gemini-2.5-flash-image

Specs & limits

• Unusually for an image model, publishes a knowledge cutoff (June 2025)
• function calling and thinking not supported

per Google official docs ↗

One prompt, measured through the gateway

PROMPT A weathered enamel diner mug on a steel counter, the words "OPEN 24H" stencilled on the mug in worn paint, low winter sun raking in from the left, shallow depth of field.

gemini-2.5-flash-image

Model returned 1024×1024 latency 7 s

One prompt, one request per model, no retries and no cherry-picking - the first result each model returned. Neither size nor duration was pinned: each model used its own default, because a request shaped to fit all of them would flatter none. Files here are re-encoded for the web, so judge composition and prompt adherence, not compression.

Use gemini-2.5-flash-image in 30 seconds

OpenAI-compatible: swap the base_url, keep your SDK. POST /v1/images/generations

from openai import OpenAI

client = OpenAI(
    base_url="https://synthorai.io/v1",
    api_key="sk-syn-...",
)

resp = client.images.generate(
    model="gemini-2.5-flash-image",
    prompt="a watercolor lighthouse at dawn",
    size="1024x1024",
)
print(resp.data[0].b64_json[:80])

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://synthorai.io/v1",
  apiKey: "sk-syn-...",
});

const resp = await client.images.generate({
  model: "gemini-2.5-flash-image",
  prompt: "a watercolor lighthouse at dawn",
  size: "1024x1024",
});
console.log(resp.data?.[0]?.b64_json?.slice(0, 80));

curl https://synthorai.io/v1/images/generations \
  -H "Authorization: Bearer sk-syn-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-2.5-flash-image",
    "prompt": "a watercolor lighthouse at dawn",
    "size": "1024x1024"
  }'

package main

import (
	"context"
	"fmt"

	"github.com/openai/openai-go/v3"
	"github.com/openai/openai-go/v3/option"
)

func main() {
	client := openai.NewClient(
		option.WithBaseURL("https://synthorai.io/v1"),
		option.WithAPIKey("sk-syn-..."),
	)
	resp, _ := client.Images.Generate(context.TODO(), openai.ImageGenerateParams{
		Model:  "gemini-2.5-flash-image",
		Prompt: "a watercolor lighthouse at dawn",
		Size:   "1024x1024",
	})
	fmt.Println(resp.Data[0].B64JSON[:80])
}

import com.openai.client.OpenAIClient;
import com.openai.client.okhttp.OpenAIOkHttpClient;
import com.openai.models.images.*;

OpenAIClient client = OpenAIOkHttpClient.builder()
    .baseUrl("https://synthorai.io/v1")
    .apiKey("sk-syn-...")
    .build();

ImagesResponse resp = client.images().generate(
    ImageGenerateParams.builder()
        .model("gemini-2.5-flash-image")
        .prompt("a watercolor lighthouse at dawn")
        .size("1024x1024")
        .build());
System.out.println(resp.data().orElseThrow().get(0).b64Json().orElseThrow().substring(0, 80));

About gemini-2.5-flash-image

• It is designed for high-volume scenarios, conversational image editing, and low-latency workflows, combining native multimodal understanding of image and text input with image-and-text output at 1K, 2K, and 4K, with aspect-ratio control added when the model reached general availability.
• Editing follows the family pattern: multi-turn conversational refinement that carries the previous interaction forward, adding, removing, or modifying elements, semantic masking for inpainting, style transfer, and composition from several reference images.
• Token limits are 65,536 for input and 32,768 for output, and structured outputs, caching, batch processing, and Flex and Priority inference are supported, while function calling, Search grounding, and thinking are not, which is the practical ceiling on how much reasoning it will do about a layout before drawing it.
• Every generation carries a SynthID watermark.
• Google now describes it as the legacy pioneer of the Nano Banana series and recommends new work move to Nano Banana 2 Lite for better quality, faster speeds, and lower cost, so treat it as the compatibility option rather than the default.
• Its knowledge cutoff is June 2025.
• Synthorai serves it behind the same OpenAI-compatible endpoint as its chat models.

FAQ

Related models

Compare