The numbers

Meta released Muse Ember on Monday, The Next Web reports, and aimed it at the routine work AI agents do over and over, such as calling tools, routing requests and pulling data out of long documents. Ember lists at $0.10 per million units of input and $0.50 per million units of output. Rivals Anthropic and OpenAI charge exactly that for Claude Haiku 5.5 and GPT-6 Luna, according to The Next Web. On the Intelligence Index from Artificial Analysis, which The Next Web describes as an independent benchmarking firm, Ember and Haiku 5.5 each score 36 at their highest setting and Luna scores 28. Artificial Analysis says the index combines 9 tests, from office work to coding.

$0.14

Cost per task

36

Intelligence Index score

3.2

Minutes per task

37.8%

AutomationBench score

Source: as reported by The Next Web, Meta, Artificial Analysis, VKTR, Startup Fortune and MacRumors.

Cost per finished task

Artificial Analysis also prices a whole task, not only a million units. It works out a cost per task from everything each tool reads and writes while it completes the index, weighted across the tests. On that measure Ember costs $0.14 per task, Haiku 5.5 costs $0.41 and Luna costs $0.15, according to Artificial Analysis and The Next Web.

Working through the index, Haiku 5.5 produced 298M units of output, Ember 115M and Luna 110M, Artificial Analysis reports.

Meta's launch post says its own cost figures for Ember are estimates at proposed pricing. Meta's AI chief Alexandr Wang put the claim more bluntly on X: “Our teeny model muse ember is now on the efficient frontier, matching haiku at 3x cheaper cost per task, and smarter than luna”.

Minutes per task

At their highest setting, Artificial Analysis times Ember at 3.2 minutes per index task, Luna at 12.1 minutes and Haiku 5.5 at 14.8 minutes.

The wait before the first word of an answer is 4.37 seconds for Ember, 71.56 seconds for Luna and 182.25 seconds for Haiku 5.5, Artificial Analysis reports.

Artificial Analysis describes Claude Haiku 5.5 at its highest setting as notably slow per task.

Scores on agent and coding tests

On AutomationBench, which tests whole business workflows, Ember scored 37.8%, Luna 20.2% and Haiku 5.5 13.9%. In Meta's other agent and coding results, Ember comes out ahead of both, The Next Web reports.

Haiku 5.5 wins on FrontierCode 2.0, which checks whether the people who run a software project would accept a tool's code changes. It scored 34.9% there, against 27.2% for Ember, according to The Next Web. The Next Web notes that several of Meta's other results come from Meta's own internal tests.

How much each can read at once

Artificial Analysis lists a 1M window for all three, meaning each can hold about a million units of text in one go. Meta says Ember can also write up to 512K units in a single answer.

Ember can take in text, images, audio and video, according to The Next Web.

Above 100,000 units, Haiku 5.5 charges $0.50 per million units of input and $2.50 per million units of output, VKTR and Startup Fortune report. Haiku 5.5 also counts the same text as roughly 30% more units than Haiku 4.5 did, according to VKTR.

Where you can use each, and since when

Ember is available through Meta's own paid service for software builders and through Muse Code, Meta's coding tool, The Next Web reports. Claude Haiku 5.5, released on 7 October, is priced from $0.10 per million units of input, according to VKTR. It runs on Claude Platform, Amazon Bedrock, Google Cloud and Microsoft Foundry, according to VKTR, and Anthropic has set a date until which it will keep Haiku 5.5 available. OpenAI launched GPT-6 Luna on 22 September, Startup Fortune reports. MacRumors says Luna went to ChatGPT Work and Codex users on Plus, Pro, Business, Enterprise and Edu plans, and to Free and Go users in the desktop app.

Meta lists calling tools, small coding jobs, routing requests, pulling out data and reading long documents for Ember. Anthropic names classification, routing, extraction and helper agents that a bigger agent hands work to, VKTR reports. MacRumors says OpenAI tuned Luna for fast responses and higher volume.

Two price cuts before Ember

Haiku 5.5 costs a tenth of Haiku 4.5, which is listed at $1 and $5, VKTR reports. GPT-6 Luna costs $0.10 per million units of input and $0.50 per million units of output, down from $0.20 and $1.20, MacRumors reports. Writing about Anthropic, OpenAI and xAI, Startup Fortune says the labs now compete as much on price as on test scores.

Which one fits which job

On these figures, Ember is the cheapest and quickest way to reach a score of 36 on routine agent work, if Meta's estimated pricing holds. Luna costs about the same per task as Ember but scores lower on the index. Haiku 5.5 costs the most per task and is the slowest of the three at its highest setting, even though Anthropic aims it at high-volume work where speed matters, VKTR reports. Yet it is the one to beat on code that a project would accept. Ember is also the only one of the three that takes audio and video, while Artificial Analysis lists text and image input for the other two. For most businesses the sensible answer is a split: send high-volume sorting and data work to the cheapest tool that passes your own test, and keep code changes on the tool that scores best on them, with a person checking.

How we found this story

We started from Meta's launch post and The Next Web's report, then asked how each rival compares on the measures an owner pays for. Artificial Analysis publishes a page for each of the three tools with the same tests, costs and timings, so we used those three pages for the head to head. VKTR and Startup Fortune filled in Haiku 5.5's launch, outlets and long-document prices, and MacRumors covered GPT-6 Luna's launch and where it can be used. In all we read eight pages from six publishers, and where a figure comes only from Meta's own tests, we say so.

Our take

Our view, drawn from the reporting in this story. It is analysis, not news.

The price per million units has stopped telling buyers anything useful. All three makers now list the same price, so the bill depends on how much a tool writes and how long it runs to finish your job. On the independent figures, that puts Ember ahead for routine agent work and leaves Haiku 5.5 as the stronger coder. But Ember is the newest of the three, its cost figures are Meta's estimates, and it is sold through fewer outlets than Haiku 5.5. A business that tests each tool on its own jobs, and budgets per finished task rather than per headline price, will spend less than one that picks a brand.

What we don't know yet

Meta calls its cost figures for Ember estimates at proposed pricing, so the final bill per task is not yet known. None of our sources says whether Ember or Luna charge more for very long requests, as Haiku 5.5 does above 100,000 units. None lists the countries where each tool can be used. The Artificial Analysis figures are for each tool at its highest setting; at lower settings the scores, costs and timings will differ. Only Anthropic gives a date until which its tool stays on sale.

What to do in your business tomorrow

  1. Compare Muse Ember, Claude Haiku 5.5 and GPT-6 Luna on one routine job your agent already does, and record the cost and minutes for each finished task.
  2. Try your three longest documents on Claude Haiku 5.5 and note whether any request passes 100,000 units, where its price rises to $0.50 in and $2.50 out per million.
  3. Set a rule this week that a person checks every code change a cheap AI tool makes before it goes live, since the best of the three scored 34.9% on a test of whether its code changes would be accepted.

Sources

  • The Next Web: Meta's Muse Ember matches Claude Haiku at a third of the cost per task
  • Meta: Introducing Our Smallest, Most Efficient Model: Muse Ember
  • Artificial Analysis: Muse Ember: Intelligence, Performance and Price Analysis
  • Artificial Analysis: Claude Haiku 5.5: Intelligence, Performance and Price Analysis
  • Artificial Analysis: GPT-6 Luna: Intelligence, Performance and Price Analysis
  • VKTR: Anthropic Cuts Small-Model Pricing 90% With Claude Haiku 5.5
  • Startup Fortune: Anthropic launches Claude Haiku 5.5 with prices cut up to 90 percent
  • MacRumors: OpenAI's New GPT-6 Sol and Luna Models Bring Astra Improvements to Cheaper Tiers