Introducing ChatGPT Images 2.0

TL;DR

OpenAI’s official April 21, 2026 release of ChatGPT Images 2.0 (API model id gpt-image-2, snapshot gpt-image-2-2026-04-21). The official blog page itself is a thin capability showcase — examples, no architecture, no benchmarks, no metrics — so most of what is concrete about the release comes from third-party coverage and the API model card. Two headline claims define the launch: (a) the first OpenAI image model with built-in reasoning (“Thinking mode”) that can plan layout, search the web, and self-check outputs before generating, and (b) near-perfect multilingual text rendering across Japanese, Korean, Chinese, Hindi, Bengali, and Arabic. Available to all ChatGPT and Codex users immediately, with the API rolling out in early May. The identity of this release was previously surfaced on the wiki as [[tweet-levelsio-2026-05-24]] — the LMArena leak under codenames maskingtape-alpha / gaffertape-alpha / packingtape-alpha.

Key claims

ChatGPT Images 2.0 is OpenAI’s first image generation model with native reasoning (“thinking”) capabilities built into the architecture, with a two-mode access split — Instant (default, free-tier eligible) and Thinking (Plus / Pro / Business / Enterprise only) [official blog; corroborated coverage].
Thinking mode enables three behaviors that Instant does not: web search for real-time fact verification, layout-aware planning before generation, and output self-verification [official blog; MacRumors §1].
Multilingual text rendering is the most-touted concrete capability — character-level accuracy on non-Latin scripts (Japanese, Korean, Chinese, Hindi, Bengali) is called out explicitly [official blog; multiple secondary reports].
Up to 2K output resolution and up to 8 coherent images from a single prompt with character/object continuity across the batch [official blog].
API pricing per million tokens: $8 input /$ 2 cached input / $30 output (image);$ 5 / $1.25 /$ 10 (text). Per-image estimates land around $0.006 (low quality) to$ 0.211 (high quality) at 1024×1024 [OpenAI pricing page].
Knowledge cutoff is December 2025 [official blog].
DALL-E 2 and DALL-E 3 are being deprecated and retired on May 12, 2026; gpt-image-2 replaces them as the default image model across ChatGPT and the OpenAI API [official blog].
No architecture, no training data, no benchmark numbers, and no per-task accuracy figures are disclosed in the official announcement [official blog body — capability showcase only].

Method

Not disclosed. OpenAI describes gpt-image-2 only as a “generalist model” or “GPT for images”; researchers have declined to confirm whether the underlying generator is diffusion or autoregressive. The announcement page is structured as a dense capability showcase — categorized example galleries across photorealism, retro collage, Bauhaus, fashion editorial, B/W documentary, children’s-book, manga, infographics, product grids, and character reference sheets — rather than a technical report. The “Thinking mode” behavior is described functionally: the model can plan layout, invoke web search, generate up to 8 coherent images in a single batch, and self-check outputs. No mechanism is given for any of these. The API surface (v1/images/generations, v1/images/edits, v1/responses, v1/chat/completions) is consistent with gpt-image-1.5; the snapshot id gpt-image-2-2026-04-21 is pinnable for production code.

Results

No quantitative results are provided in the official announcement. Numbers below come from third-party coverage.
Third-party reports place the model at #1 across every category on Image Arena within 12 hours of launch, by a +242-point margin — described in the coverage as the largest lead ever recorded on that leaderboard [secondary; buildfastwithai.com].
Third-party comparison vs Google Nano Banana 2 (Gemini 3.1 Flash Image) reports ~99% character accuracy on text rendering for gpt-image-2 versus continued failures by Nano Banana 2 on dense text in multi-element compositions [secondary; vidmuse.ai].
Tier-1 API users are capped at 5 images/min; Tier-5 (requires $1,000 spent + 30-day-old account) at 250 images/min [OpenAI rate limits page].

Why it’s interesting

This is the official artifact behind the [[tweet-levelsio-2026-05-24]] LMArena leak from a week earlier — the three *tape-alpha codenames now have a verified destination. It is the closed-source frontier baseline that several open-image-gen releases in this batch explicitly position against: [[blog-bfl-flux-2-klein-towards-interactive-2026-01]], [[blog-qwen-qwen-image-2-0-2026-02]], [[2509.23951-hunyuanimage-3-0-technical-report]], and [[blog-microsoftai-introducing-mai-image-2-2026-04]] are all targeting this competitive cohort, alongside [[tweet-officiallogank-2026-02-26]] (Nano Banana 2 / Gemini 3.1 Flash Image). The “Thinking mode” framing is the more research-interesting axis — it is the first productized example of [[thinking-with-modalities]] substrate (c) (external-tool / verification loop) embedded directly in a frontier image generator, complementing the academic substrates filed there (Video-Thinker’s tagged autoregressive CoT, DiffThinker’s diffusion-native parallel exploration, Gemini 3 Flash Agentic Vision’s Python-tool loop). What is not interesting: the official page itself ships zero benchmark numbers, zero architecture detail, and zero ablations — typical for an OpenAI product launch but a sharp contrast to the open-source releases this filing sits next to.