GPT Image API Review

$10 gets you 1,000 credits, roughly 1,000 low-quality images, valid one year. Every tier claims savings versus OpenAI directly — 10% at Starter, up to 25% at the $5,000 Ultra tier. That's the whole pitch: same model, markup arbitrage on volume.

The product itself is competent. Async task handling via POST /v1/async/images/generations, seeded reruns for review consistency, and dedicated CJK guidance for literal text placement are real, documented features, not vague promises. The image edit endpoint for updating brand-approved assets without full regeneration is a genuinely useful primitive for production teams.

Tradeoff: you're dependent on OpenAI's model roadmap and pricing, with this vendor as a pass-through layer. An announcements page exists, no pricing-page trial, no visible security or compliance detail for teams handling client assets. Fine for prompt iteration and batch social content. Riskier if you need contractual guarantees on uptime or data handling.

The endpoint design is unremarkable in a good way: POST /v1/images/generations, an async task/poll route for batch jobs, and an edits endpoint that lets you patch a brand-approved asset instead of regenerating it. That's the right shape for production pipelines — edit-in-place saves both cost and review cycles. The documented prompt patterns (quoted literal copy, CJK validation guidance, seeded reruns for reproducibility) show someone has actually shipped text-heavy generation work before, not just wrapped an OpenAI call.

What's missing is any SLA or rate-limit documentation — I found docs, a pricing page, an announcements page and a blog. Credits expire at 1 year, sold in one-time packs from $10 to $5,000, which is a strange billing model for something calling itself usage-based.

If you adopt this, in three years you have a dependency on a third party's access to GPT Image 2, with your own reliability capped by theirs. Fine for prototyping or a bounded campaign pipeline; risky as the only path to a model you can't call directly.

$10 buys 1,000 credits. At low-quality 1K resolution, that's ~1,000 images — $0.01 each. Scale to Business at $100 for 10,500 images, same math, better bonus credits. No monthly fee, no seat pricing. That's rare in this category.

TCO math gets fuzzy fast. Quality tiers multiply credit cost 4x from low to high. A team running high-fidelity packaging comps burns through Growth's 20,000 credits quicker than the marketing math implies. Credits expire in 1 year — use them or lose them. No published overage rate beyond buying the next tier up.

Each tier claims savings "vs OpenAI" — 10% at Starter, 25% at Ultra. No named competitor pricing shown to verify that math myself. Billing is one-time purchase, not invoiced — fine for a dev card, friction for procurement wanting POs and terms.

Endpoint shape is sane: sync POST /v1/images/generations for quick calls, async with GET /v1/tasks/{id} polling for batch work. That's the correct split — I've been burned by APIs that only offer sync and time out on real workloads. Edit route (POST /v1/images/edits) instead of full regen for brand-approved assets is a good call; saves credits and keeps layout stable.

Seeded reruns plus 'store prompts and seeds alongside each request' is the kind of advice that only shows up when someone got burned in a review cycle. CJK guidance is oddly specific — validating spelling/placement/legibility before publish reads like a postmortem, not marketing copy.

Friction: pricing is credit packs valid for 1 year, not metered usage despite being labeled usage-based — $10 gets 1,000 credits, that's a budgeting exercise, not a bill. No webhook confirmation in the docs excerpt beyond a buyer-question mention, so I'd want to see the actual payload before trusting production batch runs. No pricing tier for pay-as-you-go if you're prototyping small.

This is a dev tool, so 'daily polish' means: does the doc read like someone actually shipped with it. It mostly does — POST /v1/images/generations, async polling via GET /v1/tasks/{id}, an edit route so you're not regenerating a whole poster to fix one typo. That's the right shape for real production work, not just demo prompts.

Onboarding is a scoped API key and a curl command, which is homework, but familiar homework if you've used any image API before. No free trial though — you're in at $10 for 1,000 credits before you know if CJK text guidance actually holds up on your fonts. That's a real cost of finding out.

The pricing ladder (up to $5,000 for 500K credits) tells me this is built for teams who already know they need volume, not folks kicking tires. Seeded reruns for reproducibility is a nice touch — most teams forget review workflows until month three, and here it's documented on day one. Mobile isn't really the point; this lives in your build pipeline, not your pocket.

"Save 10% vs OpenAI" is on every single tier. That's not a differentiator, that's an admission — this is GPT Image 2 behind a markup arbitrage, not a distinct model or capability. Fine if the convenience is worth it. Not fine if you think you're buying something OpenAI doesn't already offer directly.

The docs are genuinely specific: POST /v1/images/generations, async polling via /v1/tasks/{id}, seeded reruns, CJK guidance. That's real engineering, not vapor. Credits expire after 1 year, which is a fair constraint most vendors bury.

Exit portability is the real question. If this shuts down, you're rewriting to hit OpenAI's endpoint directly — probably a day of work, not a migration project. An announcements page and a blog exist, no pricing-page tiers beyond credits. Fine for a thin layer. Just know what you're paying the markup for.