On August 28, 2026, Meta CAO Alexandr Wang announced that Muse Image is live on the Meta Model API at $0.01 per image:
"muse image is live on the meta model api — $0.01/image — one of the best price-to-quality ratios for production volumes."
This is a distinct event from the July 7 Muse Image launch, which introduced the model inside Meta's consumer apps as an agent that searches, codes, and self-refines. What changed on August 28 is developer access and a price: builders and agents can now call Muse Image programmatically through the same Meta Model API that shipped with Muse Spark 1.1 in July.
TL;DR — what people are asking
| Question | Answer |
|---|---|
| Price? | $0.01 per image on the Meta Model API |
| Quality? | ~1,280 Arena Elo (arena.ai, snapshot 2026-08-25) |
| When? | Announced August 28, 2026 by Meta CAO Alexandr Wang |
| API? | Meta Model API — same OpenAI-compatible surface as Muse Spark 1.1 |
| Top scorer on the chart? | GPT Image 2 medium — ~1,383 Elo at ~$0.055 |
| Cheapest on the chart? | Z-Image Turbo — ~$0.005 at ~1,085 Elo |
| Big gaps? | Aggressive violence filter, no image-to-text, sub-2K output |
Where Muse Image lands on the price-to-quality frontier
Wang's post included a "Text-to-Image Elo vs. Cost" chart — Elo from arena.ai, pricing from Artificial Analysis plus vendor pages, snapshot 2026-08-25. Muse Image sits on the Pareto frontier: for its ~1,280 Elo, nothing on the chart is cheaper, and for its $0.01 price, nothing scores higher.
| Model | Approx. price/image | Approx. Arena Elo |
|---|---|---|
| Z-Image Turbo | $0.005 | ~1,085 |
| Muse Image | $0.01 | ~1,280 |
| FLUX.2 [dev] | $0.012 | ~1,147 |
| Nano Banana 2 Lite | $0.02 | ~1,250 |
| Grok Imagine | $0.02 | ~1,172 |
| Seedream 5 Lite | $0.028 | ~1,138 |
| FLUX.2 [pro] | $0.03 | ~1,155 |
| Nano Banana | $0.035 | ~1,150 |
| Grok Imagine 2.0 (low) | $0.04 | ~1,315 |
| Qwen Image 3 Pro | $0.04 | ~1,255 |
| MAI Image 2.5 | $0.045 | ~1,253 |
| GPT Image 2 (medium) | $0.055 | ~1,383 |
| FLUX.2 [flex] | $0.065 | ~1,157 |
| Nano Banana 2 | $0.07 | ~1,262 |
| Seedream 5 Pro | $0.09 | ~1,258 |
| Reve 2.1 | $0.20 | ~1,300 |
The frontier points are Z-Image Turbo (cheapest), Muse Image ($0.01 / 1,280), Grok Imagine 2.0 low ($0.04 / 1,315), and GPT Image 2 medium ($0.055 / ~1,383, the top scorer). Everything else on the chart is dominated — something is both cheaper and better.
The practical read: at production volume, the gap between $0.01 and $0.055 is 5.5x on the image-gen line item. For an agent generating thousands of images a day — product mockups, ad variants, thumbnail tests — Muse Image trades roughly 100 Elo points for that 5.5x. Whether that trade is right depends entirely on whether those 100 points matter for your prompts, which Arena's blind-preference number cannot tell you.
Arena Elo is preference, not task accuracy
Arena Elo is a blind human-preference duel, the same caveat that applied to the July Muse Image Arena ranks and to Reve 2.1's #2 text-to-image claim. It does not measure:
- Text rendering — does the model spell the words on the sign correctly?
- Instruction adherence — did it place the object where you asked?
- Brand safety — will it refuse or mangle your specific content category?
- Consistency — do 50 generations from one prompt hold a coherent style?
For any of those, run your own suite. See the AI benchmarks complete guide for how to build one, and how diffusion image generation works for the mechanics Muse Image's agentic layer sits on top of.
The gaps builders are hitting
Community feedback since the API opened flags three concrete limits:
- The content filter is overly aggressive. Any hint of fighting or injury gets blocked — a problem for game studios, sports media, comics, and news illustration. The NSFW guardrails, by contrast, are described as more reasonable and predictable.
- No image-to-text. Muse Image generates and edits; it does not caption or describe an input image. If your pipeline needs a describe-then-generate loop, you need a separate vision model for the first half.
- Output resolution is below 2K. Multiple requests for larger and 2K outputs. For hero images or print, you are upscaling downstream.
None of these are dealbreakers for the core use case — cheap, high-volume generation for builders and agents — but they narrow the fit. A filter that blocks "fighting" is a hard stop for some content categories no matter how good the price is.
How this fits the Meta Model API story
Meta is assembling an agentic stack behind one API surface:
| Model | Role | API status |
|---|---|---|
| Muse Spark 1.1 | Reasoning, tools, computer use, 1M context | Meta Model API (public preview, July 9) |
| Muse Image | Image generation and editing | Meta Model API, $0.01/image (Aug 28) |
| Muse Video | Audio-native video | Preview, not yet on the API |
The generate-then-act workflow the July Muse Spark 1.1 coverage described — Spark plans, Image renders, Spark executes in a browser — now has both halves callable from one OpenAI-compatible endpoint. That is the same loop engineering pattern applied across models: plan, generate, verify, revise.
It also lands the same week Meta's own consumer agent work was reported to run on Claude while Meta spends heavily on Anthropic inference — Meta ships its own models on its own API for builders while routing some internal agent work elsewhere. Both things are true at once.
What builders should do this week
- Request Meta Model API access if you run OpenAI-compatible image pipelines.
- Run 20-50 of your real prompts through Muse Image, Z-Image Turbo, FLUX.2 [dev], and GPT Image 2 medium — compare on your criteria, not Arena's.
- Test the content filter early with your actual content categories. A launch demo will not surface a "fighting" block.
- Price the full pipeline, not just the generation call — upscaling, a separate captioning model, and retry cost on filter refusals all add up.
- Compare against Microsoft's route — MAI Image 2.5 in Foundry sits at ~$0.045 / ~1,253 Elo, and GPT Image 2 / ChatGPT Images 2.0 is the current top scorer if quality outranks cost for your workload.
Honest limits — read before switching
- Elo is Meta-selected framing — the chart is real (arena.ai, Artificial Analysis) but was shared by Meta's CAO to make Muse Image look good. It does.
- Pricing snapshots move — the $0.005-$0.20 spread was accurate on 2026-08-25; competitors cut prices monthly.
- The filter is a real constraint — not a rough edge that gets patched next week; it reflects a policy choice.
- Sub-2K output caps the use cases without a downstream upscale step.
- "Agentic" is a harness property — the API gives you the model; the search-code-refine loop from July is something you build around it or get inside Meta's apps.
Related on explainx.ai
- Meta Muse Image and Muse Video — the July 7 agentic launch
- Muse Spark 1.1 and the Meta Model API (July 9, 2026)
- Reve 2.1 — #2 Arena text-to-image, top 4K model
- ChatGPT Images 2.0 and GPT Image 2
- MAI-Image-2.5-Pro and MAI-Voice-2-Flash in Foundry
- How diffusion image generation works
- What is loop engineering?
- AI benchmarks complete guide 2026
- Meta's "Hatch" agent runs on Claude while Meta spends on Anthropic inference
Official: Build with Muse Image — Meta for Developers · meta.ai
Pricing, Arena Elo figures, and the frontier chart reflect Meta's August 28, 2026 announcement and the arena.ai / Artificial Analysis snapshot dated 2026-08-25. Verify current pricing, rate limits, resolution caps, and content-filter behavior in Meta developer documentation before committing production volume.