5-minute quickstart — three ways to use it

By now you know what gpt-image-2 is and what it can do. And you want to actually try it.

But the moment you land on OpenAI's site, you'll pause: where do I even start? ChatGPT works. The Codex developer preview works. The API works. And in China there's a stack of third-party proxies on top of that. Each entry point has totally different pricing, quotas, and generation speed. Pick the wrong one and you'll either pay several times more, or just fail to get it running.

⚠️ One common mix-up to clear first: Sora's been shut down. It's not a gpt-image-2 entry point. If a tutorial tells you to "open Sora to generate images," that's a 2024 article and it's stale.

This chapter sorts the three real entry points out, gives you a complete prompt you can paste in directly, and breaks down monthly cost in actual US dollars for three scenarios — personal, team, and accessing from China.

---

1. Three Entry Points Compared

For mainland China access, you go through third-party proxies like fal.ai or Apiyi: do the compliance paperwork (use a USD card or top up via reseller), pricing is usually 15–30% higher than official API, but you don't need a VPN and there's a CNY settlement channel.

role hint — inside the prompt, use English role tags like Headline / Subhead to tell the model whether a piece of text is the title or the subtitle. The prompt breakdown below uses this.

---

2. Your First Complete Prompt (Paste It In)

Open ChatGPT (Plus/Pro) and paste this whole block:

Vertical 3:4 social media cover, 1242×1660 pixels.

A young Chinese woman sitting in a Sydney café,
bright daylight from the window,
candid laptop session, blurred coffee cup foreground.

Headline (top center, large bold, white with subtle shadow): "30 天学会 AI 画图"
Subhead (directly below, smaller, soft yellow): "保姆级 prompt 公式"

Style: warm autumn palette, slight vintage film grain, real-person aesthetic.
Exact Chinese text only. No extra words. No duplicate copy.

30–60 seconds later you'll get a vertical cover ready to post on Xiaohongshu. The Chinese text comes out right almost every time, and the mood is warm-coffee-shop.

---

3. Prompt Breakdown (What Each Line Does)

Remember the official OpenAI Cookbook order: setting → subject → text → style → constraints. Treat it like a fill-in-the-blank, not a freewriting prompt.

---

4. Real Pricing Across Three Scenarios

Pricing in actual dollars. None of that vague "cheap / expensive" talk.

Scenario A: Xiaohongshu blogger, 30 posts/month × 5 carousel images = 150 images

→ Solo bloggers should just go with ChatGPT Plus. Less hassle.

Scenario B: e-commerce ops, 500 white-background product shots/month

→ At volume, run API low tier. Ten-plus times cheaper, no rate limit.

Scenario C: poster designer, 50 high-quality KVs/month

→ Per-image price doesn't matter here, quality does. API high tier is roughly ¥1.5 per image. Compared to ¥800 per design from a freelance designer, of course you go with high.

---

5. Notes on Accessing From China

Direct connections to OpenAI are blocked from mainland China. Three workarounds:

1. fal.ai — overseas platform, takes USD card top-ups, model versions sync with official. Medium pricing is around $0.06 per image (about $0.007 more than official).
2. Apiyi and similar relay proxies — CNY settlement, works out of the box. But the model version may lag (when gpt-image-2 launched, some proxies were still serving 1.5). Always confirm the model ID before placing an order.
3. Self-hosted proxy — buy your own USD card and proxy through to the official API. Cheapest, but you need to know what you're doing.

On compliance: when you're calling overseas AI services for commercial content, go through proper channels (company-level filing, content review). Don't run unaudited calls on the company's main network.

---

6. Things That Went Wrong

Trap 1: Thinking Plus = unlimited generations

First week I got hyped and pumped out 200+ images. Friday afternoon, suddenly: "You've reached the limit for this period, please try again later." OpenAI's fair-use policy doesn't spell out N images per minute, but bursting traffic triggers a soft throttle — about 5–10 images per minute and it slows you down. Heavy users have to upgrade to Pro $200 or move to API.

Trap 2: API "model not found" on day one

On 2026-04-25 we tried to call the API right out the gate. model: "gpt-image-2" kept 404'ing. Took us a while to figure out: the API only opens to developers in early May. The launch date (April 21) was just an announcement. Don't read "available now" in a release post and assume the API is live too — works in the web UI ≠ works in API.

Trap 3: Third-party proxy takes the money and disappears

A friend topped up $500 on some proxy site to do a batch run. Halfway through, the site started 503'ing and the community group dissolved. Use big proxies (something like fal.ai, with VC backing). Way more reliable than backroom shops. Even if it's 20% more expensive, don't chase the bargain.

Trap 4: ChatGPT account banned when used from China

Log into ChatGPT directly from a Chinese IP and you're risking a ban. Only accounts registered with a stable overseas IP and an overseas phone number are safe. If you get banned, every image goes with it, and your conversation history is gone too — back up important prompts locally.

---

7. Our Cheat Sheet

The internal team at JR Academy ran this for a month and arrived at this table:

Think of "which entry should I use" like ordering food: just feeding yourself? Convenience store (ChatGPT). Whole family eating? Order delivery (Plus with multiple accounts). Opening a restaurant? Now you're sourcing wholesale (API). The scenario picks the plan, not the other way around.

---

What's Next

Once your prompt is dialed in, the real challenge is consistency — why does the same prompt nail it once and flop the next time?

The next chapter walks through the OpenAI Cookbook's six-component formula (Subject / Setting / Style / Composition / Lighting / Technical specs). Turns "rolling the dice" into "snap together by formula." After that you stop second-guessing what to write — you just fill the six slots.

If you want to crank out a few images right now: take the prompt from §2, swap in your own subject and headline text, and generate four variants. Pick one. Then come back and learn the formula.

---

📷 Simple Prompts Can Make Great Images Too

Two real generations from absurdly simple prompts. Proof that you don't need to write 100-word essays from day one. From awesome-gpt-image (CC BY 4.0).

Case 1: Simple Chinese prompt — childhood family snapshot

Prompt (Chinese):

生成一张 2003 年，用数码相机拍摄的，5 岁的我和爸爸妈妈在中国的家，小区院子里的照片。

One Chinese sentence, and the result captures five things at once: 2003 texture, digital camera grain, 5-year-old kid, Chinese residential courtyard, family vibe. The model fills in the 2003 image quality, the period clothing, the neighborhood feel — that's the semantic compression power of reasoning.

📷 Creator: @pangyusio · Curated by: awesome-gpt-image

Case 2: Simple English prompt — subway candid

Prompt:

A beautiful woman looking at her phone on the subway; a candid photo.

Ten words. Yet the output ships with all the right ingredients — the industrial feel of a subway car, candid-shot quality, the lighting, the wardrobe. Whether short prompts produce usable images is a hard test of model reasoning — and gpt-image-2 leaves MJ / Flux a generation behind on this metric.

📷 Creator: @AntCaveClub · Curated by: awesome-gpt-image