How does the budget cap in compare_models work?

You (or your agent) pass a maximum spend; the server orders candidate models to fit under it and refuses to exceed the cap mid-run. The response itemizes what each model actually cost — the cap is enforced server-side, not on the honor system.

What comes back from a bakeoff?

Per model: the generated output, its real billed cost, and its wall-clock latency. That turns model choice into a data decision — an agent can codify rules like “fastest under $0.05” or “best text rendering regardless of price”.

Why not just call generate_image several times?

You can, but compare_models runs candidates in parallel, enforces one shared budget, and returns a structured comparison in a single tool result — one round-trip for the agent instead of N calls it has to orchestrate and total up itself.

What does a sweep of every image model cost?

Adding every text-to-image model's list price — flux-schnell $0.003, z-image-turbo $0.01, nano-banana $0.039, flux-2-pro $0.03, seedream and recraft-v3 $0.04, nano-banana-2 $0.067, nano-banana-pro $0.134, gpt-image-2 $0.128 — comes to $0.491 for all 9. The two edit-only models (runway-gen4-image-turbo $0.02, qwen-image-edit-plus $0.03) need a reference image, so they sit out a text-only sweep. Failures refund, so you pay only for outputs that arrived.

Can bakeoffs cover video models too?

Video is priced per second — an 8-second clip runs $0.288 on seedance-lite up to $3.20 on veo-3.1 — so sweeps get real money fast. Compare on images to pick a style direction first, then commit one or two video models. estimate_cost will price the video leg before the agent runs it.

Which client is best for running comparisons?

All of them work; Claude Code and Codex are strongest because the agent can codify the winner into your repo (e.g., writing the chosen model into a config), while claude.ai and ChatGPT are best for eyeballing results interactively.

Create free key

Model bakeoffsStorm

Model bakeoffs, under a budget cap.

compare_models is the tool the others are jealous of: one prompt, several models, a hard spending cap — and every result comes back with its true cost and latency. Your agent stops guessing which model is best.

Connect in 60 seconds

compare_models — race the field under one budget4 engines · fixed prices · pick the keeper

A real Z-Image Turbo run through NinjaChat — **Z-Image Turbo**1.1s

A real GPT Image 2 run through NinjaChat — **GPT Image 2**3.4s

A real FLUX.2 Pro run through NinjaChat — **FLUX.2 Pro**4.1s

A real Seedream 4.5 run through NinjaChat — **Seedream 4.5**3.8s

Paste this. Watch it run.

First ask

Brand mascot, flat vector style — run it across recraft-v3, nano-banana-2, and z-image-turbo, cap at $0.20.

compare_models

3 models · $0.117, under a $0.20 cap

Use case · bakeoff

Four engines. One budget.

compare_models races the field. Keep the light through the wave — then animate the winner.

Golden light through the wave — **Storm**02 · kept

Four engines. One budget. Keep the light through the wave.

Models

The field your budget picks from.

An editorial shot of a cobalt coat in white dunes — a real FLUX.2 Pro run

FLUX.2 ProFLUX.2 Pro

A flamenco dancer mid-turn under a single spotlight — a real Seedream run

Seedream 4.5Seedream

A red-sailed boat crossing a mirror-still turquoise fjord — a real Nano Banana 2 run

Google Nano Banana 2Nano Banana 2

A matte espresso machine in a shaft of morning light — a real Nano Banana Pro run

Google Nano Banana ProNB Pro

A red fox mid-leap over a frozen creek at sunrise — a real Nano Banana run

Google Nano BananaNano Banana

A hummingbird frozen mid-air at a fuchsia flower — a real Z-Image Turbo run

Z-Image TurboZ-Image

GPT Image 2's take on the tiny astronaut asleep on a crescent moon

GPT Image 2GPT Image

Tools

21 tools. These do the work here.

Make

generate_imageCreate images and download the full-resolution results.

Decide

compare_modelsTry one prompt across models with a spending limit.
search_modelsCheck model capabilities and prices.

Money

estimate_costCheck the price before generating.
get_account_balanceSee available credit and monthly spending.

Also on the server

create_mediagenerate_speechedit_imagegenerate_videoimage_to_videoedit_videoextend_videoget_generation_statussearch_creationsget_creationsave_creation_aslist_creative_entitiesupload_assetprepare_scrub_videopublish_pageadd_credits

Setup

Connected in about a minute.

Pick your client — the steps for the others stay one tap away.

Claude Code4 steps

Create an API key
Sign in at ninjachat.ai and create a key under Developers → Keys. Keys look like nj_sk_… and carry your spending limit.
2
Add the server with one command
Register NinjaChat as an HTTP MCP server. Use --scope user to make it available in every project.
bash · template
Replace nj_sk_YOUR_KEY with your API key before using this template.
```
claude mcp add --transport http --scope user ninjachat https://www.ninjachat.ai/api/mcp --header "Authorization: Bearer nj_sk_YOUR_KEY"
```
3
Verify the connection
Run /mcp inside Claude Code — ninjachat should list as connected with the ninjachat tools.
text
```
/mcp
```
Put it to work
Start a task in Claude Code, then send your prompt.

Claude4 steps

Open connector settings
In claude.ai, go to Settings → Connectors and choose “Add custom connector”. Your account and workspace must allow custom connectors.
2
Paste the NinjaChat endpoint
Name it NinjaChat and paste the server URL — nothing else to configure.
text
```
https://www.ninjachat.ai/api/mcp
```
Sign in when prompted
Sign in to NinjaChat and approve the connection. No API key needed. Enable NinjaChat in each chat or project where you want to use it.
Ask Claude to make something
Start a new chat and enable NinjaChat in the tools menu.

Want the full walkthrough with troubleshooting? Claude Code guide · Claude guide

Questions, answered.

Keep exploring.

imageAI Image Generation over MCP

11 image models from $0.003/image, callable from any MCP agent

alternativesReplicate MCP Alternative

Curated models, per-output pricing, OAuth for web agents — vs raw model zoo

videoKling Video in Claude

Generate Kling videos from claude.ai — no Kuaishou account needed

NinjaChat MCP in Claude

Add every creative model to claude.ai in 60 seconds — OAuth, no API key

Model Context Protocol NinjaChat MCP docs

Storm · Bakeoff

Stop guessing. Race them.

compare_models under one budget. Keep the winner.

Connect in 60 seconds

Every frame of Storm was generated through the NinjaChat MCP server.