Skip to content
AI Radar by Wagner Solution
CS
Menu
major New models verified update

Muse Wonder Full scored 56 points in Artificial Analysis Intelligence Index v4.3

confirmed by 4 independent sources updated 1 h ago

Muse Wonder Full scored 56 points in Artificial Analysis Intelligence Index v4.3, and Muse Wonder scored 52 points. Only Muse Wonder is available through Meta Model API; Muse Code subscribers can use both.

According to The Decoder, Muse Wonder Full scored 56 points in Artificial Analysis Intelligence Index v4.3 and Muse Wonder scored 52 points; the previous model, Muse Spark 1.3, scored 48 points. The article also reports an improvement in the autonomous coding test Terminal-Bench 4.0 from 33 to 57 %. However, the results are not better across the board: in the factual knowledge test AA-Omniscience, both new models trail the cited competing models.

Meta released both models for Muse Code subscribers and also made Muse Wonder available through Meta Model API. TechCrunch clarifies that Muse Wonder Full will not be offered through an API. The sources differ on both the timing and scope of access: Wired speaks generally of future availability through an API and Muse Code, while The Verge and The Decoder report current availability, with the stronger variant restricted to Muse Code.

Both models support a context of one million tokens and accept text, images, video and audio. According to Meta, Active Memory retains relevant information across conversations, and Active Focus helps the models continue working on difficult tasks, try different approaches and recover from errors. The company promises to publish a full report with additional safety evaluations later.

The price of Muse Wonder through the API is 4 USD per million input tokens and 20 USD per million output tokens. The Decoder reports rates of 1.25 USD per million input tokens and 4.25 USD per million output tokens for the previous model, Muse Spark 1.3. According to measurements by Artificial Analysis, the cost per task in the cited index rose from 1.60 USD to 4.52 USD; these are the costs of a specific test, not a universal price per task.

What changed

Why it matters

The results suggest an improvement in autonomous coding, but do not by themselves guarantee the same benefit on your own tasks. The long context and multiple input types expand the possibilities for working with extensive source material. For occasional use, costs depend on the token usage of the specific task; at high usage volumes, higher rates for both input and output affect the budget. In addition, the stronger model, Muse Wonder Full, cannot be selected for integration through an API.

Release card

Muse Wonder Full

Meta

Context
The context window is one million tokens long.
Inputs
The model accepts text, images, video and audio.
Availability
The model is available to Muse Code subscribers. According to TechCrunch, it will not be offered through an API.
Documented measurements
  • Artificial Analysis Intelligence Index v4.3 56 bodů The model scored 56 points in the overall capability index.
  • Coding Agents Index 68 bodů In the coding agent evaluation, the model achieved the same score as Claude Code with Sonnet 5.5.
  • AA-Omniscience 30 bodů In the factual knowledge test, the model trailed Claude Opus 5.5, GPT-6 Astra and Gemini 4 Argon.
  • ProofBench 94 ze 100 úloh The model solved 94 out of 100 problems in a non-public mathematical proof test.
According to the sources, it is suitable for
  • Coding with agents.
  • Solving mathematical proof problems.
  • Working with information across conversations, according to the description of Active Memory from Meta.
Documented limits
  • Tasks requiring extensive factual knowledge are a weaker area for the model, according to its AA-Omniscience result.

The Decoder compares the model with the previous model, Muse Spark 1.3, and describes improvements primarily in coding and mathematics. However, in the factual knowledge test, the model trails Claude Opus 5.5, GPT-6 Astra and Gemini 4 Argon.

The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →

What was added since the original report

Verified updates

  1. New verified information

    Muse Wonder Full scored 56 points in Artificial Analysis Intelligence Index v4.3.; Muse Wonder scored 52 points in the same index.; Both models support a context of one million tokens.; Both models accept text, images, video and audio.; Performance in Terminal-Bench 4.0 rose from 33 to 57 percent compared with the predecessor.

    • Muse Wonder Full scored 56 points in Artificial Analysis Intelligence Index v4.3.
    • Muse Wonder scored 52 points in the same index.
    • Both models support a context of one million tokens.
    • Both models accept text, images, video and audio.
    • Performance in Terminal-Bench 4.0 rose from 33 to 57 percent compared with the predecessor.
  2. Correction or contradiction

    Muse Wonder costs 4 USD per million input tokens.; Muse Wonder costs 20 USD per million output tokens.; Muse Wonder Full is available only to Muse Code subscribers and will not be offered through an API.

    • Muse Wonder costs 4 USD per million input tokens.
    • Muse Wonder costs 20 USD per million output tokens.
    • Muse Wonder Full is available only to Muse Code subscribers and will not be offered through an API.
  3. Correction or contradiction

    The new article describes availability through an API and Muse Code as future availability, while the event lists it as current.; According to the company, the best version of the model is intended to power features across Meta platforms, including the personal agent Muse.

    • The new article describes availability through an API and Muse Code as future availability, while the event lists it as current.
    • According to the company, the best version of the model is intended to power features across Meta platforms, including the personal agent Muse.
  4. New verified information

    Muse Wonder Full is available to Muse Code subscribers.; Meta plans to deploy Muse Wonder in the Muse agent for subscribers first.; Meta plans to develop specialized models for all Muse users.

    • Muse Wonder Full is available to Muse Code subscribers.
    • Meta plans to deploy Muse Wonder in the Muse agent for subscribers first.
    • Meta plans to develop specialized models for all Muse users.

Two audiences, two different impacts

What this means

01

For individuals

Muse Code subscribers can try Muse Wonder Full for more demanding coding tasks. However, according to the cited tests, better coding results do not imply equally strong factual knowledge.

What to do If you have a Muse Code subscription, compare both models on one of your own coding tasks and check the result for correctness.
More practical updates →
02

For a business

Muse Wonder is available for business integrations through Meta Model API, while Muse Wonder Full remains restricted to Muse Code subscribers. Higher rates compared with Muse Spark 1.3 change the costs of automated tasks.

Development
What to decide Before deploying Muse Wonder, recalculate the costs of representative tasks based on input and output token usage separately.
More business impacts →
Active Focus Active Memory Artificial Analysis MemoryBench Meta Meta Model API Muse Code Muse Wonder Muse Wonder Full Terminal-Bench 4.0

Check the original

Event sources

confirmed by 4 independent sources · 5 publishers, 4 independent. We count feeds from the same owner only once.

5
The Verge AI independent context · first detected Meta has some Wonder Full new AI models Wired — AI section independent context Muse Wonder Puts Meta Back at the AI Frontier Meta Newsroom — AI primary source Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents TechCrunch AI independent context Meta releases its first frontier model, Muse Wonder The Decoder (daily AI news) independent context Meta's new frontier model Muse Wonder Full breaks into the AI top tier