Model and Modality PlaybooksSeptember 13, 2026Flatkey Team

GPT Image vs Imagen API: Pricing Units and Request Checks

A current comparison of GPT Image and Google's post-Imagen image-generation path, with pricing units, request checks, rate-limit checks, and migration guidance.

If you are comparing GPT Image vs Imagen API: Pricing Units and Request Checks in September 2026, start with one correction: Imagen is no longer the current Google image-generation path in the Gemini API. Google's deprecations page lists the Imagen 4 API models as shut down on August 17, 2026, with gemini-3.1-flash-image as the recommended replacement. Google now presents Nano Banana as Gemini's native image generation family.

That changes the buying question. A useful comparison is not "which active endpoint has the nicer sample gallery?" It is: how do OpenAI's GPT Image units compare with Google's current Nano Banana units, and what request checks should a media or product team run before moving traffic?

This guide uses first-party docs checked on September 13, 2026:

OpenAI's current image generation and pricing docs for gpt-image-2.5-sunburst, gpt-image-2.5-flare, and gpt-image-2.
Google's current Gemini API docs for Nano Banana, Gemini image pricing, rate limits, and Imagen deprecations.
Flatkey's live model/pricing feed and public product pages for routing, billing, and endpoint-family checks.

The Quick Answer

For new work, compare GPT Image with Google's current Nano Banana models, not with retired Imagen 4 routes.

Decision point	GPT Image	Google image generation after Imagen	What to check before launch
Current image family	`gpt-image-2.5-sunburst`, `gpt-image-2.5-flare`, and `gpt-image-2` in OpenAI docs	Nano Banana models such as `gemini-3.1-flash-image`, `gemini-3.1-flash-lite-image`, and `gemini-3-pro-image`	Confirm model ID, route, account access, and deprecation status on the day you ship
Pricing unit	Text and image token rates per 1M tokens	Text/image token rates plus resolution-derived image output equivalents	Normalize cost by accepted image, not only by provider list price
Single-shot generation	OpenAI Image API	Gemini Interactions API with image output	Validate required fields, aspect ratio, image size, output format, and output count behavior
Iterative edits	OpenAI Responses API image tool or Image Edits endpoint	Gemini multi-turn interactions with `previous_interaction_id`	Store previous response or interaction IDs only when you need iteration
Rate-limit shape	RPM, RPD, TPM, TPD, IPM, and account/project limits can apply	RPM, TPM, RPD, and IPM can apply, with IPM for Nano Banana image models	Treat rate limits as part of the routing policy, not an afterthought
Migration risk	Model/version changes inside the GPT Image family	Legacy Imagen API routes are already past their August 17, 2026 shutdown date	Block old `imagen-4.0-*` routes in production checks

If you need a practical image API rollout guide, pair this comparison with Flatkey's Image Generation API guide. If you are modeling spend across several AI families, also use the Gemini API pricing guide and the multimodal routing playbook.

Why Imagen Needs A Migration Check Now

Searchers still use "Imagen API" because Imagen was Google's named text-to-image API line. But Google's current docs tell teams to move the center of gravity to Nano Banana.

The migration check is concrete:

Legacy model family	Google status checked September 13, 2026	Recommended replacement in Google docs	Practical result
`imagen-4.0-generate-001`	Shut down August 17, 2026	`gemini-3.1-flash-image`	Do not build new production traffic against this route
`imagen-4.0-ultra-generate-001`	Shut down August 17, 2026	`gemini-3.1-flash-image`	Treat old request templates as migration debt
`imagen-4.0-fast-generate-001`	Shut down August 17, 2026	`gemini-3.1-flash-image`	Use current Gemini image models for fresh evaluation
`gemini-2.5-flash-image`	Deprecated, shutdown date October 2, 2026	`gemini-3.1-flash-image` or `gemini-3.1-flash-lite-image`	Migrate before the shutdown window closes

That is the first request check in any GPT Image vs Imagen API: Pricing Units and Request Checks workflow: make sure the "Imagen" side of the comparison is not an already retired model ID.

Pricing Units: GPT Image

OpenAI prices current GPT Image models by modality. The pricing table checked on September 13, 2026 lists prices per 1M tokens for image and text surfaces.

OpenAI model	Modality	Input	Cached input	Output
`gpt-image-2.5-sunburst`	Image	$8.00 / 1M tokens	$2.00 / 1M tokens	$30.00 / 1M tokens
`gpt-image-2.5-sunburst`	Text	$5.00 / 1M tokens	$1.25 / 1M tokens	Not listed
`gpt-image-2.5-flare`	Image	$8.00 / 1M tokens	$2.00 / 1M tokens	$30.00 / 1M tokens
`gpt-image-2.5-flare`	Text	$5.00 / 1M tokens	$1.25 / 1M tokens	Not listed
`gpt-image-2`	Image	$8.00 / 1M tokens	$2.00 / 1M tokens	$30.00 / 1M tokens
`gpt-image-2`	Text	$5.00 / 1M tokens	$1.25 / 1M tokens	Not listed

OpenAI's pricing page also shows Batch pricing for gpt-image-2, where image input, cached input, and output are listed at $4.00, $1.00, and $15.00 per 1M tokens. Do not apply that batch row to gpt-image-2.5-sunburst or gpt-image-2.5-flare unless the live pricing page lists those models under the Batch tab on your review date.

The important point is the unit shape: GPT Image pricing is not one flat "per request" number. Product teams need to estimate prompt text, image input if editing or using references, generated image output, cache behavior, and whether the workload can tolerate batch latency.

Pricing Units: Google After Imagen

Google's current Gemini image-generation docs describe Nano Banana as Gemini's native image-generation capability. The current pricing page lists image output in tokens, then gives per-image equivalents based on output resolution.

Google model	Paid input unit	Paid output unit	Per-image equivalent shown by Google
`gemini-3.1-flash-image` / Nano Banana 2	$0.50 / 1M tokens for text/image input	$60.00 / 1M tokens for images	$0.045 at 0.5K, $0.067 at 1K, $0.101 at 2K, $0.151 at 4K
`gemini-3.1-flash-lite-image` / Nano Banana 2 Lite	$0.25 / 1M tokens for text/image/video input	$30.00 / 1M tokens for images	$0.0336 at 1K
`gemini-3-pro-image` / Nano Banana Pro	$2.00 / 1M tokens for text/image input, with an image-input equivalent of $0.0011 per image	$120.00 / 1M tokens for images	$0.134 at 1K/2K, $0.24 at 4K

Google also lists Batch pricing for current Gemini image models. For example, the pricing page shows gemini-3.1-flash-image Batch image output at $30.00 per 1M tokens and gemini-3.1-flash-lite-image Batch image output at $15.00 per 1M tokens. That can matter for high-volume creative variant jobs that do not need interactive latency.

The unit lesson is similar to GPT Image but not identical. OpenAI presents GPT Image as text and image token pricing. Google presents Nano Banana as token pricing plus resolution-based image equivalents. If your dashboard only counts "requests," it will hide the cost drivers that actually move the bill.

Request Checks Before You Compare Results

Before you run a quality bakeoff, make both sides pass the same request checklist.

Check	GPT Image	Google current image path	Why it matters
Active model ID	Use `gpt-image-2.5-sunburst`, `gpt-image-2.5-flare`, or another currently documented GPT Image model	Use a current Nano Banana model such as `gemini-3.1-flash-image`; block retired `imagen-4.0-*` IDs	Prevents comparing a live route to a deprecated or shut-down route
API surface	Use Image API for one-shot generations/edits; use Responses API for multi-turn flows	Use Gemini Interactions API for image generation, editing, and multi-turn workflows	The API shape changes state handling and retry policy
Output constraints	Check `n`, quality, size, format, compression, and whether organization verification is required	Check `response_format`, `aspect_ratio`, `image_size`, output modality, and model-specific limitations	Bad parameter defaults can distort cost and quality tests
Reference images	Confirm image input handling for edits or Responses API workflows	Confirm MIME type, base64 payload, and reference-image count limits for the selected model	Creative and ecommerce workflows often depend on references
Safety and rights	Apply policy checks for uploaded assets and generated output	Apply Google's rights/safety guidance and SynthID watermark expectation	Media teams need a review path, not just generated files
Rate limits	Inspect RPM, RPD, TPM, TPD, IPM, project/org limits, and `Retry-After` behavior	Inspect RPM, TPM, RPD, IPM, project-level limits, and preview-model limits	A winning test prompt is not a production capacity plan
Usage accounting	Record input text, image input, image output, cache, failures, and accepted-image status	Record input units, output image resolution, image-token equivalents, search grounding charges if used, failures, and accepted-image status	Finance needs cost per accepted asset, not cost per attempt

Run these checks first. If they fail, the model comparison will be misleading.

Cost Per Accepted Image Beats Cost Per Request

Image generation cost is not finished when the provider returns a file. For a media workflow, a "successful" request usually means the image passed review and can be used.

Use this normalized metric:

cost_per_accepted_image =
  (all_generation_cost + edit_cost + retry_cost + grounding_or_tool_cost)
  / accepted_images

Then segment it:

Segment	What to measure
Draft generation	Cost per usable concept, latency per first result, prompt rejection rate
Image editing	Cost per accepted edit, reference fidelity, number of iterations
Campaign variants	Cost per approved variant, batch queue time, human review load
Product images	Wrong-object rate, brand-rule failures, required retouching
Agent-generated media	Cost by task, fallback rate, and whether the agent retries blindly

This is where a multi-model router matters. Flatkey's positioning is one key, one balance, and one usage ledger across model and tool calls. For a team running image generation next to chat, video, speech, or agent workflows, the operational advantage is not just model access. It is seeing every request, token type, endpoint family, and failure in one place.

How Flatkey Fits This Comparison

Flatkey should not be presented as a magic discount switch for every image workflow. The stronger claim is simpler and source-backed: Flatkey gives teams one API layer for many official models and tools, one balance, one invoice, and a ledger for model and tool usage.

For this topic, the useful Flatkey checks are:

Flatkey check	September 13, 2026 live feed result	How to use it
Endpoint family	`image-generation` is listed with `POST /v1/images/generations`	Confirms an image-generation route family exists in the public router feed
GPT Image rows	`gpt-image-2.5-sunburst`, `gpt-image-2.5-flare`, and `gpt-image-2` appear under image-generation/OpenAI-compatible endpoint types	Use Flatkey when you want GPT Image traffic under the same billing and ledger layer as other AI routes
Gemini image rows	`gemini-3.1-flash-image`, `gemini-3.1-flash-lite-image`, `gemini-3-pro-image`, and `gemini-2.5-flash-image` appear under Gemini/OpenAI-compatible endpoint types	Use the feed to validate current Google-side model availability and avoid stale Imagen assumptions
Retired Imagen rows	No current `imagen-4.0-*` image-generation route was confirmed in the feed check	Do not imply Flatkey routes retired Imagen API traffic unless a future feed check confirms it

If the workflow is one standalone GPT Image prototype, a direct provider integration may be enough. If the workflow sits inside a broader production stack with chat prompts, image edits, usage caps, retries, and cost review, a gateway like Flatkey gives the team one place to standardize request checks and spending.

For model availability and billing review, start with the Flatkey model directory and Flatkey pricing.

A Practical Evaluation Workflow

Use this workflow for GPT Image vs Imagen API: Pricing Units and Request Checks before routing production traffic.

Replace stale Imagen language. If a spec, ticket, or prompt says imagen-4.0-*, update it to the current Google model being tested, usually a Nano Banana model.
Freeze three representative tasks. Include one simple generation, one reference-based edit, and one multi-turn creative revision.
Normalize output requirements. Fix aspect ratio, image size, format, resolution, and review criteria before testing.
Estimate provider units. For GPT Image, model text and image tokens. For Nano Banana, model input tokens and output-resolution equivalents.
Record failed attempts. Count invalid requests, safety blocks, rate limits, timeouts, and review failures separately.
Calculate accepted-image economics. Use cost per accepted asset, not cost per raw request.
Check route availability. Confirm model ID, endpoint family, account tier, organization/project limits, and deprecation status on launch day.
Decide routing policy. Keep direct integration for narrow single-provider tests; use a gateway when several models, teams, or budgets share the workflow.

Which API Should A Media Team Choose?

Choose GPT Image first when your team wants OpenAI's current image models, when your existing app already uses OpenAI SDK patterns, or when the workflow needs the Image API and Responses API split.

Choose Google's current Nano Banana path first when your team is already building in Gemini, needs Gemini's image-specific features, or wants to test Google's current image models after the Imagen shutdown.

Use Flatkey when the answer is not one provider. That is common for media teams because real production stacks mix image generation, edits, chat prompts, asset review, campaign variants, and cost reporting. A shared routing layer lets the team test providers while keeping usage controls and accounting consistent.

FAQ

Is Imagen API still active in the Gemini API?

Google's deprecations page checked on September 13, 2026 lists Imagen 4 generate models with a shutdown date of August 17, 2026. For new Google image-generation work, use the current Nano Banana model family documented under Gemini image generation.

Is GPT Image priced per image or per token?

OpenAI's current GPT Image pricing table lists image-generation models by modality, with text and image rates per 1M tokens. Cost estimates should include prompt text, image input for edits or references, generated image output, cache behavior, and any batch assumptions.

Why does Google's pricing page show per-image numbers?

Google lists Nano Banana image output by tokens and then gives per-image equivalents based on output resolution. For example, gemini-3.1-flash-image image output is priced at $60 per 1M tokens, with listed equivalents for 0.5K, 1K, 2K, and 4K images.

What should I log for an image API comparison?

Log model ID, endpoint, prompt text length, reference-image count, output size, image-token output or per-image equivalent, rate-limit headers or errors, retry reason, review outcome, and accepted-image status.

Where does Flatkey help?

Flatkey helps when image generation is part of a broader AI product workflow. Instead of tracking separate provider keys, invoices, and usage dashboards, teams can route model calls through one key and review usage from one shared balance and ledger.