GPT Image 2 API

In internal benchmarks published alongside the model, fal-ai/gpt-image-2 clears roughly 99 percent on short Latin phrases and stays in the mid-90s on longer copy and CJK glyphs. In practice that means brand names, UI labels, menu items, and comic captions render spelled and kerned on the first pass. You still want to keep a quoted phrase under about 12 words per region for the highest reliability, and wrap the exact text in single quotes inside your prompt so the model treats it as a literal rather than a hint.

Quality on fal-ai/gpt-image-2 toggles the sampling budget behind the scenes. Low is good enough for thumbnails, moodboards, and rapid ideation and lands in a few seconds. Medium is the default for most production work and trades roughly 50 percent more time for noticeably cleaner detail, especially on faces, hands, and small type. High is reserved for hero stills and print-ready output; it takes the longest and costs the most, but it is the tier you want for anything that is going on a billboard, a cover, or a pitch deck shown at resolution.

Pricing on fal-ai/gpt-image-2 is structured by output resolution and one of three quality tiers. A 1024x1024 render costs $0.01 at low, $0.04 at medium, and $0.13 at high. Step up to 1024x1536 and the row becomes $0.02 / $0.08 / $0.29. 1920x1080 lands at $0.03 / $0.10 / $0.37. 2560x1440 is $0.05 / $0.17 / $0.66. 3840x2160 tops the table at $0.10 / $0.38 / $1.48. The tier gap is ~4x between low and medium and ~3x between medium and high, so budget by intent: draft at low, review at medium, deliver at high. fal.ai/pricing is the source of truth.

Yes, and CJK support is one of the reasons to pick fal-ai/gpt-image-2 over a generic diffusion model. Stroke order holds up on common characters, radicals connect properly, and mixed-script lines with Latin brand names next to Chinese body copy render without obvious glitches. If the brief calls for traditional versus simplified Chinese or for specific Japanese kana styles, be explicit in the prompt; the model respects the distinction. For very dense blocks of text, keep each character larger than about 24 pixels in the final canvas to avoid compression artifacts.

Sign in at fal.ai, open the API keys page, and create a new key. Store it as FAL_KEY in your environment, then call fal.config({ credentials: process.env.FAL_KEY }) once at startup. From there, fal.subscribe('fal-ai/gpt-image-2', { input: { ... } }) is all you need. Keys are scoped to your account and billed per request, so rotate them when you swap collaborators off the project. Never commit a FAL_KEY to git; your starter's gitignore already blocks .env files, and the ship check in this repo greps for leaked keys before every deploy.

Edit mode on fal-ai/gpt-image-2 accepts a reference image plus an optional mask and returns a variant that honors your edit instruction. It is ideal for changing a label on a product shot, swapping copy inside a UI mockup, or fixing a misspelled caption on a previously rendered image. Edit requests cost slightly more than a fresh generation because the model conditions on the input. Keep the edit instruction specific and scoped to the masked region; broad prompts on a narrow mask tend to produce inconsistent results around the edit boundary.

Most DALL-E 3 prompts port cleanly to fal-ai/gpt-image-2, but two things tend to shift. First, color comes out more neutral by default, so prompts that relied on DALL-E 3's warm, saturated bias will look cooler; add explicit warmth cues if you need the old palette. Second, parameters renamed: the size string now accepts 2048 variants, quality replaces the older style switch, and the content policy surface is stricter on celebrity and branded content. Expect about 80 percent of your prompt library to work untouched and the rest to need a one-line revision.

Yes, fal-ai/gpt-image-2 is stable enough for production workloads, but wrap the call behind your own abstraction so you can swap in fal-ai/flux/dev or another model if the endpoint schedule changes. Log every request ID; when a render looks off, fal's queue keeps a trace you can replay. For latency-sensitive flows run quality=low on the hot path (1024x1024 at $0.01) and reserve high only for user-triggered downloads where the cost tolerance is real.

Use fal-ai/gpt-image-2 when the image must contain exact text, a real product label, or CJK copy. Reach for Flux 2 Pro when the brief is photoreal and textless and you want the richest depth-of-field response on fal.ai. Reach for Imagen 4 when you want clean Google-style output, strong faces, and neutral fabric texture without paying the GPT Image 2 premium. If the brief is a typographic poster with no photo content, Ideogram 3 can undercut all three. Decide by text load first, then by price ceiling.

Eight reasons stack up when you host GPT Image 2 on fal.ai. One, the same FAL_KEY calls fal-ai/gpt-image-2 and 600 other models, so you keep one credential surface. Two, requests land on a serverless queue with no cold starts. Three, webhooks let you fan out batches and pick the results up asynchronously. Four, billing is consolidated into a single invoice. Five, fal's CDN serves result URLs fast worldwide. Six, the playground on this site lets your team test prompts without writing code. Seven, fallbacks such as fal-ai/flux/dev are one line away when the preview endpoint hiccups. Eight, the async queue has retry and dead-letter semantics so you do not lose a failed job.

OpenAI has not posted a date. What we know: three codename models (packingtape-alpha, maskingtape-alpha, gaffertape-alpha) briefly appeared on LM Arena on April 4, 2026, then were pulled. ChatGPT is running an internal A/B of something that behaves like GPT Image 2 on a small cohort. DALL-E 2 and 3 retire on May 12, 2026, which creates a hard pressure to ship a replacement before that date. Community bookmakers put the launch window at late April to mid May 2026 with roughly 40 percent probability. Set a Slack reminder for May 12, monitor the fal model page, and ship the fal-ai/gpt-image-2 endpoint switch behind a feature flag so you can enable it the moment it goes live.

Blind A/B tests circulated on LM Arena gave GPT Image 2 about a 75 percent win rate on text-heavy prompts, 65 percent on photoreal scenes, 60 percent on world knowledge (real brand layouts, landmark details), 55 percent on speed, and a tie near 50 percent on pure creative composition. Nano Banana Pro wins on stylised illustration and on aspect ratios the OpenAI model cannot natively produce, and it beats GPT Image 2 on prompts that need a single cohesive painterly palette. Most production teams shipping UI mockups or packaging will still want GPT Image 2 as primary and keep Nano Banana Pro wired in for creative overflow.

Yes. fal.ai passes through OpenAI's commercial terms for image output, which let you use, distribute, and sell generated images as long as the prompt does not ask for trademarked IP, the likeness of a living person without consent, or content that violates the content policy. Keep the request ID from every production render so you have a paper trail if a piece gets challenged. For highly regulated use cases (pharma packaging, political creative, financial disclaimers) maintain a human review step before anything ships.

On fal.ai you can request PNG or JPEG output. Aspect presets include 1024x1024 (square), 1024x1536 or 1536x1024 (3:2 or 2:3), 1024x1792 or 1792x1024 (9:16 or 16:9), and auto (the model chooses). High-tier also exposes a 2048 and native 4K mode (3840x2160) for hero stills. Transparent backgrounds are available via the background=transparent flag when output_format is png; JPEG transparent is a no-op and will quietly drop the alpha channel.

fal.ai gives new accounts a small starter credit that is enough to render a few hundred low-tier images on fal-ai/gpt-image-1.5 today (the same endpoint you will migrate off when 2 ships). Beyond that every request is paid per render with no minimum. There is no hard monthly subscription: you pay for what you generate. For prototyping, set a daily spend cap in the fal dashboard so a runaway loop cannot empty the account.

Yes, noticeably better than 1.5. Arena blind tests have over 70 percent of participants misclassifying GPT Image 2 portraits as real photographs versus somewhere near 55 percent for the previous generation. Expect cleaner skin micro detail, consistent catchlights, and fewer hand and finger anomalies. Edit mode respects facial identity when input_fidelity is set to high, which is the tool you want for corporate headshots that need a wardrobe or background swap. The content policy still refuses likenesses of real public figures without a consent reference.

Seed-style reproducibility is limited on GPT Image 2; the model does not expose a public seed parameter today. The reliable path is a reference image plus the edit endpoint. Render one canonical shot at high quality, then send every subsequent request through fal-ai/gpt-image-2/edit with that canonical as image_urls and a prompt that describes only the delta. Color grade, lens feel, and overall composition carry forward. Batch this in parallel at concurrency 8 to 10 and you will finish a 200-asset campaign in a working afternoon.

Natively no; the model takes a single prompt plus optional reference images and returns a single-pass render. What you can build on top of that is a stateful conversation by feeding each new turn the previous image as image_urls and the diff prompt as text. fal's async queue makes this cheap because each call returns a request ID your app can await. The conversational feel in the ChatGPT UI is the frontend holding that state for you, not a server-side multi-turn mode.

At quality=low, a 1024x1024 typically returns in 2 to 4 seconds on fal.ai. Quality=medium at the same size runs 4 to 8 seconds. Quality=high at 1536x1024 sits in the 10 to 18 second band, with the 4K tier reaching 20 to 30 seconds on hero renders. Latency is stable after the first warm call; the fal queue does not cold start. For user-facing UX that cannot sit behind a spinner, render a low-tier placeholder immediately and upgrade to medium in the background.

That is one of its strongest surfaces. GPT Image 2 will produce a browser screenshot with a correctly laid out nav, button copy that reads the way you wrote it, and microcopy below input fields that honors the quoted text in the prompt. Wrap every UI string you need preserved in single quotes inside the prompt and the model treats it as a literal. Complex multi-state UIs (hover, focus, error) still benefit from describing them as separate renders rather than trying to pack three states into one frame.

fal.ai returns a structured error when the underlying model refuses a prompt; the HTTP status is typically 400 with a detail that surfaces the policy category (sexual content, violence, hateful imagery, public figure likeness, trademarked IP). Your app should log the category, show the user a human-readable message, and not retry automatically. Building an auto-retry on a policy refusal is the fastest way to land on a rate-limit list. For borderline creative work, run the prompt through a quick moderation check before calling the image endpoint.

Put the exact string in single quotes and keep each quoted block under roughly 12 words. Separate distinct copy blocks in the same scene into their own quoted strings so the model does not merge them. Specify the typeface feel (for example: 'a bold condensed sans'), the color, and the alignment. For long paragraphs on a poster, describe the hierarchy explicitly: 'a headline reading X, a supporting sentence reading Y, a caption reading Z.' Avoid mixing fonts across the same sentence; the model will pick one and commit.

fal.ai returns 429 when you exceed the queue's per-account limit. The canonical pattern is exponential backoff with jitter: wait 1 second, then 2, then 4, then 8 before giving up, each time adding a random 0 to 500 ms jitter so you do not synchronize with other retrying clients. The fal.subscribe helper already retries transient network errors; add your own wrapping layer for 429 and 5xx so your worker pool does not starve on one slow job.

Historically yes, at least slightly. Codename preview endpoints have a higher refusal rate on edge prompts and tend to run at lower sampling budgets to keep Arena costs down. When the public fal-ai/gpt-image-2 lands, expect the same capability ceiling but noticeably better edge case coverage and a small but measurable jump in quality at the high tier. Bench your prompt library on the preview so you know which ones to re-bench on the public release.

OpenAI's docs cover the abstract model surface. This site focuses on fal.ai specifically: endpoint paths, pricing tiers as actually billed, real TypeScript and Python examples that compile, migration notes from other providers, and prompt patterns tested against the fal-ai/gpt-image-2 preview. We update the pricing table and the FAQ whenever fal.ai/pricing shifts, so the numbers on this page are the ones you will actually see on your invoice.