Source: OpenAI’s official launch video Introducing ChatGPT Images 2.0 (Sam Altman + research team Gabe / Kuan / Kenji / Alex), plus four creator coverage videos: GPT Image 2 just dropped — WOAH (LM Arena ranking), Nano Banana Finally Dethroned — GPT-Image 2.0 FULLY tested (head-to-head testing), ChatGPT Images 2 Tutorial for Beginners (UI walkthrough), 40+ INSANE Ways to Use ChatGPT Image for FREE (use-case catalog). Refreshed 2026-09-29 with the ChatGPT Images 2.5 release from raw/AI_Model_Month_is_Already_Delivering_Big_Gains.md (The AI Daily Brief, Nathaniel Whittemore relaying OpenAI’s announcement). Launch date corrected 2026-10-27 from OpenAI’s own post, ai-research/introducing-chatgpt-images-2-0-openai-2026-10-27.md (https://openai.com/index/introducing-chatgpt-images-2-0/). 2026-11-09: Images 2.6 section added from The Neuron Daily (raw/newsletter-theneurondaily-com-ef2052c870.md, raw/newsletter-theneurondaily-com-69aefd9ea9.md), Matt Wolfe (raw/AI_News_-_New_ChatGPT_Google_Models_14_Updates_Worth_Knowing.md) and Matthew Berman (raw/Kimi_K3.1_weights_OUT_open_source_computer_use_SOTA_image_model_science_model_Pika_H3_-_AI_NEWS.md, for PH3 Image). 2026-11-11: Images 2.6 checked against OpenAI’s post (ai-research/openai-introducing-images-2-6-2026-11-11.md, https://openai.com/index/introducing-images-2-6/), with API pricing and MiniMax-H3-Image from Matthew Berman (raw/Gemini_Carbon_NVIDIA_CEO_AFFAIR_Fable_5.2_Meta_Muse_Ember_GPT_Image_2.6_Minimax_H3_-_AI_NEWS.md) and MiniMax’s model card (ai-research/minimax-h3-image-huggingface-2026-11-11.md).

OpenAI shipped ChatGPT Images 2 (a.k.a. GPT Image 2, internal codename “duct-tape” per community sources) on April 21, 2026 — Sam Altman positions it as “GPT-3 to GPT-5 all at once” for image generation. The model is a thinking image generator (it researches and can search the web for accurate facts before generating), supports multilingual text, generates multiple distinct images at once (full magazines, manga comics with consistent characters, room-by-room renovation plans), runs at 2K resolution natively, and 4K via API. Available in ChatGPT and the API at launch. LM Arena text-to-image ranking: instant #1 with a 250+ ELO jump from previous leader Nano Banana 2 (Gemini 3.1 Flash Image Preview) — 1270 → 1512.

Key Takeaways

  • Launch date: April 21, 2026. OpenAI’s post “Introducing ChatGPT Images 2.0” is dated April 21, 2026, and OpenAI’s launch video was uploaded the same day (YouTube upload date 2026-04-21).
    • Correction (2026-10-27): this line previously read “~early May 2026”, inferred from the launch video saying “today” and the 40+ Ways video saying “last week” when the transcripts were fetched on 2026-05-08. The fetch date was taken for the release date. The April date also matches the wiki’s internal Codex image-generation notes (gpt-image-2 released 2026-04-21) and Jay’s remark on 2026-04-30 that it was “released just a few days ago” (Seedance fal skill).
  • Sam Altman’s framing: “Imagen 2.0 is a huge step forward. This is like going from GPT-3 to GPT-5 all at once.” (^[ambiguous] — the audio transcript reads “Imagen 2.0” but OpenAI’s product is consistently named “ChatGPT Images 2.0” / “GPT Image 2” elsewhere; treating “Imagen” as a whisper mistranscription of “Image” — the wiki has used “GPT Image 2” as canonical since 2026-05-05.)
  • Thinking-level image generation. Unlike GPT Image 1, this model “isn’t just generating, it’s thinking” — researches, searches the web, fills in gaps from world knowledge. Per OpenAI’s blog quoted in coverage: “It understands the world. So you get smarter images with less prompting.”
  • Multi-image generation in one shot. “First time in image generation” you can create multiple distinct images at once with shared structure / characters: full magazines with structured typography + photorealistic photos, manga comics with recurring characters and evolving storylines, full renovation plans for every room in a house.
  • 2K native, 4K via API. Multiple aspect ratios. “Extraordinary micro detail” — every grain of rice in a zoom-and-render demo looked individually rendered.
  • LM Arena #1, big margin. Text-to-image leaderboard — 1512 ELO vs Nano Banana 2 (Gemini 3.1 Flash Image Preview) at 1270. 250+ ELO jump. Prior leaderboard was Nano Banana 2 → GPT Image 2 leapfrog.
  • Multilingual text + accurate text rendering. Generates infographics, manga lettering, math proofs with rendered equations, and signage in non-English scripts — the kind of output where prior image models would silently corrupt text.
  • Infographics + math + structured layouts. Demoed: explanatory infographics for complex systems; image generation that includes a mathematical proof rendered correctly; YouTube thumbnail concept boards with consistent typography.
  • Available in ChatGPT + API + Higgsfield. Sam: “available right now in ChatGPT and in the API.” Higgsfield uses GPT Image 2 as a model option (covered in their gpt_image_2 mode); 4K API access works through Higgsfield too.
  • Big tip for realism: add the word photorealism. Per the Nano-Banana-tester: words like “realistic photo / iPhone photo / cinematic” produced mediocre realism; adding photorealism to the prompt was a step-change improvement. “Same prompt, just add the word — completely changes the result.”
  • Image editing is granular and accurate. The Beginners Tutorial highlights the select feature in the image canvas: hover over a region, click, then describe the edit (e.g., “replace with tail”). Much higher precision than describing what to remove in plain text. “Saves time and compute.”
  • Templates note. ChatGPT’s image template feature uses the template’s style, not the template image itself, as a driver. Common beginner gotcha — the template image is a style reference, not a starting frame.
  • Strong character consistency across scenes. Edit-and-evolve a single subject across many scenes (volcano boarding → surfing → skydiving → walking through haunted house) maintains face + identity. Better than prior best (Nano Banana 2 still loses face-fidelity in 4K renders, per the head-to-head).
  • Works for production marketing assets. Use cases catalogued in 40+ Ways: YouTube thumbnail concept boards (2×3 grids of distinct concepts), Instagram carousels, slide decks, infographic explainers, comparison images. The model “is just actually way more useful” than Image 1 for everyday marketing/content workflows.

What’s new vs GPT Image 1

CapabilityImage 1Image 2
Multi-image batchSingle image per generationMultiple distinct images in one shot (magazines, manga, room-by-room)
ResolutionStandard2K native, 4K via API
ReasoningDirect generationThinking model — researches, searches web, fills gaps from world knowledge
Multilingual textLimitedNative multilingual including non-Latin scripts
Text renderingOften corruptedAccurate text in infographics, equations, signage
Editing precisionWhole-prompt rewritesClick-region select + targeted prompt; whole-image rewrites still available
Aspect ratiosLimitedMultiple aspect ratios; still some difficulty with 16:9 (per creator testing)

Capability deep-dive

Thinking-level intelligence

Per OpenAI’s blog (quoted in WOAH coverage): “ChatGPT Image 2.0 is a step change in detailed instruction following — placing related objects accurately, rendering dense text, with the ability to generate across aspect ratios. It’s also accurate across languages and uses its expanded visual and world knowledge to fill in the gaps for you. So you get smarter images with less prompting.”

The “thinking” framing is consistent with frontier-model trends — image generation has joined LLM reasoning as a domain where a thinking step before generation matters. Practical impact: prompts can be terser; the model fills in correctness (e.g., a math proof’s equations, a manga panel’s recurring character costume) instead of demanding the prompter spell every detail.

Multi-image generation

OpenAI’s launch demo: “entire magazines with structured typography and photorealistic photos, full renovation plans for every room in your house, manga comics with recurring characters and evolving storylines.” First image generator with first-class multi-image batch where the images share state — character consistency across panels, typography consistency across magazine pages, room-by-room continuity.

Practical use: client storyboards, multi-asset campaign launches, comic strips, paginated reports. Previously these required either hand-stitching with character-LoRA fine-tunes (Stable Diffusion ecosystem) or Higgsfield’s Soul Character training pipeline (see Higgsfield Soul ID). Now in-the-box on ChatGPT.

Photorealism prompt trick

The single highest-leverage tip from creator coverage: add the word photorealism to any prompt asking for realistic output. Words like “realistic photo,” “iPhone photo,” “cinematic” did less than expected; photorealism was the consistent unlocking keyword. Worth standardizing in any GPT-Image-2 prompt template aiming for non-illustrated output.

Granular editing via select

Beginners Tutorial demo: instead of “remove the [thing]” prompts (often miss because the model doesn’t know which “thing”), click the select button in the canvas, hover over the region, then describe the edit (“replace with tail”). The select-feature scopes the edit to the visual region the user chose, not whatever the model parses from text. Reduces token spend and avoids whole-image regeneration.

LM Arena context

The 250+ ELO leap from 1270 (Nano Banana 2 / Gemini 3.1 Flash Image Preview) → 1512 (GPT Image 2) is the largest single-model jump on the text-to-image arena since the leaderboard launched. Comparable to GPT-4 → GPT-4o on the LMs side, but compressed to a single release. Implies GPT Image 2 will hold #1 against Nano Banana 2 follow-ons until Google ships Nano Banana 3 or Gemini 4 image-mode.

How to access it

SurfaceHow
ChatGPT (web/app)Just generate — default image engine for paid tiers from launch day. Templates feature available; use select for granular edits.
ChatGPT APIgpt-image-2 model. 4K supported via API. See OpenAI Images API docs.
HiggsfieldGPT Image 2 mode + 4K rendering. See [[ai-video-content/higgsfield-overview
Codex CLIVia the image_generation tool over ChatGPT OAuth (internal codex-OAuth notes)
RaycastYouMind-OpenLab’s Raycast integration with dynamic argument substitution — see [[chatgpt-image/awesome-gpt-image-2

Use cases (from 40+ Ways)

The 40+ Ways catalogue — distilled into the categories the creator demonstrated:

  • YouTube thumbnail concept boards — “create six 16:9 thumbnail concepts in a 2×3 grid” prompts produce useful divergent options, though the model still misses 16:9 aspect ratios cleanly.
  • A/B thumbnail test sheets — three concepts with different hooks for split-testing.
  • Instagram / LinkedIn carousels — multi-image generation with shared style produces native carousel content.
  • Slide decks — produce structured slide images with consistent typography for slide-deck workflows.
  • Infographic explainers — explain a complex system in an image with accurate text.
  • Comparison images — head-to-head visual comparisons (model A vs model B, before vs after).
  • Magazine spreads — full magazines with covers + photorealistic photos + structured typography.
  • Manga comics — recurring characters across panels.
  • Room renovation plans — every room in a house with consistent visual style.
  • Math + science visualizations — equations + diagrams rendered accurately as images.

Try It

  1. Confirm access. Open ChatGPT (paid tier). Start a new image — confirm you’re on Image 2 (interface labels image generation; some accounts may roll out gradually).
  2. Add photorealism to your realistic-image prompt as the only change. Compare against same prompt without it. Replicate the creator-validated unlock.
  3. Generate multi-image content. Try “create a 4-panel manga comic with recurring character X” or “create a full magazine cover + 3 inside spreads in style Y.” Multi-image batch is the biggest workflow shift; test it on your actual content needs.
  4. Use the select editing feature. Generate any image, click select, hover over a region, edit. The precision shift vs whole-prompt rewrites is real.
  5. Test 4K via Higgsfield or API. For client deliverables where ChatGPT’s 2K is short, run the same prompt through Higgsfield’s gpt_image_2 mode at 4K and measure the quality lift. The Nano Banana tester reports the gap is significant for face fidelity.
  6. Read awesome-gpt-image-2 (4,430 prompts in 16 languages) and Depikt for prompt libraries to short-circuit the prompting curve.

Open Questions

  • Pricing details. ChatGPT pricing tier behavior + API price-per-image not extracted from the launch video. Confirm against OpenAI pricing.
  • Rate limits / safety filters. No mention of new safety filters or rate limits in the launch video; coverage videos didn’t probe the safety surface. Check ChatGPT Image policies.
  • EU rollout. OpenAI features have varied EU release dates. Whether GPT Image 2 is in EU at launch unconfirmed.
  • Template-image-as-driver mode. Tutorial creator notes the templates feature uses the template’s style, not its image as a driver. Whether OpenAI plans to add image-as-driver mode (Img2Img-style anchor) is unstated.
  • Long-form text generation. Renders accurate text in graphics, but longest demo’d block is a paragraph. Whether full-page text (book covers, brochures) renders correctly at scale not tested.
  • Sora / video model overlap. Whether GPT Image 2 outputs feed into Sora directly — likely yes given the OpenAI consolidation pattern, but transcripts don’t probe.

Field-test refresh — 50-example comparison vs Nano Banana Pro (2026-05-28)

[YouTube signal — wiki-inbox 2026-05-28] Source: raw/OpenAI_JUST_Dropped_GPT_Image_2_and_It_s_WILD_Nano_Banana_Pro_Killer.md (YouTube Tj3jFnx5IW0, creator had early access). 50-example head-to-head field test from a creator with pre-launch early access. Key reconfirmations + new observations beyond the original launch coverage:

  • Character consistency = solved vs Nano Banana Pro on the same reference. Six-character composition retained tattoos, piercings, hairstyle across all six references; Nano Banana Pro dropped the tattoos and changed hairstyle drastically. GPT Image 2 also held a Mona-Lisa-shaped reference better.
  • Multi-reference composition — character + location + 2 objects (one with a brand label) → editorial photo. Both models worked; GPT Image 2 trended more editorial, Nano Banana Pro more cinematic.
  • The “duct-tape” codename confirmed — the model surfaced on AI image arena under that hidden codename in AB tests against itself + competitors before public rollout.
  • Pairs well with Seedance 2 for image-to-video chains. Creator: “GPT Image 2 gets even stronger when combined with C Dance 2” — pattern alignment with HeyGen + Seedance 2.0 cinematic prompting (cross-model chain producing better results than single-model).
  • Real-time internet browsing confirmed — model can search the web for accurate facts before generation; produces art about real-world events with current detail.
  • Self-correcting thinking flagged — model catches its own errors before the operator does. Same direction as the launch coverage but the field-test confirms it survives 50 stress-test cases.
  • Multi-image consistency at scale — whole magazines, 10 different camera angles of a scene, 7-page manga with one prompt → multi-page output all confirmed working under stress.

Operator implication: the launch coverage above remains correct on capability claims; the field test removes the “marketing demo” caveat from the character-consistency and multi-image claims. The Seedance 2 pairing tip is the new operator move worth lifting into a downstream workflow article.

Images 2.5 (September 2026)

OpenAI released ChatGPT Images 2.5, the next model behind built-in image generation in ChatGPT, on a Tuesday in early September 2026 (the same Tuesday Nathaniel Whittemore dates to September 8).^[inferred] Source: The AI Daily Brief, “AI Model Month is Already Delivering Big Gains” (raw/AI_Model_Month_is_Already_Delivering_Big_Gains.md), with NLW relaying OpenAI’s announcement. These are vendor claims, not tests.

  • Usage. OpenAI says ChatGPT users generate “more than 3 billion images a week.”
  • Claimed gains: sharper detail, more precise editing, and faster generation with “a 50% reduction in latency.”
  • Two variants. Flare, “the fast version designed for quick iteration,” and Sunburst, “optimized for professional workflows that require better control across edits.”
  • Sketch. A new ChatGPT feature that takes a drawn input to guide generation; you can add a text prompt for style or detail.
  • The theme is control. NLW compares it to Nano Banana’s finer editing control. He quotes Higgsfield’s head of product: “What impressed us most about GPT image 2.5 Flare is how well it understands what not to change. You can make a meaningful edit without losing the character, composition, or visual identity of the original image.” Higgsfield offers OpenAI’s image models inside its own product (see the access table above), so read that as a partner endorsement.
  • New genre. NLW points to stop-motion animation as something the edit consistency makes viable for the first time.
  • Codex angle (NLW’s opinion). He still prefers the look of Fable-built websites, but calling GPT Image from Codex for UI assets and aesthetics makes him use Codex’s integrated image generation more often.

How to use it. For iterative edits of one approved image (a product shot, a character across a campaign), try Sunburst and tell it what must not change; use Flare when you are exploring many directions quickly.^[inferred] A hands-on test of the release is in ChatGPT Images 2.5 field test.

Second hands-on test (added 2026-10-30). AI Master ran GPT Image 2 against 2.5 Sunburst (through the API) on reference drift, a local edit, a six-turn edit chain and an overloaded ad brief, then used 2.5 on a poster, an ad set, a dashboard mockup and a flat vector illustration (raw/GPT_Image_2.5_Tutorial_-_How_to_Use_ChatGPT_Image_Generator_Like_a_PRO.md). He found references hold better, edits stay more local and long prompts are easier to control, but on his one timed prompt “that speed advantage didn’t show up.” Method and prompt habits: GPT Image 2.5 — Comparison Tests and Prompt Habits.

Images 2.6 (November 2026)

Added 2026-11-09. Sources: The Neuron Daily, November 5 (raw/newsletter-theneurondaily-com-ef2052c870.md) and November 8 (raw/newsletter-theneurondaily-com-69aefd9ea9.md); Matt Wolfe, “AI News - New ChatGPT, Google Models, 14 Updates Worth Knowing” (raw/AI_News_-_New_ChatGPT_Google_Models_14_Updates_Worth_Knowing.md, https://www.youtube.com/watch?v=IrAHINf0M6E). OpenAI’s announcement was not fetched; speed and usage figures are OpenAI’s, as relayed.

OpenAI released Images 2.6 in the first week of November, in ChatGPT on every subscription tier, in Codex, and in the API. The Neuron calls it mostly a fix for the “change one thing, change everything” problem.

  • Usage. About 4 billion images a week across OpenAI’s products, up from “more than 3 billion” at the 2.5 launch.
  • Speed. About 50% faster, per OpenAI. The API’s speed-focused model can make an image in as little as six seconds at low quality (early coverage, via The Neuron). Wolfe timed a standard image (a San Diego skyline at sunset) at 20.4 seconds in ChatGPT. This is a new claim for 2.6, separate from the 50% latency cut claimed for 2.5.
  • Animated GIFs. Short loops from characters, mascots, product shots, sketches or storyboards. Wolfe’s “monkey roller skating” GIF arrived in about 20 seconds, and two follow-up edits (add a baseball hat, move him to the street) were applied in about 30 seconds. ChatGPT placed the first version in San Diego, which he puts down to its memory of where he lives.
  • Saved references. Store a picture of yourself, a pet, a character or a product once and reuse it across requests, aimed at consistent characters, products and brand details across edits.
  • For developers: image markup (mark the exact spot to change instead of describing it), SVG export (shapes that stay sharp at any size and stay editable), and a complexity setting for how much the model plans before drawing; OpenAI recommends low for quick edits and medium or high for crowded scenes with many constraints.
  • Two API models, same names as 2.5: GPT-Image-2.6 Flare for speed and layouts such as posters; GPT-Image-2.6 Sunburst for detail and consistent edits.
  • Unknown: how well SVG export handles photo-like images. The Neuron suggests trying a logo first.

Checked against OpenAI’s post (added 2026-11-11). The announcement is dated November 4, 2026 and confirms: 50% faster generation; Flare “in as little as six seconds” at low quality; Sunburst for higher quality and consistency across edits at medium or high quality; GIFs; image markup and SVG export in the API; a complexity parameter (low for most edits, medium or high for scenes with many constraints); 4 billion images a week; availability in ChatGPT and Codex on all tiers and both models in the API. OpenAI’s post does not mention saved references; that feature comes from the newsletter and creator coverage.

  • API price (Berman): 8 per million image-input tokens and $30 per million image-output tokens for both Flare and Sunburst.
  • The official MiniMax-H3-Image weights followed Pruna’s PH3 Image: 33.1B-parameter DiT, about 16-megapixel output in six steps, generation and editing in one checkpoint, 68.65 GB at BF16, under the MiniMax-H3-Image Community License (model card). Berman calls it a phenomenal open image model and gives about 18 GB for a quantized build, a figure not on the card.

How to test it (The Neuron’s test): put one character in five different scenes with a few edits in between. If scene five still looks like scene one, saved references work; if not, “50% faster doesn’t help much when you’re regenerating the same image three times.” For brand work, save the logo, product shot and mascot as references before the first campaign request, and use Sunburst for edit chains.

An open-weight alternative the same week. Pruna AI’s PH3 Image turns MiniMax’s H3 video model into an image model: eight steps instead of 50, about five times faster (a 4-megapixel image in 2.6 s instead of 12.4 s), seven times smaller than H3 (55 GB at BF16, 30 GB at INT8) and strongest at photorealism and short quoted text (Berman, raw/Kimi_K3.1_weights_OUT_open_source_computer_use_SOTA_image_model_science_model_Pika_H3_-_AI_NEWS.md). Wolfe tried its free Hugging Face Space and found the quality “nowhere near” ChatGPT’s; its advantage is that you can fine-tune and run it yourself, at about 21 GB of VRAM with the 4-bit build for an 8-megapixel image (56 GB at 16-bit), per Wolfe.