OpenAI Took First and Second on Artificial Analysis's Image Boards. Images 2.6 Is Priced by the Token, Not the Image, and Flare Is About Twice as Fast as the Last Flare.
OpenAI shipped Images 2.6 on November 4 across ChatGPT, Codex, and the API as GPT-Image-2.6 Flare and Sunburst. Artificial Analysis puts Sunburst and Flare at first and second on both its text-to-image and editing boards. The API bills by token, $5 per million text in, $8 per million image in, $30 per million image out, the same for both models, so AA's figure of about $0.21 per 1024-square max-quality image is a derived unit, not an OpenAI price. OpenAI's 50% faster, 4 billion images a week, and six-second Flare claims are vendor numbers; AA timed Flare at about 30 seconds median at max settings, roughly twice as fast as 2.5 Flare, and Sunburst at about two minutes fifteen. We read the launch post, the model docs, the system card, and AA's boards.
By T.W. Ghost
OpenAI shipped Images 2.6 today, November 4, 2026, in ChatGPT, Codex, and the API, as two models: GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst. The vendor pitch is speed and surface area: image generation 50% faster, Flare at about six seconds on low quality, animated GIFs, SVG export, image markup in the API, a complexity parameter, availability on every ChatGPT and Codex tier, and a usage claim of 4 billion images a week across the three surfaces. Two days later Artificial Analysis put both models at the top of its boards: Sunburst first and Flare second on AA-Image-T2I v2.1, and the same order on AA-Image-Editing v2.1.
The number to read before the Elo is on the pricing docs. Images 2.6 is billed by the token: $5 per million text tokens in and $1.25 cached, $8 per million image tokens in and $2 cached, $30 per million image tokens out. Flare and Sunburst share the same rates. There is no flat per-image price. AA's figure of about $0.21 per 1024-by-1024 image at max quality, or $210.72 per thousand on its board, is AA's arithmetic at those token rates, and it moves with resolution, quality, and every editing round. We read OpenAI's launch post, both model pages, the system card, AA's two boards, and AA's November 6 thread. Here is what held up.
What Shipped, and What It Costs
| GPT-Image-2.6 Flare and Sunburst (November 4, 2026) | |
|---|---|
| Where | ChatGPT, Codex, API; all ChatGPT and Codex tiers (vendor) |
| New | GIFs; SVG export; image markup in the API; complexity parameter (vendor) |
| Speed claims | 50% faster generation; Flare about 6 seconds at low quality (vendor) |
| Usage claim | 4 billion images a week across ChatGPT, Codex, API (vendor) |
| Token pricing, both models | Text in $5 per 1M, cached $1.25; image in $8 per 1M, cached $2; image out $30 per 1M |
| Flat per-image price | None |
| AA cost unit | About $0.21 per 1024-square max-quality image; $210.72 per 1,000 on the board (AA's derivation) |
| Provenance | SynthID and C2PA on output; system card says neither is unremovable |
| Preparedness | Neither model crosses Bio High or Cyber High; treated with High-bio-style mitigations (system card) |
Token pricing is the catch for anyone budgeting. A $30-per-million output rate means a bigger image, a higher quality setting, or a third edit pass each costs more, and the same two models with the same rates means picking Sunburst over Flare is a speed and quality trade, not a price one.
The six-second Flare and the thirty-second Flare are both true. Six seconds is OpenAI's low-quality figure. Thirty seconds is AA's median at the max-quality, high-complexity settings that earned the board rank. Pick the setting before you quote the time.
The Boards, and Who Measured Them
| Claim | Number | Who |
|---|---|---|
| AA-Image-T2I v2.1 | Sunburst (max) first, Elo 1,259 plus or minus 15; Flare (max) second, 1,228 plus or minus 12 | Independent (Artificial Analysis board) |
| AA-Image-Editing v2.1 | Sunburst (max) first, 1,209 plus or minus 14; Flare (max) second, 1,206 plus or minus 13 | Independent (AA board) |
| Flare speed vs 2.5 Flare | Median text-to-image about 55 s to about 30 s; editing about 58 s to about 32 s | Independent (AA, Nov 6 thread) |
| Sunburst speed | About 2 minutes 15 seconds per image at max quality, high complexity | Independent (AA) |
| 50% faster; 6-second Flare; 4 billion images a week | Vendor (OpenAI) | |
| Safety eval, 3,077 adversarial prompts | Flare unsafe-not-blocked 2.14% vs 2.5 Flare 1.49%, not statistically significant overall; Sunburst 1.72% vs 1.92% baseline | Vendor (system card) |
Three honest observations:
- First and second is an independent result, at max settings. The Elo is AA's, with confidence intervals, and it is the strongest line in the launch. It was earned at the slow end of the quality dial.
- Sunburst pays for the crown in time. Two and a quarter minutes an image is the price of the top row; Flare sits at 1,206 to Sunburst's 1,209 on editing and lands in about 30 seconds.
- GIFs are the new risk surface, by OpenAI's own card. The system card adds GIF-specific safeguards and says motion risks are still hard to catch. The safety numbers are OpenAI's.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Launched Nov 4 in ChatGPT, Codex, API; Flare and Sunburst model ids | Verified (OpenAI) |
| Token pricing; same rates for both models | Verified (OpenAI model docs, rechecked Nov 7) |
| First and second on T2I and Editing boards with the Elo figures above | Verified (AA boards) |
| About $0.21 per max-quality image | AA's derivation at OpenAI's token rates |
| Flare about 2x faster than 2.5 Flare; Sunburst about 2 min 15 s | Independent (AA timing) |
| 50% faster; 6-second Flare; 4 billion images a week | Vendor |
| SynthID and C2PA; GIF safeguards; Preparedness below High | Vendor (system card) |
Who Should Use It
Use Flare to iterate and Sunburst when the rank matters; both cost the same per token, so the only thing you are choosing is time against the last few Elo points.
Price production at token math, and if you need a per-image number, use AA's $0.21 for max quality and expect it to move with your settings.
The frontier is elsewhere, and for once it is here: on the independent image boards OpenAI holds both top rows this week, which is more than Canva's Craft or FLUX 3 Image Klein could say last month. Our model comparison has the current standings on the language models side by side.
Sources
- OpenAI: introducing Images 2.6
- OpenAI API docs: gpt-image-2.6-flare
- OpenAI API docs: gpt-image-2.6-sunburst
- OpenAI deployment safety: ChatGPT Images 2.6 system card
- Artificial Analysis: text-to-image leaderboard
- Artificial Analysis: image editing leaderboard
- Artificial Analysis on X: Images 2.6 timing and cost
*Not sure whether a token-priced image model fits a budget built on per-image fees? Take the free 2-minute quiz and get matched. Then read about the image model that debuted fourth behind three OpenAI rows: Canva Craft.*