GPT Image 2 Slow? It's the quality Parameter

If GPT Image 2 is taking two minutes, quality is set to high. We measured 117–132 seconds at high against 17–20 seconds at low — same model, same prompt, same 1024×1024 output. The parameter is not just a quality dial. It is the speed control and the price control, and most people never set it.

Model ID:   openai/gpt-image-2
Endpoint:   /v1/images/generations
Parameter:  quality = low | medium | high
Measured:   low 17-20s · medium ~41s · high 117-132s
Cost span:  $0.0059 -> $0.2107 per image (36x)
Default:    low, on our route (see below)

Measured against the live endpoint on 7 September 2026, 1024×1024, identical prompt, two runs per tier.

The Measurement

Generation time tracks the token count, not the prompt. High produces 36× the tokens of low and takes roughly 7× as long. That is the whole mechanism: quality decides how much the model actually renders, and everything else follows from it.

Cost is computed from the measured tokens at $30 per million output image tokens, the live catalog rate. You can verify any row yourself: 7,024 × $30 ÷ 1,000,000 = $0.2107.

Note the spread. Several published comparisons put low-to-high at 15×. The endpoint returns 36×. If you sized a budget off the 15× figure, the high tier costs more than twice what you planned.

What the Default Actually Is

Omit quality and you get the low tier. A request with no quality and no size returned 201 output tokens — within noise of low’s 196 — and completed in the same time bracket.

This matters because the published guidance contradicts itself. Searching for the default turns up three different answers: that it is high, that it is medium, and that it behaves like auto and varies with prompt length. Those cannot all be true, and none of them match what we measured.

So the practical rule is: set it explicitly. Not because low is wrong, but because a default you did not choose is deciding both your bill and your latency, and the documentation you would check to confirm it is unreliable.

This also explains two opposite complaints about the same model. People who set high say it is unusably slow. People who set nothing say the output looks soft. Both are the same parameter, at opposite ends.

What to Set

• Drafts and iteration: low. 17–20 seconds and just over half a cent. At this price you can generate ten variations for less than one high render, which is usually a better use of the money than one careful expensive attempt.
• Anything user-facing but not final: medium. About 41 seconds and $0.05. This is the tier most work should sit at.
• Finals only: high. Two minutes and 21 cents. Worth it for a hero asset, wasteful in a loop.
• Never leave it unset in production. You are shipping the low tier by accident.

If two minutes is unacceptable and high is genuinely required, the fix is architectural rather than parametric: generate asynchronously and poll, rather than blocking a request on it.

curl -X POST https://api.ofox.io/v1/images/generations \
  -H "Authorization: Bearer $OFOX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-image-2",
    "prompt": "A single red apple on a white table",
    "size": "1024x1024",
    "quality": "low"
  }'

Read usage.output_tokens_details on the response. It reports image_tokens directly, so you can confirm which tier you actually got rather than assuming.

The Comparison That Explains the Confusion

Nano Banana 2 has no quality parameter at all. It generates a flat 1,120 tokens at 1K regardless of what you send it — we tested an 8-token prompt and a 51-token prompt and both billed identically.

These are two different billing shapes, not two prices. Nano Banana 2 is predictable per resolution and gives you no lever. GPT Image 2 gives you a 36× lever and no protection against pulling it the wrong way. Cross-model comparisons that quote a single GPT Image 2 price are picking one of three tiers and usually not telling you which.

Read that table by tier, not by brand: GPT Image 2 at low is the cheapest option here by a wide margin, and at high it is the most expensive by a wide margin. Same model.

Related

• Nano Banana 2 token billing — the flat-rate counterpart, plus the reasoning tokens its Pro sibling adds.
• GPT Image 2 generation failures — when the call errors rather than merely dragging.
• GPT Image 2 release guide — capabilities and the rest of the parameter surface.

Sources

• https://ofox.io/models/openai/gpt-image-2
• https://ofox.io/models/google/gemini-3.1-flash-image

Token counts and timings are from live /v1/images/generations calls on 7 September 2026: two runs per quality tier at 1024×1024 with an identical prompt, plus one call with quality and size omitted to establish the default. Timings were measured wall-clock from the client and include network and queueing, so treat them as the shape of the difference rather than as model inference time — your absolute numbers will differ by route, region and load. Costs are computed from the measured token counts at the live catalog rate of $30 per million output image tokens. This is a small sample on one route on one day, not a specification. The conflicting default-quality claims referenced above are from third-party guides, which we did not re-test; we report only what our own route returned.