Suppose an agent pipeline runs 10,000 Artificial Analysis intelligence index tasks a month. According to the single-question cost announced by AA, Muse Ember is US$1,400, GPT-6 Luna is US$1,520, and Claude Haiku 5.5 is US$4,110.
According to AA, the input and output prices of the three companies are the same: US$0.10 per million input tokens and US$0.50 per million output.
The same price list, but the difference in the bill at the end of the month was almost three times.
The price difference is hidden in two places
The first place is cache. Ember's cache input is only $0.004 per million tokens, which is 60% cheaper than Haiku 5.5's cache price. The characteristic of the agent task is that the same system prompt, the same tool definition, and the same code base context are fed repeatedly, and the cache hit rate is very high. This 60% difference will fall on every call.
The second point is whether you talk too much or not. According to AA statistics, when doing the same question, Ember's output token is only about one-third of Haiku 5.5. The output unit price is five times that of the input. If it is reduced by two-thirds, the bill will naturally come down.
Luna is the other way around. Its single question costs $0.152, which is just over one cent behind Ember, but its intelligence index is only 28 points, 8 points lower than the other two. Cheap is cheap, and you get less stuff.
The origin of the three models
Muse Ember was just released by Meta AI on November 9. The model ID is muse-ember-1.0. Meta calls it its “smallest and most efficient” model. There are also Muse Spark 1.3 and the largest Muse Wonder. Officially, Ember is the first model trained with Muse Wonder Full. The parameter quantities and weights are not disclosed. The context has 1 million tokens, and a maximum of 512,000 tokens can be output at a time. Text, images, audios, and videos can all be consumed. The first partners are Box, Cline, Oracle, Cloudflare and You.com.
The Claude Haiku 5.5 is the current model in the Anthropic small model range, the cheapest and fastest in the Claude family.
GPT-6 Luna is the lightest one in the OpenAI GPT-6 family, and its positioning is the same as the previous two.
What can be explained clearly in a table
| project | Muse Ember | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|---|
| AA Intelligence Index | 36 | 36 | 28 |
| AA single question cost | $0.14 | $0.41 | $0.152 |
| Terminal-Bench 4.0 | 40.9% | 39.2% (Meta table) / 32.8% (AA actual measurement) | 16.4% |
| FrontierCode 2.0 | 27.2% | 34.9% | 25.2% |
| Coding Agent Index | 46 | 37 | 31 |
| OSWorld 2.1 (partial score) | 69.5% | 68.9% | 50.5% |
| GDP.pdf | 22.8% | 20.8% | 22.8% |
| AA-Omniscience accuracy | 21.1% | 34.7% | not disclosed |
You should pay attention to this line of Coding Agent Index: each of the three companies uses their own framework for testing. Ember runs in Muse Code, Haiku runs in Claude Code, and Luna runs in Codex. The comparison is a complete set of models plus own tools, not entirely the model itself.
Two numbers that don’t match
AutomationBench-AA is an end-to-end business process test. In the table when Meta was released, Ember was 37.8%, Haiku was 13.9%, and Luna was 20.2%. Many Chinese reports directly quoted this group, so there was a saying that "the business process is more than twice the lead." But in Artificial Analysis’s own post, Ember is written at 68.1%.
The difference is 30 percentage points.
The more likely reason is that the test version or scoring caliber is different, but neither side has come out to explain it so far. Using this to prove that Ember crushes its opponents in business processes is not a solid foundation.
Terminal-Bench 4.0 does something similar. In the Meta table, Haiku 5.5 is 39.2%, marked as the "best published framework" score; AA uses its own unified framework to test, and Haiku is only 32.8%. Ember is at 40.9% on both sides. The same model dropped more than 6 points when changing the test environment, indicating that terminal tasks are very sensitive to the framework. When selecting a model, it is best to retest it with your own tool chain, don't just look at the list.
Ember’s problem: If you don’t know, don’t tell.
AA-Omniscience tests knowledge. Ember's accuracy is only 21.1%, and Haiku 5.5's is 34.7%. Among the models with the same smart gear, Ember is at the bottom.
Its hallucination rate is 17.4%, about half that of Haiku. It sounds like an advantage, but it's a bit subtle when you break it down: Ember directly refused to answer 64% of the knowledge questions.
It rarely compiles because it rarely answers questions.
Put into actual scenarios, this temperament is a mixed bag of good and bad. When doing tool invocation, information extraction, and code subtasks, the answer is in the context, and the model’s own knowledge is not needed. The high rejection rate is safe. If you use it to directly answer open questions from users, more than six out of ten will reply with "not sure", and the experience will be very poor.
Haiku 5.5 is expensive and worth it
Three times the cost of a single question is Haiku’s most obvious shortcoming. And under the premise of the same price, this difference comes from its large output and expensive cache. Users have almost no room for optimization unless they manually reduce the output length.
What it maintains is the quality of writing code. On FrontierCode 2.0, Haiku is 34.9%, Ember is 27.2%, and Luna is 25.2%, leading by nearly 8 points; in AA-Briefcase for document office, Haiku is 1574 Elo, and Ember is 1521. The knowledge aspect is even more than 13 percentage points ahead.
If the subtask is "understand a piece of unfamiliar code and then correct it" rather than "adjust the tool once according to a fixed format", Haiku's extra money will be accounted for.
Luna's position is a bit awkward
Luna's single-question test is a bit more expensive than Ember, but its intelligence index is 8 points lower. Terminal-Bench 4.0 is only 16.4%, less than half of Ember; OSWorld 2.1 is 50.5%, and the other two are around 69%.
The only tie was GDP.pdf, at 22.8%.
The main reason for choosing it is ecological. For a team that is already betting on OpenAI API and Codex, changing models means changing billing, monitoring, and auditing processes. These migration costs cannot be covered by the price difference of one penny per question. But if you are building a new assembly line from scratch, it is difficult to find a technical reason to choose Luna.
Fall into the assembly line
The answers to routing, classification, extracting fields from long documents, and adjusting tools according to schema are all in the context. Ember is two-thirds cheaper, and many rejections are not a problem. It is worth cutting off a small amount of traffic to try it out. The founder of Cline has publicly said that it is very good at sub-agent work, which is an early circumstantial evidence.
Haiku 5.5 is still more stable when it comes to understanding unfamiliar code, modifying patches that can pass the review, or requiring models with built-in common sense.
Luna is reserved for teams that cannot live without the OpenAI ecosystem.
One more thing: Ember has only been released for one day, and Meta’s model API has not been tested by large-scale production traffic. You have to test the current limit, stability, and regional availability yourself. Don’t rush to include the more than $2,700 saved in the first 10,000 missions in your budget.
Related comparisons: "Kimi K3.1 Pro and DeepSeek V4.1 Flash 1017: The difference in each question is eight times, and the extra money only buys terminal programming", "GPT-6.1 Terra and Claude Sonnet 5.5: Terra is five times cheaper at the medium level and slightly better, but it is overtaken at high scores", "LongCat-2.5, GLM-5.3, DeepSeek V4.1 Flash: On Meituan’s 31-line report card, one column is all self-assessment scores.