GPT Image 2 pricing and rate limits

Standard and Batch token rates

USD per million tokens, from OpenAI's pricing table. These rows refer to gpt-image-2; GPT Image 2.5 is a separate model family.

Batch is a separate asynchronous processing path. Sending several ordinary requests together does not make them Batch requests. Confirm endpoint eligibility and scheduling needs before applying the discounted column.

The image generation guide limits cached input pricing to the image generation tool in the Responses API. Direct Images API requests, including edits, do not receive these cache rates. The guide also says cached token counts are not exposed in the Responses output, so that output alone cannot verify an image-tool cache hit.

1024×1024 image-output estimates

OpenAI's public calculator, set specifically to GPT Image 2 at 1024×1024, gives the following estimates. All exclude text and image input. Explicit quality and size are required to reproduce them; auto can produce different usage.

For medium quality, 1,756 × $30 ÷ 1,000,000 = $0.05268. One thousand such outputs cost $52.68 before inputs. These are calculator estimates, not results of a paid image-generation test.

Prompt and reference-image inputs add to the bill

Suppose a Standard request consumes 100 text input tokens, 1,000 image input tokens and the estimated 1,756 medium output tokens. With no cache discount, its cost is (100 × $5 + 1,000 × $8 + 1,756 × $30) ÷ 1,000,000 = $0.06118. The input counts are illustrative; actual reference images can use different numbers of tokens.

At the published Batch rates, the same hypothetical token counts cost $0.03059. A larger reference image or a second reference image can increase input usage even if output quality and size stay fixed. Budget all attempts that are billed, including revisions you choose not to keep.

GPT Image 2 processes reference-image inputs at high fidelity automatically. The official guide says to omit input_fidelity for this model. You cannot assume a cheaper low-fidelity setting will reduce an edit's cost. Requests through Responses may also incur the main model's token usage in addition to image-generation charges.

Quality and resolution are budget choices

GPT Image 2 supports low, medium and high quality, plus auto selection. The size rules allow flexible dimensions: both edges must be multiples of 16, the longest edge cannot exceed 3840 pixels, the aspect ratio cannot exceed 3:1, and total pixels must be between 655,360 and 8,294,400. Common sizes include 1024×1024, 1536×1024 and 3840×2160.

The square examples above do not price every allowed shape. Use the official calculator for your chosen dimensions and compare actual usage after generation. A quality label is model-specific: GPT Image 2.5 can consume different output tokens at the same label, so do not copy its estimates into a GPT Image 2 budget.

“Rate” can mean price or calling capacity

Token prices determine money spent. Rate limits determine how quickly an organization can call the model. The official model page lists these current limits; check your account's limits before planning a launch.

At 20 images per minute, 1,000 images require at least 50 minutes on image capacity alone. That is an arithmetic lower bound, not a completion-time guarantee: token limits, request latency and failures can slow the queue. A larger tier changes capacity, not the Standard token prices above.

Keep the exact model identifier in your job records. The model page lists gpt-image-2 and the snapshot gpt-image-2-2026-04-21. For other suppliers, use the image API comparison, whose GPT Image 2 row now labels the output-only estimate explicitly.