OpenAI's third GPT Image generation of 2026, launched November 4 in ChatGPT, Codex and the API: Flare (fast) and Sunburst (detailed), with animated GIF generation, SVG export and a complexity control added. Token rates are identical to 2.5 — but per-image token use at low/medium quality is 4.2x/3.3x higher, so bills at those tiers are several times larger. OpenAI claims 50% faster generation; third-party tests show only Flare delivering it, while Sunburst 2.6 is slower than 2.5 Sunburst. AA has 2.6 Sunburst at #1 on both boards (T2I 1261, editing 1210); LMArena still lists 2.5 Sunburst first. This page collects official specs, pricing, API status and first readings, and marks what is still unverified.
The confirmed picture first: 2.6 is the third GPT Image generation of 2026, the dual-model split is not new (2.5 had it too), and the three genuine additions are GIF, SVG and complexity. "Same price" refers to token rates only — per-image bills at low/medium quality are several times higher — and the "50% faster" claim is only borne out for Flare.
GPT Image 2 to 2.6 took only 20 weeks, then 8 weeks between generations; the name leaked a week before launch in a public OpenAPI commit; six days after launch, GPT Image 2 was marked deprecated.
GPT Image 2 released (the first generation of 2026).
GPT Image 2.5 released, also as a Sunburst + Flare pair — the dual-model split is not new to 2.6.
gpt-image-1 retired.
An OpenAI public OpenAPI-repo commit adds the markup parameter, described as "Supports GPT Image 2.5 and 2.6" — the first official appearance of "2.6"; an Oct 31 commit reverts the description to "Supports GPT Image 2.5".
Official announcement "Introducing Images 2.6"; X announcement at 23:35 UTC; developer forum post at 23:49 UTC; API changelog and system card the same day. ChatGPT, Codex and the API go live simultaneously.
LMArena T2I and editing boards gain 2.6 entries (marked Preliminary); fal ships 2.6 endpoints the same night.
Recraft Studio and Runware add it; Recraft runs 12 prompts vs V4.1 Pro and 5 vs Nano Banana 2.1.
Artificial Analysis adds both 2.6 models to its boards.
GPT Image 2 marked deprecated, removal set for 2027-02-10, designated replacement GPT Image 2.6.
gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest retire (sources disagree on the designated replacement, so this page records only the retirement fact).
This topic page goes live.
Per the official API documentation (read directly Nov 13–14, 2026). Note: 2.6 supports only the Image API's two endpoints; the guide's size section names the 2.5 models, and 2.6's 4K support is confirmed by third-party tests.
POST /v1/images/generations and POST /v1/images/edits; Responses API image tools, Chat Completions and Batch API are all marked unsupportedlow / medium / high / xhigh / max / auto (xhigh/max were added in the 2.5 generation and carry over to 2.6)low/medium/high, default high; independent of quality (rendering fineness); no separate fee; the guide recommends low for most edits and medium/high for scenes with many elements or constraints; high complexity may add latency and introduce unrequested changesoutput_compression 0–100%); background supports transparent/opaque/auto, transparency requires PNG/WebP; size, quality and background all support autooutput_format: "gif" on both generation and editing endpoints; frame_count 2–16 frames (default 9); frame_duration_ms 20–2,000ms per frame (default 200ms, rounded to the nearest 10ms); the default combination is 9 × 200ms = a 1.8s loop; requires GIF preview access for the chosen model; no streaming, partial images, custom compression or maskoutput_format: "svg" on the generation endpoint, n must be 1; the editing endpoint does not support SVG output, and SVGs cannot be used as editing input; two-stage flow: the API generates a reference image first, then produces the SVG from it (svg.model specifies the author model), billed as two parts; returns static vector content (editable paths, shapes and colors; no scripts, external resources or embedded bitmaps) with text kept as live editable textmask; supported 2.5 models got markup too — it is not a 2.6 exclusivepartial_images 0–3, each partial image costs an extra 100 image-output tokens; SVG/GIF outputs don't support streaming, partial images or custom compressionmoderation set to auto (default) or low (more permissive)safety_identifier per request (a stable, de-identified user identifier, ≤128 characters); the older user parameter is deprecated but still supported; C2PA metadata and SynthID watermarkFastest is ChatGPT (free tier included); for API access go direct to OpenAI or via OpenRouter; Codex users should note image turns consume quota 3–5x faster than text-only turns; to try Sunburst free, Recraft Studio's free plan covers its default settings.
Live globally, all tiers, desktop, mobile and web. The free tier works — multiple reports and hands-on tests confirm free accounts can use it (the Medium tester used a free account) — but OpenAI publishes no fixed image count for any tier; the only number is a single-person test that hit the cap after ~8 operations (images, GIFs and edits combined), with a prompt to try again in 24 hours. New: saved reference images (save a photo as a named reference, then reuse it with @name in other chats, no re-upload); GIF generation in ChatGPT is still rolling out gradually, not all accounts have it yet. Note: OpenAI does not disclose whether ChatGPT calls Flare, Sunburst or switches by context.
Model IDs gpt-image-2.6-sunburst / gpt-image-2.6-flare, Image API generations/edits endpoints only. No free tier on the API. GIF requires GIF preview access for the chosen model. SVG is API-only.
The announcement says globally available on all tiers; image generation counts against the same usage quota as normal conversation, and per OpenAI's Codex pricing FAQ, turns with image generation consume quota 3–5x faster than comparable text-only turns (depending on image quality and size). The Codex CLI ships a built-in $imagegen skill, with generated images saved by default to ~/.codex/generated_images/ (that detail comes from a gpt-image-2-era article; Codex-specific behavior for 2.6 beyond the announcement and the FAQ multiplier has not been separately verified).
openai/gpt-image-2.6-sunburst and openai/gpt-image-2.6-flare, listed Nov 4, sole provider OpenAI (no routing); prices match the official ones ($8/$30 per 1M tokens); EU-region routing not supported yet.
Replicate (per-image: low $0.05, medium $0.09, high $0.128, max $0.50; GIF and SVG billed per output token), fal (live on launch night, $30/1M output tokens), Runware (both models on Nov 5), BudgetPixel ($0.12/image retail), SYNTX.AI, Treza, LaoZhang API (own VIP model at a fixed $0.03/call — aggregator pricing, not OpenAI's).
Added Nov 5: Sunburst available at default settings on the free plan, Flare and all quality tiers require a paid plan; 1K or 2K; credits 4–57 (default 10); model comparison puts two models side by side on the same canvas. This is the only confirmed route outside OpenAI's own channels for trying Sunburst free.
Token rates are identical to 2.5, but per-image token use at low/medium quality is 4.2x/3.3x higher — the official guide warns explicitly: "Equal token rates don't mean equal cost per image." High and max are the same price.
| Item | GPT Image 2.6 | GPT Image 2.5 |
|---|---|---|
| Text input / 1M tokens | $5.00 (cached $1.25) | $5.00 (cached $1.25) |
| Image input / 1M tokens | $8.00 (cached $2.00) | $8.00 (cached $2.00) |
| Image output / 1M tokens | $30.00 | $30.00 |
| Batch pricing | None (2.6 is not in the Batch table) | Half price (image output $15/1M) |
2.6 does not support the Responses API, so the guide's caching section — which covers GPT-Image-2 and 2.5 inside the Responses API only — does not apply to 2.6. The pricing page lists cached rates for 2.6, but whether they can actually be realized is unverified. Don't budget cache discounts.
| Quality | 2.6 tokens | 2.6 per image | 2.5 per image | Multiple |
|---|---|---|---|---|
| low | 823 | $0.02469 | $0.00588 | 4.2x |
| medium | 1,449 | $0.04347 | $0.01317 | 3.3x |
| high | 1,756 | $0.05268 | $0.05268 | Same |
| xhigh | 3,122 | ~$0.094 | ~$0.094 | Same |
| max | 7,024 | $0.21072 | $0.21072 | Same |
Other sizes measured: 1536×1024 (mxchat, Nov 7, 46 images): low 664 tokens $0.0203, medium 1,132 tokens $0.0343, high Flare 1,353 tokens $0.0410 / Sunburst 1,372 tokens $0.0415, xhigh $0.0741, max $0.1650. 2048×2048 (official calculator): low $0.050, max $0.428. 3840×2160 low $0.047 (LaoZhang). SeedDance measured 4K at ~3.4x the price of 1K, and a 3:1 panorama at about one-third of a square image.
QuantSlant same-prompt comparison (medium): 2.5 Sunburst 343 tokens ($0.011) vs 2.6 Sunburst 1,132 tokens ($0.034), about 3x the bill; a three-round editing flow cost $0.046 / ~27s on 2.6 vs $0.022 / ~19s on 2.5.
SVG is billed as two parts: the reference image at image-model rates + the author model at its token rates, with no fixed per-image price (per the guide); CellCog measured $0.03–$0.23 per returned SVG. GIF is billed per output token (Replicate's framing). Streaming partial images cost an extra 100 image-output tokens each.
Official pages: OpenAI announcement · OpenAI pricing · Image generation guide
In the ten days after launch, testing concentrated on three questions: is the 50% speed-up real, does SVG export actually work, and is text rendering better. The answers: Flare is genuinely faster, Sunburst is slower; SVG reliability is poor; text looks good in small samples.
12 identical prompts (medium): 2.5 Sunburst median 17.9s, 2.6 Sunburst 26.3s, 2.6 Flare 8.0s — "50% faster" holds only for Flare, and Sunburst 2.6 was slower than 2.5 Sunburst on all 12 prompts (the author notes launch-day server congestion may have affected the data). Quality ladder (3:2 bakery prompt): low 7.4/19.0s, medium 7.4/23.7s, high 10.1/39.2s, max 24.7/118.3s (Flare/Sunburst). Same prompt at medium: 2.5 Sunburst 343 tokens ($0.011) vs 2.6 Sunburst 1,132 tokens ($0.034).
Only 9 of 48 valid requests returned an SVG (19%), all failures bare HTTP 502s (after 11–91s); prompts with text succeeded 3 of 41, prompts without text 6 of 7; successful calls median 20.9s. The same batch on Recraft V4.1 was 9/9, median 8.3s, fixed $0.08/image. CellCog says it re-runs the same batch weekly (no updated result as of Nov 14).
Recraft compared 2.6 against its own V4.1 Pro on 12 prompts and against Nano Banana 2.1 on 5, concluding GPT 2.6 leads on text accuracy (Recraft's own account). The same day Recraft Studio added 2.6: Sunburst available at default settings on the free plan.
Flare travel ad at low: 9.7s vs 2.5 Flare's 15.0s (1.5x), living-room edit 16.6s vs 21.4s (1.3x); Sunburst 2.6 took 43–61s per image vs 2.5 Sunburst's 32–42s. The "Flare faster, Sunburst slower" finding matches QuantSlant.
Across five hands-on tests, the ChatGPT browser test produced a genuine 12-frame GIF, a 2.16s loop. Syntx's two-day hands-on reached a similar conclusion: clean loop seams and stable subject size across frames. OmniaKey requested 9 × 200ms on Nov 5 and got 9 × 180ms, 1.62s total — a slight frame-duration discrepancy.
The author tested on a free account: Korean text 4/4 correct (the three local models compared got only 1–2 right); ~13–16s per image; the cap hit after ~8 operations (images, GIFs and edits combined), with a prompt to retry after 24 hours (counted from the first operation). This is the only quantitative evidence available — a single-person test, not an official quota.
Six in-depth articles from this site (Chinese + English), one question each, rewritten from public sources with a reference link.
On AA's boards 2.6 Sunburst is already first; on LMArena's boards 2.5 Sunburst still leads — but 2.6 has only a few thousand votes and is marked Preliminary, so the two boards don't contradict each other; it's a sample-size difference. On Rapidata's overall board the two 2.6 models take the top two spots.
| Model | T2I Elo | Editing Elo | Notes |
|---|---|---|---|
| GPT Image 2.6 Sunburst (max) | 1261 ±14 (#1) | 1210 ±13 (#1) | 5,282 T2I votes |
| GPT Image 2.6 Flare (max) | 1226 ±12 (#3) | 1206 ±13 (#2) | 7,169 T2I votes |
| GPT Image 2.5 Sunburst (max) | 1229 ±10 (#2) | 1201 ±11 (#3) | 32 T2I points behind 2.6 |
| Nano Banana 2.1 | ~1176–1177 (#6) | ~1152 | — |
| Seedream 5.5 | 1135 (#10) | 1159 (#6) | See the Seedream 5.5 topic |
T2I: 2.5 Sunburst still #1 (1429 ±5, 63,818 votes); 2.6 Sunburst (Medium) #2 (1406 ±13, 2,697 votes); 2.5 Flare #3 (1398); 2.6 Flare (Medium) #4 (1384 ±13, 2,532 votes); GPT Image 2 (Medium) #5 (1382) — the top five are all OpenAI. Editing: 2.5 Sunburst #1 (1519 ±5, 131,648 votes); 2.6 Sunburst (Medium) #3 (1476 ±11, 3,493 votes); 2.6 Flare (Medium) #6 (1460 ±10, 3,534 votes). All 2.6 rows are marked Preliminary, with only a few thousand votes and wide confidence intervals (±10–13), so they can't be directly compared against 2.5's six-figure vote counts.
7,679,896 human responses, 57 models, 1,514 prompts: overall Sunburst #1 (1308), Flare #2 (1284), 2.5 Sunburst #3 (1212). By dimension: Sunburst alignment 1445 (#1), preference 1326 (#1), coherence 1132 (#6), text coherence 1254 (#2); Flare text coherence 1348 (#1).
The official announcement says "Images 2.6 generates images 50% faster than before" without naming the comparison baseline (not which model, quality or workload); OpenRouter's model description says "about 50% faster than GPT Image 2.5". Third-party tests agree: Flare is faster, Sunburst 2.6 is slower than 2.5 Sunburst in multiple tests. The correct phrasing is "officially 50% faster (baseline unstated); third-party tests show Flare ~1.5–2.6x faster, Sunburst 2.6 slower than 2.5 Sunburst in multiple tests" — never present 50% as a general speed-up for both models.
No controlled GenEval, DPG-Bench or DrawBench results exist (kingy.ai searched Nov 7–9 and found none; that is a limitation of the search, not proof they don't exist). AA scores move slightly each day (tech-insider read 1262/1230/1227 on Nov 13).
On AA's boards 2.6 Sunburst is first, but on LMArena 2.5 Sunburst still leads; against rivals, 2.6's edge is text accuracy and formats (GIF/SVG), its weaknesses SVG reliability and low-quality per-image price.
| Dimension | GPT Image 2.6 Sunburst | GPT Image 2.5 Sunburst | Seedream 5.5 |
|---|---|---|---|
| AA T2I | 1261 (#1) | 1229 (#2) | 1135 (#10) |
| AA editing | 1210 (#1) | 1201 (#3) | 1159 (#6) |
| LMArena T2I | 1406 (#2, Preliminary) | 1429 (#1) | Not listed |
| Exclusive features | GIF, SVG export, complexity | Responses/Batch workflows | 100-round editing, native RGBA, layer decomposition |
| Max resolution | 4K (confirmed by tests) | 4K | 2K |
| Billing | Per token (image output $30/1M) | Per token (same rates; Batch half price) | Per image (Volcano ¥0.13/¥0.23) |
| 1024×1024 low per image | $0.02469 | $0.00588 | ~$0.02 (third-party) |
| Weights | Not public | Not public | Not public |
Nano Banana 2.1 (Google Gemini): AA T2I 1176–1177 (#6), editing 1152; Google lists a 1K standard image at ~$0.0336, 1K/2K/4K output, search grounding, up to 14 reference images; in Recraft's five-prompt comparison GPT 2.6 led on text accuracy (Recraft's own account). FLUX 3 Image (Black Forest Labs): Rapidata 1109; ~$0.048/1K image; native 2K/4K and bounding-box control. Recraft V4.1: the SVG reliability gap is large (CellCog's test: 2.6 returned SVGs on 9/48 requests, Recraft 9/9); V4.1 vectors $0.08/image; Recraft generates natively vector, 2.6 converts bitmap to vector. Ideogram 4.5: API prompt generation $0.03/$0.06/$0.10 (low/medium/high), high reference and editing $0.22; native 2K.
Same name, different model: Microsoft's MAI-Image-2.6 has nothing to do with GPT Image 2.6 — LMArena T2I #6 (1330); its editing preview MAI-Image-2.7-Preview sits at 1195 on the AA editing board (#2–4 range). The Trends query "mai image 2.6 vs gpt image 2" comes from exactly this confusion. For detail, see this site's 2.6 vs 2.5 upgrade guide and the Seedream 5.5 vs GPT Image 2.6 comparison.
Ten days after launch, these are still unknown: some await an OpenAI statement, others need more hands-on testing.
$imagegen come from gpt-image-2-era articlesConfusion around 2.6 comes mostly from four places: the baseline of "50% faster", what "same price" actually means, the old docs' GIF frame count, and the same-named Microsoft model. One by one.
The official announcement says "Images 2.6 generates images 50% faster than before" with no named baseline; OpenRouter's model description says "about 50% faster than GPT Image 2.5". Third-party tests: Flare is ~1.5–2.6x faster (QuantSlant median 8.0s vs 2.5 Sunburst's 17.9s), but Sunburst 2.6 is slower than 2.5 Sunburst in multiple tests (QuantSlant median 26.3s vs 17.9s; OverChat 43–61s vs 32–42s). "50%" cannot be written as a general speed-up for both models.
Token rates are indeed the same (text $5, image input $8, image output $30 per 1M), but per-image token use at low/medium quality is 4.2x/3.3x that of 2.5: low 823 tokens $0.02469 (2.5: 196 tokens $0.00588), medium 1,449 tokens $0.04347 (2.5: $0.01317). High and above cost the same. The official guide also warns "equal token rates don't mean equal cost per image".
Launch-day docs and early coverage said 2–64 frames; after Nov 7 the docs read 2–16 frames (default 9); kingy.ai published a correction on Nov 9. The current figure is 2–16; 64 is an old docs number.
The guide is explicit about the two-stage flow: the API first generates a reference image (quality controls that image), then produces the SVG from it; the svg.model field names the author model that writes the vector code; billing adds the two stages. Recraft V4.1 is the native vector generator.
Both the official changelog and the guide say supported 2.5 models support markup too; markup was already available in ChatGPT, and this release brings it to the API. The 2.6 additions are GIF, SVG and complexity (changelog wording).
MAI-Image-2.6 is Microsoft's image model and has nothing to do with OpenAI's GPT Image 2.6. Same name, different model — the comparison searches in Trends come from exactly this confusion.
The system card is explicit: SVG export is API-only; there is no SVG export in ChatGPT. Set output_format: "svg" on the generation endpoint; the editing endpoint doesn't support SVG output.
The free tier working is true (multiple reports and hands-on tests confirm it), but "unlimited" is false: OpenAI publishes no fixed image count for any tier, and the only number is a Medium author's single-person test — the cap hit after ~8 operations, with a prompt to retry after 24 hours. A single-person test, not an official quota.
November 4, 2026: the official announcement, API changelog, system card, ChatGPT release notes and X announcement (23:35 UTC) all landed the same day; ChatGPT, Codex and the API went live simultaneously. It's the third GPT Image generation of 2026 (GPT Image 2 in April, 2.5 in September). See the section 02 timeline.
Token prices are identical and token use at the same settings measures the same — the choice trades time for detail, not money. Flare is "fast iteration": as fast as ~6s at low quality, improved layout quality, good for posters, product shots and other fast-turnaround work; Sunburst is "detailed visual work": higher quality at medium/high and stronger multi-round editing consistency. Measured speed gap: Flare is ~2–3x faster at the same settings; Sunburst at max quality can take around 2 minutes. See this site's Flare vs Sunburst article.
It's billed per token: image output $30 per 1M tokens. At 1024×1024: low $0.02469, medium $0.04347, high $0.05268, xhigh ~$0.094, max $0.21072. Note low/medium cost 4.2x/3.3x more than the same tiers on 2.5 (they use more tokens); high and above cost the same. No Batch pricing. See this site's pricing guide.
Three additions: animated GIF, SVG export (API-only), and the complexity planning parameter. Markup editing went to 2.5 at the same time, so it's not a 2.6 exclusive. Token rates are the same, but low/medium images cost 4.2x/3.3x more. Flare is faster, Sunburst is often slower. On AA's boards 2.6 Sunburst has passed 2.5 Sunburst; on LMArena's boards 2.5 Sunburst still leads (2.6 has far fewer votes and is Preliminary). Note: 2.6 doesn't support the Responses API or Batch API, so existing integrations that need those workflows can't migrate directly. See this site's upgrade guide.
Yes: ChatGPT launched globally on all tiers including free, and multiple reports and hands-on tests confirm free accounts can use it. But OpenAI publishes no fixed image count for any tier; the only number is a Medium author's single-person test — a free account hit the cap after ~8 operations (images, GIFs and edits combined), with a prompt to retry after 24 hours (counted from the first operation). That's a single-person test, not an official quota. The API has no free tier; Recraft Studio's free plan offers Sunburst at default settings.
In ChatGPT: a text prompt, a photo or a hand-drawn sketch can produce a looping GIF; playback speed is adjustable, frames are editable one by one; but you can't edit an existing GIF you upload; GIF generation is still rolling out gradually, and not all accounts have it yet. API: output_format: "gif" on both the generation and editing endpoints; frame_count 2–16 frames (default 9), frame_duration_ms 20–2,000ms per frame (default 200ms); the default combination is 9 × 200ms = a 1.8s loop; GIF preview access is required. See this site's GIF & SVG article.
API only (no SVG export in ChatGPT): set output_format: "svg" on the generation endpoint, n must be 1; the editing endpoint doesn't support SVG output. It's a two-stage flow: the API generates a reference image first, then produces the SVG from it, billed as two parts. The result is static vector content, with text kept as live editable text. Reliability is a problem, though: CellCog measured only 9 SVGs returned from 48 requests (19%), and just 3 of 41 on prompts containing text. See this site's GIF & SVG article.
gpt-image-2.6-sunburst and gpt-image-2.6-flare. Only the Image API's POST /v1/images/generations and POST /v1/images/edits; the Responses API, Chat Completions and Batch API are all unsupported. On OpenRouter they're openai/gpt-image-2.6-sunburst and openai/gpt-image-2.6-flare, listed Nov 4 at the official prices. See this site's API access article.
Hands-on tests say yes: mxchat generated 3840×2160 images, and LaoZhang reports native 4K tests passing; 2048×2048 costs low $0.050 / max $0.428, 3840×2160 low $0.047. But the guide's size section names the 2.5 models (single side ≤3840px, total pixels 655,360–8,294,400) and doesn't list a 2.6-specific ceiling. Above 2560×1440 is experimental.
On AA scores 2.6 Sunburst is higher (T2I 1261, editing 1210 vs 5.5's 1135/1159); on features 2.6 has native GIF, SVG export and markup editing, while 5.5 has 100-round editing, native RGBA and layer decomposition. MaverickFrame and BeatAPI agree: long feedback loops + transparent assets + dense Chinese text favor 5.5; GIF/SVG/precise markup favors 2.6. See this site's comparison article and the Seedream 5.5 topic.
On AA T2I, 2.6 Sunburst scores 1261 vs Nano Banana 2.1's 1176–1177; on editing 1210 vs 1152 — 2.6 leads on both. In Recraft's five-prompt comparison GPT 2.6 led on text accuracy (Recraft's own account). Nano Banana 2.1's advantages are price and reference images: a 1K standard image at ~$0.0336, up to 14 reference images, and search grounding.
Yes: the official product is "ChatGPT Images 2.6" / "Images 2.6", and the API models are "GPT-Image-2.6 Flare" / "GPT-Image-2.6 Sunburst"; the search variants "chatgpt image 2.6", "gpt image 2.6" and "gpt images 2.6" all refer to the same model. Don't confuse it with Microsoft's MAI-Image-2.6 — that's an unrelated model.
Every fact on this page comes from the public sources below (checked Nov 14, 2026).