ChatGPT Makes GIFs Now. The Cheapest Images on OpenAI’s API Cost Four Times More.

a close up of a computer screen with a purple background

OpenAI did not change the price of its image API on November 4. The cheapest images it makes now cost about four times as much anyway.

That is the part of the Images 2.6 launch that got buried under the headline feature, which is GIFs. ChatGPT can now turn a prompt, a photo or a rough sketch into a short looping animation, and it can remember a reference image (your face, your dog, your mascot) so you stop uploading it to every new chat. The release also ships two new API models, gpt-image-2.6-flare for speed and gpt-image-2.6-sunburst for detail, with OpenAI claiming generation is 50% faster than Images 2.5, which came out only two months ago.

If you make images inside ChatGPT, most of this is free upside. If you run anything on the API (thumbnail batches, a blog image pipeline, a product photo script), read the cost section before you change the model string.

The price per token stayed put. The tokens per image did not.

OpenAI’s image generation guide lists 2.6 at the same $30 per million output tokens as 2.5, and it says plainly that the models “can use different token counts for the same quality setting and share the same price per image output token.” That sentence is doing a lot of work. When eesel pulled the numbers out of OpenAI’s own cost calculator, the 2.6 models used roughly 4.2 times as many tokens at low quality and 3.3 times as many at medium. At high quality and above, the counts match.

Here is what that looks like for a square 1024×1024 image, using eesel’s calculator figures:

Quality Images 2.5 Images 2.6 Change
Low $0.0059 $0.0247 about 4.2x
Medium $0.0132 $0.0435 about 3.3x
High $0.0527 $0.0527 none
Max $0.2107 $0.2107 none

In batch terms, eesel puts 10,000 low-quality square images at about $59 on 2.5 and about $247 on 2.6. Same quality label, same per-token rate, four times the bill. These are calculator estimates rather than invoices, so check the usage field on your own responses, but two developers quoted by Fello AI said their real requests matched the calculator “within a few tokens.”

The likely explanation is that “low” on 2.6 is a better picture than “low” on 2.5, and you are paying for that. Fine. But a lot of creator automation sits on low or medium precisely because those images are disposable: draft thumbnails you pick from, placeholder art, A/B variants. Those are the workloads that just got more expensive. Two more details matter here. OpenAI’s Batch price table had no 2.6 row at launch, so the half-price batch discount only applies to 2.5 for now. And the new complexity setting (how much the model plans before it renders) carries no separate fee, according to the guide, and defaults to high. Dropping it to low for simple background swaps costs nothing to try.

What a 2.6 GIF actually is

The ChatGPT version is simple: describe a motion, upload an image or sketch, adjust the speed, and edit individual frames. The API version has real numbers, and they tell you what this feature is for. Per OpenAI’s guide:

  • frame_count runs from 2 to 16 frames, default 9.
  • frame_duration_ms runs from 20 to 2,000 milliseconds per frame, default 200.
  • The defaults produce a 1.8 second loop. OpenAI’s own example is a monkey holding a lantern that returns to its starting pose after 1.8 seconds.
  • GIFs work on both the generations and edits endpoints, but not with streaming, partial images or custom compression.
  • All frames share one pixel budget, so a large frame size forces fewer frames.

Sixteen frames is a hard ceiling. Even stretched to the slowest setting, that is a slideshow, and at a watchable speed you get a couple of seconds of motion. This is a sticker, a reaction, a waving mascot, a loading animation. It is not a clip. OpenAI’s launch partners point the same way: OpenArt for social reactions and memes, and a startup called Town for an animated typing indicator, per eesel’s writeup.

Early testers have been blunt. One Hacker News commenter quoted by Fello AI said “the splotchiness on that simple chess animation makes this GIF feature a nonstarter,” while others found simpler subjects held up much better. Consistency across frames is the hard part of any animation, and a model that redraws each frame is going to wobble on detail.

OpenAI has not published a per-GIF price, a per-plan GIF limit, or a maximum length for GIFs made in ChatGPT. Pasquale Pillitteri, citing heise, also reports the feature is rolling out gradually, so it may not be on your account yet.

The useful job for a two-second loop

The best creator use I can see is not the GIF itself. It is the test before the expensive thing.

AI video is billed by the second, and a bad take costs the full clip. I walked through what that does to retries in the Kling 4.0, Wan 3.0 and Seedance 2.5 comparison. A nine-frame loop inside a ChatGPT plan you already pay for is a cheap way to check whether a character’s pose, silhouette and gesture read before you spend video credits. OpenAI even pitches “storyboards for testing character movement before video production” as a use case, per Indian Television’s launch report. It sits in the same spot as blocking a shot in 3D before rendering, which is the idea behind Seedance 2.5’s previsualization mode.

The other honest use is the small stuff creators make by hand today: an animated logo sting for a newsletter header, a Discord or Twitch emote, a reaction loop for replies. If you need a web animation you can actually control (exact timing, small files, crisp at any size), a GIF is the wrong format and a tool like Lottie Creator 2.0 is the better path. GIFs are heavy and capped at 256 colors per frame, and that matters once you put one on a page.

Which creators should switch, and which should wait

ChatGPT users on any plan: nothing to decide. Images 2.6 is live on every tier, Free included, though OpenAI only describes plan limits in general terms (Free gets “limited” creation, Pro gets “unlimited”). Try saved references for a recurring mascot or product. Get consent before saving anyone else’s face, and know that OpenAI hasn’t said how many references you can keep or how to delete them. One more date to watch: Pillitteri notes ChatGPT and Codex move to a single shared usage limit in December, and nobody knows yet how much a GIF draws from it.

API pipelines at high quality or above: switch. The cost per image is identical, Sunburst ranked second on Arena’s text-to-image board at launch, per Fello AI, and you get the 50% speed claim (unverified, but free if true).

API pipelines at low or medium quality: do not swap the model string blindly. Run 20 of your real prompts on both versions, compare the usage field and the output, and decide whether the better “low” is worth roughly four times the price. For bulk drafts, staying on 2.5 with batch pricing is still the cheapest option OpenAI offers. If you want speed without the token jump, Krea 3’s sub-eight-second edits are worth pricing against it.

Anyone quoting the “4 billion images a week” figure: don’t lean on it. explainx traced it to OpenAI’s own announcement as relayed by The Verge, with no breakdown of GIFs, edits or API calls. It is a usage claim, not a benchmark.

Every 2.6 image carries C2PA metadata and a SynthID watermark, per eesel’s reading of the system card, so platforms that check provenance will see it as AI-made. For reaction GIFs that hardly matters. For a YouTube thumbnail or a client deliverable, it is worth knowing before you ship.

Ty Sutherland

Ty Sutherland is the Chief Editor of Full-stack Creators. Ty is lifelong creator who's journey began with recording music at the tender age of 12 and crafting video content during his high school years. This passion for storytelling led him to the University of Regina's film faculty, where he honed his craft. Post-university, Ty transitioned into the technology realm, amassing 25 years of experience in coding and systems administration. His tenure at Electronic Arts provided a deep dive into the entertainment and game development sectors. As the GM of a data center and later the COO of WTFast, Ty's focus sharpened on product strategy, intertwining it with marketing and community-building, particularly within the gaming community. Outside of his professional pursuits, Ty remains an enthusiastic content creator. He's deeply intrigued by AI's potential in augmenting individual skill sets, enabling them to unleash their innate talents. At Full-stack Creators, Ty's mission is clear: to impart the wealth of knowledge he's gathered over the years, assisting creators across all mediums and genres in their artistic endeavors.

Recent Posts