GPT Image 2.6 vs GPT Image 2.5: what changed and which to use

GPT Image 2.6 vs 2.5 at a glance

Token counts and API times are from CellCog's launch-night test (80 generations, one machine, 5 November 2026); token counts repeated exactly across reruns, so they are the firmest numbers. Prices and features are from OpenAI's model pages and image generation guide.

Cost: the same from high quality up, 3–4 times more below it

Both versions bill image output at $30 per million tokens, so the price of an image depends only on how many tokens it uses. From high quality upward the two versions use exactly the same number; at low and medium, 2.6 uses far more, because its cheapest tiers now do work closer to 2.5's medium-to-high range. CellCog measured a 1024×1024 image at each level:

• At high, xhigh and max, 2.6 costs the same as 2.5 and is faster, so the upgrade is free.
• At low quality 2.6 costs about four times as much, and at medium about three times. CellCog found the 2.6 low-quality image cleaner, but not by enough to justify the price for drafts.
• 2.5 also has a half-price Batch API row; 2.6 does not yet. CellCog estimates bulk work on 2.5 through Batch costs about a quarter of the same work on 2.6.
• Bigger sizes use more tokens on both versions; the ratio between them is what matters when you switch.

On GPT Image Hub the token gap does not apply: both versions run on the Saver route at a flat price per image, 10 credits for 2.6 and 8 for 2.5, at every size and quality. Picking high or max costs nothing extra there, and new accounts get 25 free credits.

Speed: Flare doubled, Sunburst did not

Through OpenAI's API, 2.6 Flare was roughly twice as fast as 2.5 Flare on every prompt CellCog tried, and the gap grew with quality (45.1 seconds down to 18.0 at max). Sunburst went the other way: 2.6 Sunburst was as fast or a little slower than 2.5 Sunburst, by about 17% on a data graphic and 27% on a mascot.

These are launch-night numbers from one location, so treat them as a direction rather than a benchmark. Through GPT Image Hub's Saver route the difference mostly disappears, because the route's queue and image delivery add more time than the model saves. In our runs on 5 November 2026, every model took 13–27 seconds:

Same prompt on both versions: side-by-side images

We sent the same prompt, size and quality to all four models on GPT Image Hub on 5 November 2026. Each caption gives the time and what we saw. The prompts are in our prompt guide.

Text-heavy poster

All four spelled the title and the full date line correctly. The differences are in finish: 2.6 Flare has the strongest contrast, 2.6 Sunburst the softest, most painterly light. For posters, 2.5 was already reliable.

Six-step infographic

Every version got all six step labels right and wrote sensible captions. The flaws were visual: 2.6 Flare left faint smudges on the title and 2.5 Sunburst left a third of the page empty, while 2.5 Flare and 2.6 Sunburst produced clean layouts. 2.6 did not beat 2.5 here; each version had one strong and one weaker result.

Mobile app screen

Every requested string was correct on all four. Both Flare versions added text nobody asked for ("Sunny" on 2.5, "Partly Cloudy" and a high/low line on 2.6), while both Sunburst versions kept to the brief. That difference is between Flare and Sunburst, not between 2.5 and 2.6.

Product photo

Both Flare versions made a convincing photo with the requested empty space on the right. Side by side they are hard to tell apart; for photographic work the upgrade is about speed and price, not looks.

Edits: both versions keep the rest of the image

We edited the 2.6 Flare product photo above with the same instruction on all four models: swap the background, keep everything on the counter.

All four kept the press, cup, steam, napkin and beans in place. The split was again by model type: both Flare versions left a plant and jar from the old window, both Sunburst versions cleared them for the plain wall the prompt asked for. Times ran 16–18.7 seconds.

CellCog's API tests point the same way. In a three-step edit chain on a mascot (add a scarf, add a skateboard, raise an arm) both versions kept the character's identity; 2.6 held a small hoodie stripe slightly better, but 2.5 did it at 42% of the cost. A markup-guided edit that turned a pair of sunglasses amber changed only the lenses on both, for $0.022 on 2.5 and $0.052 on 2.6.

What only 2.6 can do

Markup-guided edits arrived in the API with 2.6 but work on the 2.5 models too, so they are not a reason to switch on their own. Details of every new feature are in our GPT Image 2.6 guide.

Should you switch? A decision table

Not sure which of the 2.6 models to pick? See GPT Image 2.6 Flare vs Sunburst.

Moving from 2.5 to 2.6 in the API

• Change the model ID (gpt-image-2.5-flare to gpt-image-2.6-flare, and the same for Sunburst); the endpoints, sizes, quality levels and response format stay the same.
• Check the quality you request. If your pipeline runs at low or medium, your image bill can triple or quadruple; at high and above nothing changes.
• Keep Batch jobs on 2.5 until OpenAI adds a 2.6 Batch price.
• Your prompts carry over. Test a handful on both versions before switching everything, since 2.6 renders differently even at the same token count.
• Leave complexity at its default (high) unless an edit changes more than you asked; then try low.

curl https://api.openai.com/v1/images/generations \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2.6-flare",
    "prompt": "A concert poster for a jazz night titled \"BLUE HOUR\"",
    "size": "1024x1536",
    "quality": "high"
  }'

Sources

Specifications are from OpenAI's documentation as of 5 November 2026. API timings and token counts are from CellCog's independent launch-night test. Images and Hub timings are our own, from GPT Image Hub's Saver route on 5 November 2026.

• OpenAI: GPT-Image-2.6 Flare model page
• OpenAI: GPT-Image-2.6 Sunburst model page
• OpenAI image generation guide
• TestingCatalog: OpenAI launches Images 2.6 with GIFs and image reference
• CellCog: GPT Image 2.6 vs 2.5, 80 generations on launch night