Muse Ember vs Claude Haiku 5.5 vs GPT-6 Luna: Same Price Tag, Different Bills

On Monday 9 November, Meta released Muse Ember, its smallest model in the Muse family. It is built for the work an agent repeats all day: tool calls, small coding steps, routing, pulling fields out of documents. Its price is $0.10 per million input tokens and $0.50 per million output tokens. That is exactly what Claude Haiku 5.5 and GPT-6 Luna cost.

When I pick a cheap model for the repetitive part of a project, the price per million tokens is the first number I look at. With these three it tells me nothing. So I put the three official price sheets side by side and did the sums for a few typical requests. I haven't run Muse Ember myself yet, so there are no test results from me here. The numbers below come from Meta, OpenAI, Anthropic and Artificial Analysis, an independent benchmarking site. Every calculation shows its formula, so you can check it.

Short version: for short prompts the three cost the same per token, but Muse Ember finishes Artificial Analysis's test tasks for about a third of what Haiku 5.5 costs at the same score. With a large cached prompt, Ember is the cheapest of the three. Past 100K tokens Haiku gets expensive fast, and past 200K GPT-6 Luna is the cheapest. Ember's weak spot is general knowledge: it often declines to answer instead of guessing.

The price sheets, side by side

All three have a "long prompt" tier: once a request goes over a certain number of input tokens, every token gets a higher rate. The threshold is different for each, and that is where the identical price tag stops being identical. Prices are per million tokens, from the official pages, which I checked on 10 November 2026: Meta's Muse Ember announcement, OpenAI's GPT-6 Luna page and Anthropic's pricing page. Simon Willison put the same table together the morning after the launch.

"Cached input" is the part of your prompt the provider has already seen and stored, like a long system prompt or a document you send on every call. Ember charges $0.004 for it, 60% less than the other two. Cache writes, the first time a prompt gets stored, are billed separately by OpenAI and Anthropic and I've left them out below. Ember's other limits: a 1M-token context window and up to 512K output tokens per answer, against 128K for both Luna and Haiku 5.5.

The number that matters more: cost per task

These are reasoning models. Before answering they "think" in tokens you also pay for, and each model thinks a different amount. Two models with the same price per token can send you very different bills for the same job.

Artificial Analysis measures this. It runs each model through the same set of tasks for its Intelligence Index and divides the bill by the number of tasks. These are its results from 9 November, at each model's two highest reasoning settings (Meta's announcement reproduces the same chart):

Ember at max scores about the same as Haiku 5.5 at max and costs about a third as much per task. Luna at max costs a little more than Ember and scores eight points lower. Artificial Analysis gives two reasons. In its agent tests most of the bill is cached input, where Ember is 60% cheaper. And Ember at max writes roughly a third as many output tokens as Haiku 5.5 at max.

One caution: that is one benchmark's set of tasks. Your agent's tasks have their own mix of long prompts, caching and output, which is why I did the sums below.

Four worked examples

The formula is the same for every row: input tokens × input price + cached tokens × cached price + output tokens × output price, with prices per million tokens. Once a prompt crosses a model's threshold, I priced the whole request at the higher rate. That is how OpenAI words it for Luna. Meta and Anthropic list their tiers by prompt size. These examples count visible tokens only. Reasoning tokens are billed as output on top, and that is the part where Ember's per-task advantage comes from.

1. Short call, 20K tokens in, 2K out. A routing step or a tool call with some context. All three models are under their thresholds: 20,000 × $0.10 + 2,000 × $0.50 = $0.003 per call, or $30 per 10,000 calls, for each of them. Here the list price is the real price, and the only difference is how many reasoning tokens each model spends.

2. Medium document, 150K tokens in, 2K out. A contract or a long log file. Ember and Luna are still under their thresholds: 150,000 × $0.10 + 2,000 × $0.50 = $0.016 per call, $160 per 10,000. Haiku 5.5 is over its 100K threshold: 150,000 × $0.50 + 2,000 × $2.50 = $0.080 per call, $800 per 10,000. Five times the price for the same request.

The gap starts at 200K, not at 272K. At 250K tokens in and 2K out, Luna is still under its threshold: 250,000 × $0.10 + 2,000 × $0.50 = $0.026 per call. Ember is already over 200K: 250,000 × $0.30 + 2,000 × $1.50 = $0.078. Haiku 5.5: 250,000 × $0.50 + 2,000 × $2.50 = $0.130. Luna is the cheapest of the three from 200K onwards.

3. Large document set, 400K tokens in, 2K out. Now all three are over. Luna: 400,000 × $0.20 + 2,000 × $0.75 = $0.0815, so $815 per 10,000. Ember: 400,000 × $0.30 + 2,000 × $1.50 = $0.123, so $1,230. Haiku 5.5: 400,000 × $0.50 + 2,000 × $2.50 = $0.205, so $2,050. At this size Luna is the cheapest of the three.

4. Agent loop with a cached prompt, 60K cached + 2K new + 1K out. The same system prompt and tool list on every step, plus a little new text. Ember: 60,000 × $0.004 + 2,000 × $0.10 + 1,000 × $0.50 = $0.00094 per step, $9.40 per 10,000 steps. Luna and Haiku 5.5: 60,000 × $0.01 + 2,000 × $0.10 + 1,000 × $0.50 = $0.0013 per step, $13 per 10,000. Ember is about 28% cheaper, before counting reasoning tokens.

To run your own numbers for Luna and Haiku 5.5, use the Pickurai token calculator. For Ember, put its prices from the table into the same formula.

Where Muse Ember is weaker

General knowledge. On Artificial Analysis's knowledge test, AA-Omniscience, Ember gets 21.1% right, the lowest of the similarly intelligent models it compared. Haiku 5.5 at max gets 34.7%. The reason is mostly that Ember declines to answer 64% of the questions. That keeps its hallucination rate, how often it makes something up, at 17.4%, half of Haiku 5.5's. For an agent that should look things up instead of guessing, that's fine. For a chatbot that should simply know things, it isn't.

Harder coding. On FrontierCode 2.0, which judges whether real engineering changes could be merged, Haiku 5.5 scores 34.9% and Ember 27.2%. Meta itself suggests pairing Ember with its bigger Muse Wonder for the hardest problems.

Benchmarks still disagree. Some numbers differ between Meta's announcement and Artificial Analysis's write-up from the same day. Haiku 5.5 on Terminal-Bench 4.0 is 39.2% in Meta's chart and 32.8% in Artificial Analysis's text. Ember on AutomationBench-AA is 37.8% in one and 68.1% in the other. Until that is cleared up, I wouldn't choose on either figure.

Where you can use it. Ember is on Meta's own API as muse-ember-1.0, with support for the Responses API, and in Muse Code, Meta's coding agent. Meta says the Muse models are also on Google Cloud and Oracle Cloud. Luna and Haiku 5.5 are on more platforms and more tools today. If your stack only talks to OpenAI or Anthropic, switching has a cost the price table doesn't show.

Which one to pick

If you already run Haiku 5.5 or Luna as the cheap model in an agent, Ember is worth a test this week. Take one day of real requests, note their sizes and how much of each is cached, and price them with the formula above. Then run a sample through Ember and compare the bill and the results, not the price sheet. If most of your requests are short and cached, the saving should show up quickly. If they're huge, Luna is still the cheaper one.

If you're not sure which AI tool fits how you work and what you want to spend, our form asks six questions and gives you a short list.