OpenAI launches gpt-image-2 API with 1024×1024 pricing from $0.006 to $0.211
OpenAI exposed gpt-image-2 through the API with model-card details and concrete 1024×1024 pricing tiers, from $0.006 at low quality to $0.211 at high quality. The release gives engineers a new image endpoint with stronger instruction following and text rendering, but output-cost tradeoffs remain material at higher settings.
TL;DR
- OpenAI exposed
gpt-image-2in the API and Codex on April 21, with the official announcement in OpenAI's launch post and the developer rollout summarized by the launch-page evidence. - The headline numbers circulating on day one came from the HN discussion roundup, where commenters pegged 1024×1024 output at $0.006 for low quality, $0.053 for medium, and $0.211 for high.
- The official API pricing page frames that same model as token-priced, $8 per 1M image input tokens, $2 per 1M cached image input tokens, $30 per 1M image output tokens, plus $5 per 1M text input tokens, so the per-image figures are calculator-style estimates rather than the base tariff.
- OpenAI's core pitch, according to the launch-page evidence and the image generation guide, is stronger instruction following, better text rendering, and more usable structured outputs like diagrams, comics, and infographics.
You can jump from the consumer launch post to the developer model page, check the pricing page, and skim OpenAI's surprisingly practical prompting guide. The useful extra detail came from minimaxir's pricing comment, ea016's 1024×1024 comparison, and Simon Willison's CLI test, which turned the launch into a quick engineering smoke test instead of a pure marketing page.
API pricing
Discussion around ChatGPT Images 2.0
Thread discussion highlights: - minimaxir on API model card and pricing: Model card for the API endpoint gpt-image-2... API Pricing is mostly unchanged from gpt-image-1.5... The submitted page is annoyingly uninformative, but from the livestream it proports the same exact features as Gemini's Nano Bana… - ea016 on price comparison: GPT Image 2 ... Low: 1024×1024 $0.006 ... High: 1024×1024 $0.211 ... GPT Image 1 ... Low: 1024×1024 $0.011 ... High: 1024×1024 $0.167 - simonw on developer prompt testing: I've been trying out the new model like this: ... `uv run ... openai_image.py -m gpt-image-2` ... 'Do a where's Waldo style image but it's where is the raccoon holding a ham radio'
The cleanest day-one takeaway is that OpenAI launched a new image endpoint with two different pricing views. HN commenters quoted concrete 1024×1024 outputs by quality tier, while the official API pricing page lists token rates for image inputs, cached image inputs, image outputs, and text inputs.
For engineers budgeting jobs rather than reading token tables, the practical grid was:
- Low, 1024×1024: $0.006, per ea016's comparison
- Medium, 1024×1024: $0.053, per ea016's comparison
- High, 1024×1024: $0.211, per ea016's comparison
- Prior model reference from the same comment: GPT Image 1 at 1024×1024 was $0.011 low and $0.167 high, with medium not quoted there
That produces a slightly odd cost profile. Low got cheaper than the prior model in the HN comparison, while high got more expensive.
Structured outputs
Introducing ChatGPT Images 2.0
OpenAI has launched ChatGPT Images 2.0, an updated image generation model featuring advanced "thinking capabilities" that allow for web searching, improved instruction following, and the creation of complex assets like multi-paneled comic strips, infographics, and floor plans. Key technical enhancements include significantly improved text rendering—including non-Latin scripts such as Hindi, Japanese, and Korean—and the ability to maintain character and object continuity across multiple generated images. The model is available to all ChatGPT and Codex users, with advanced "thinking" features and additional capabilities provided for paid tiers. Developers also have access to the model via the gpt-image-2 API.
OpenAI's product framing was less about photorealism and more about utility. The launch post, developer announcement, and prompting guide all push the same cluster of upgrades:
- Stronger instruction following
- Better text rendering, including non-Latin scripts
- More reliable layouts for diagrams, charts, posters, comics, and infographics
- Support for more aspect ratios and resolutions, with the docs calling out
1024x1024,1536x1024, and1024x1536 - Better editing and continuity across outputs
The docs also make one implementation detail explicit: gpt-image-2 supports low, medium, and high output quality, and OpenAI positions low as the latency-sensitive setting while medium and high are for maximum fidelity.
Developer smoke tests
ChatGPT Images 2.0
For engineers, the key signal is that OpenAI exposed `gpt-image-2` through the API with pricing and model-card details, and commenters immediately started stress-testing it with structured prompts, benchmark-style grids, and code-driven workflows. The thread is most useful for understanding instruction-following, output-cost tradeoffs, and how the model compares operationally with other image systems.
The HN thread turned into fast adversarial testing. Simon Willison's example used a CLI workflow around gpt-image-2 for a "where's Waldo" style search image, and minimaxir's benchmark-style prompt pushed the model with an 8x8 Pokémon prime-number grid that mixes layout, factual recall, and heuristic following.
That is probably the most useful signal in the early reaction. Engineers immediately treated gpt-image-2 like a programmable component, not a toy canvas, and the first prompts were structured tasks that expose whether the model can keep constraints straight once the output stops looking like a generic pretty picture.