoWorkout; Hacker News reports of flicker).
  • SVG. Officially generations-only, n=1, not a valid edit input. In CellCog's SVG trial GPT Image 2.6 returned 9 of 48 files, 3 of 41 when the prompt had lettering, usually failing with HTTP 502; Recraft V4.1 returned 9 of 9 in that same write-up. Both sides of that comparison have a commercial interest. Successful 2.6 SVGs kept editable text and were smaller files. An undocumented svg.model parameter is a user report, not current official guidance.

If you do not need GIF or SVG, these features are not a reason to switch a working 2.5 pipeline.

Should you switch?

  • Switch Flare at high, xhigh, or max, if API latency matters and you are already paying the high-quality token bill. Same official output cost, and the published API timings show a large speed-up.
  • Switch if you need GIF or SVG, and treat both formats as early: keep prompts simple for GIFs, and expect SVG failures with lettering until reliability improves.
  • Stay on 2.5 for low- and medium-quality volume, where 2.6's official output-token estimates are about 4.2× and 3.3× higher, and for Batch jobs, which 2.6 does not offer.
  • Be cautious with Sunburst if latency matters. Default high complexity made 2.6 Sunburst slower than 2.5 Sunburst in the API tests cited above. Try complexity=low before you assume a regression in the model itself.
  • Don't switch on quality scores alone. AA at max and Arena at Medium disagree. Prompt-level tests so far show richer scenes and more unsolicited extras, not a spelling miracle that 2.5 lacked.
  • Don't switch for markup. It already runs on 2.5.

A practical split used by more than one of the testers above is: keep 2.5 (and its Batch row) for cheap iteration, and reserve 2.6 for final high-quality stills, Flare-speed work, or the new formats.

FAQ

Is GPT Image 2.6 cheaper, the same price, or more expensive than 2.5? Same per token. More expensive per image at low and medium quality (about 4.2× and 3.3× on the official 1024 calculator). The same at high and above. More expensive still for bulk work, because 2.6 has no Batch row.

Should I switch all traffic to 2.6? No. Switch where you need new formats, or where you already pay high-quality rates and can use Flare. Keep 2.5 for cheap drafts and Batch.

Is 2.6 better quality than 2.5? Unsettled. Artificial Analysis, at max, ranks 2.6 Sunburst first by a modest Elo margin. Arena, at Medium in an early snapshot, ranked 2.5 Sunburst higher. Those results cannot be averaged.

Why is Sunburst slower if OpenAI said 50% faster? The 50% figure is a vendor claim without a published baseline. Third-party API tests show the speed-up on Flare. Sunburst defaults to complexity=high; testers who lowered complexity reported latency closer to 2.5. ChatGPT-path tests showed almost no change.

Did markup, character lock, or text accuracy arrive only in 2.6? Markup is documented for 2.5 as well. There is no character-lock API parameter in either generation. Text accuracy was already high on 2.5 in the same kinds of poster and menu tests.

Can I use 2.6 with the Responses API? Not according to the current model pages. 2.6 is documented on /v1/images/generations and /v1/images/edits only.

Sources

Official, re-read 10 November 2026 (UTC+8) unless noted:

Third-party tests and boards (small samples except public leaderboards; several authors sell the model or a competitor):

Related on VSModels: GPT Image 2.6 guide.

Gaps

  • No VSModels-run generations and no bill from our own API key. Official calculator figures are estimates for image output tokens only.
  • CellCog's measured low-quality 2.6 usage was 824 tokens against the official calculator's 823; high was 1,758 against 1,756. We did not call the API to reconcile those one- and two-token gaps.
  • Arena figures are a 6 November 2026 snapshot and were not refreshed today. 2.6 was still preliminary there.
  • 2.5's Responses API support is described inconsistently on official pages (the 2.5 model page endpoint table versus the introductory copy). This article only relies on the clear 2.6 endpoint table and on 2.5's Batch row on the pricing page.
  • Almost every same-prompt quality test cited used one take or a handful of prompts.