By Thorsten Meyer | October 13, 2026 | ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Meta closed the capability gap. It did not close the price gap. It priced its new API model like Claude Opus.
Six weeks after Muse Spark 1.3, Meta has shipped two new models: Muse Wonder and Muse Wonder Full. Artificial Analysis published its independent evaluation on October 12, 2026, and the numbers are the most interesting Meta has posted this year.
Muse Wonder Full scores 56 on the Artificial Analysis Intelligence Index v4.3. That is eight points more than Muse Spark 1.3 (48), and it improved on all ten evaluations in the index. Of the 27 configurations I compare here, only Claude Opus 5.5 (max) scores higher, at 58. Claude Sonnet 5.5 (max) ties it at 56.
Muse Wonder, the model you can actually buy through Meta’s API, scores 52. That is exactly what GPT-6.1 Sol (max) scores. The difference is the bill: one Intelligence Index task costs $4.52 on Muse Wonder and $0.72 on Sol.
That gap is the story. Muse Wonder Full shows Meta can build a frontier model. Muse Wonder’s pricing shows Meta wants to be paid like a frontier lab before its API model has earned it.
What Meta shipped
Both models are proprietary, both have a 1-million-token context window, and both accept text, images, speech/audio and video as input. Both return text.
They are not available in the same places:
- Muse Wonder is available through Meta’s API and inside Muse Code, Meta’s coding agent. List price: $4 per million input tokens, $0.20 for cached input, $20 per million output tokens. That is the same list price as Claude Opus 5.5.
- Muse Wonder Full is available only to Muse Code users. It has no per-token price, so nobody outside Meta can calculate what it costs per task.
So the stronger model is the one you can’t buy by the token, and the model you can buy competes one tier lower.
Infographic 01 · Where Meta’s new pair lands
Muse Wonder Full ties for second. Muse Wonder ties GPT-6.1 Sol.
Artificial Analysis scores 27 frontier and open-weight configurations on ten evaluations. Meta’s two new models land near the top of the list. Six weeks earlier, Muse Spark 1.3 scored 48.
Intelligence Index score, 27 configurations
Higher is better. Reasoning effort in brackets, as Artificial Analysis tested it.
Availability: Muse Wonder runs on Meta’s API and in Muse Code. Muse Wonder Full is only available to Muse Code users and has no per-token price.
Source: Artificial Analysis, Muse Wonder analysis (12 Oct 2026) and model leaderboard, read 13 Oct 2026. Intelligence Index v4.3, ten evaluations.
Powered by Thorsten Meyer AI
As an affiliate, we earn on qualifying purchases.
Where Muse Wonder Full is genuinely strong
An index score of 56 could hide uneven results. Here the gains are spread across the board: Muse Wonder Full finishes in the top three on six of the ten evaluations.
| Evaluation | Muse Wonder Full | Context |
|---|---|---|
| Terminal-Bench 4.0 | 57% | Up from 33% for Muse Spark 1.3; tied with Gemini 4 Argon (high) |
| GDPval-AA v2.1 | 1857 Elo, #2 | Claude Opus 5.5 1866, Claude Sonnet 5.5 1838 (within confidence intervals) |
| AA-Briefcase | 1729, #3 | Up from 1581 |
| GDP.pdf | 33%, #1 | Nearest rivals at 31–32% |
| CritPt | 33%, #1 | Nearest rivals at 31–32% |
| AutomationBench-AA | 71%, #3 | |
| SciCode | 63%, #3 | |
| AA-Omniscience | 30 | Up 5 points; hallucination rate down from 33% to 23% |
The Terminal-Bench jump from 33% to 57% is the one I would pay attention to. Terminal-Bench measures whether a model can get real work done in a shell. A 24-point improvement in six weeks is not a tuning gain. It is a different class of agent.
The GDPval-AA result, which covers economically useful knowledge work, makes the same point. Muse Wonder Full sits between Opus 5.5 and Sonnet 5.5, and the three are within each other’s confidence intervals. On this measure, Meta is no longer chasing Anthropic. It is level with it.
The Coding Agents Index confirms it
Artificial Analysis also measures model-plus-harness combinations in its Coding Agents Index:
- Muse Code + Muse Wonder Full: 68
- Claude Code + Claude Sonnet 5.5: 68
- Claude Code + Claude Opus 5.5: 66
- Muse Code + Muse Wonder: 63
- Muse Code + Muse Spark 1.3: 54
Inside Muse Code, Wonder Full ties the best Claude Code configuration. That index measures the model and the harness together, so it does not say Wonder Full would beat Opus 5.5 in Claude Code. It does say that a Muse Code user now has a frontier-class coding agent.
Muse Wonder, one step down
The API model is clearly weaker but still a big step up from Spark 1.3: 52% on Terminal-Bench, 68% on AutomationBench-AA (up from 58%), an AA-Omniscience score of 25, and a 27% hallucination rate.
As an affiliate, we earn on qualifying purchases.
Knowledge and hallucination: better, not best
AA-Omniscience rewards correct answers and penalizes confident wrong ones. Meta’s improvement is real, but the leaders are still well ahead:
- Claude Opus 5.5: 46
- GPT-6 Astra: 43
- Gemini 4 Argon: 42
- Claude Sonnet 5.5: 32
- Muse Wonder Full: 30
- Muse Wonder: 25
The hallucination rates tell the same story. Among models that score 45 or more on the Intelligence Index, Gemini 4 Argon hallucinates 15% of the time, Muse Wonder Full 23% and Muse Wonder 27%. Going from 33% to 23% in one release is good progress. For any workflow where a confident wrong answer costs money, it is still a reason to keep a verification step in place.
proprietary AI models for developers
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The price of the score
This is where the release turns from impressive to awkward.
Infographic 02 · The price of the score
Same score, different bill
Muse Wonder and GPT-6.1 Sol both score 52. Running the full Intelligence Index costs $4.52 per task on Muse Wonder and $0.72 on Sol. Gemini 4 Argon scores one point higher for $1.99.
Intelligence Index vs. cost to run it
Up and to the left is better: more intelligence for less money. Hover a dot for its values.
Not plotted: Muse Wonder Full and Inkling (no per-token price) and K2 Horizon 375B A23B (shown at $0.00).
Where Muse Wonder’s $4.52 goes
Per Intelligence Index task, at max reasoning effort
- Input
- $3.43
- Cache writes
- $1.92
- Cache hits
- $1.44
- Uncached input
- $0.07
- Output
- $1.08
Muse Wonder reads 7.69M input tokens per index task, 28% more than Spark 1.3, and writes about 54k output tokens.
List price per 1M tokens
Input / output, as reported by Artificial Analysis
| Model | Input | Output |
|---|---|---|
| Muse Wonder | $4.00 | $20.00 |
| Claude Opus 5.5 | $4.00 | $20.00 |
| Claude Sonnet 5.5 · Gemini 4 Argon · GPT-6.1 Sol | $2.00 | $10.00 |
| GPT-6 Astra · Claude Fable 5.1 | $10.00 | $50.00 |
| Muse Spark 1.3 (previous) | $1.25 | $4.25 |
| Muse Wonder Full | Muse Code only | |
Cached input on Muse Wonder costs $0.20 per 1M tokens, a 95% discount.
Show the data as a table (27 configurations)
| Model | Creator | Index | Cost / task |
|---|---|---|---|
| Claude Opus 5.5 (max) | Anthropic | 58 | $5.98 |
| Claude Sonnet 5.5 (max) | Anthropic | 56 | $5.46 |
| Muse Wonder Full (max) | Meta | 56 | — |
| Claude Fable 5.1 (max) | Anthropic | 53 | $7.63 |
| GPT-6 Astra (max) | OpenAI | 53 | $3.26 |
| Gemini 4 Argon (high) | 53 | $1.99 | |
| Muse Wonder (max) | Meta | 52 | $4.52 |
| GPT-6.1 Sol (max) | OpenAI | 52 | $0.72 |
| Muse Spark 1.3 (max) | Meta | 48 | $1.60 |
| Grok 4.7 (xhigh) | SpaceXAI | 46 | $3.74 |
| MiMo-V2.6-Pro | Xiaomi | 46 | $0.13 |
| Qwen3.8 Max (0902) | Alibaba | 45 | $5.41 |
| GLM-5.3 (max) | Z AI | 45 | $2.01 |
| Step 5 Preview | StepFun | 44 | $1.03 |
| Kimi K3 (max) | Kimi | 44 | $2.00 |
| Claude Haiku 5.5 (max) | Anthropic | 43 | $0.21 |
| GLM-5.3-Flash | Z AI | 42 | $0.28 |
| Gemini 3.8 Flash (high) | 41 | $1.24 | |
| DeepSeek V4.1 Flash (max) | DeepSeek | 39 | $0.27 |
| Mistral Large 4 Preview | Mistral | 38 | $1.13 |
| GPT-6 Luna (max) | OpenAI | 38 | $0.07 |
| Qwen3.8 27B (xhigh) | Alibaba | 34 | $1.01 |
| K2 Horizon 375B A23B | Inst. of Foundation Models | 31 | $0.00 |
| MiniMax-M3 | MiniMax | 29 | $0.51 |
| Inkling (xhigh) | Thinking Machines | 25 | — |
| Nemotron 3 Ultra | NVIDIA | 23 | $0.60 |
| Muse Glimmer (high) | Meta | 17 | $0.06 |
Source: Artificial Analysis, Muse Wonder analysis (12 Oct 2026) and model leaderboard, read 13 Oct 2026. Intelligence Index v4.3, ten evaluations. Cost per task = total API cost to run the Intelligence Index divided by its tasks.
Powered by Thorsten Meyer AI
Here is what it costs to run one Intelligence Index task, set against the score:
- GPT-6.1 Sol scores 52 for $0.72.
- Gemini 4 Argon scores 53 for $1.99.
- GPT-6 Astra scores 53 for $3.26.
- Muse Wonder scores 52 for $4.52.
- Claude Sonnet 5.5 scores 56 for $5.46.
- Claude Opus 5.5 scores 58 for $5.98.
Muse Wonder costs 6.3× as much as Sol for the same score, 2.3× as much as Argon for one point less, and 2.8× as much as its own predecessor, Muse Spark 1.3 ($1.60). For under a dollar more per task, Sonnet 5.5 scores four points higher.
The list prices explain part of it. Muse Spark 1.3 cost $1.25 per million input tokens and $4.25 per million output tokens. Muse Wonder costs $4 and $20, the same as Opus 5.5, while Sonnet 5.5, Gemini 4 Argon and GPT-6.1 Sol each charge $2 and $10.
The bill is in the input, not the output
The more useful finding is about tokens. Muse Wonder writes about 54,000 output tokens per index task. That is fewer than Muse Spark 1.3 (60k), Gemini 4 Argon (62k) or Claude Fable 5.1 (78k), and more than GPT-6.1 Sol (38k) or GPT-6 Astra (27k).
Input is a different matter. Muse Wonder reads 7.69 million input tokens per task. That is 28% more than Muse Spark 1.3 (6.03M) and two to seven times what Argon, Sol and Astra consume. Muse Wonder Full reads 6.44M.
Here is how Artificial Analysis breaks down the $4.52:
- Input: $3.43 (76%), made up of $1.92 in cache writes, $1.44 in cache hits and $0.07 in uncached input
- Output: $1.08 (24%)
Cached input gets a 95% discount ($0.20 instead of $4). At the conventional 7:2:1 blend of cached input, fresh input and output, that comes to $2.94 per million tokens. The discount helps, but at 7.69 million tokens per task, even cheap tokens add up.
For anyone running agents, this is the practical lesson. Muse Wonder’s cost depends on how much context your agent re-reads, not how much it writes. Long agent loops that keep reloading a large repository or document set are the workloads where this model gets expensive fastest. Before committing, measure your own input-token profile. A per-token price comparison won’t tell you this.
As an affiliate, we earn on qualifying purchases.
Speed
Artificial Analysis measures Muse Wonder at roughly 90–105 output tokens per second, with about 20 seconds to the first token at max reasoning effort. The model page shows 104.1 t/s and 19.73 s; the leaderboard shows 88 t/s and 19.27 s. That is fine for background agents and batch work. In an interactive chat, twenty seconds before anything appears is a noticeable wait.
Who should use which
If you already use Muse Code, Muse Wonder Full is the headline. It ties the best Claude Code configuration on the Coding Agents Index and leads two of the ten intelligence evaluations outright. You can’t price it per token, but you don’t have to.
If you buy intelligence through an API, Muse Wonder is hard to justify on these numbers. GPT-6.1 Sol gives you the same index score for about a sixth of the cost, Gemini 4 Argon gives you one point more for less than half, and Claude Sonnet 5.5 gives you four more points for about 20% more money. Muse Wonder needs a workload-specific reason that the index doesn’t show, and you should prove that reason with your own evaluation before switching.
If you are deciding what to run agents on, look at input tokens first. A model that reads two to seven times as much context as its rivals changes your cost curve more than its list price suggests.
None of this changes my own setup. In my October 2026 AI stack, Opus 5.5 builds and GPT-6.1 Sol digs, and this chart is a good picture of why: Opus at the top, Sol at the cheapest point of the 52 line. It also explains why Mistral Large 4 Preview, at 38 for $1.13 per task, still trails the frontier, and why I would not run agents on it, even though it is the best model from outside the US and China.
Verdict
Meta has done the hard part. Muse Wonder Full is a frontier model by any reasonable definition: second-highest score among the 27 configurations here, first on two evaluations, top three on six, and level with Claude on coding agents. Muse Spark 1.3 was not that.
Then Meta did something harder to defend. It put its best model behind Muse Code and priced the API model like Opus, even though that model scores like Sol. Unless that pricing changes, Meta’s frontier performance stays inside Muse Code, and API buyers get a 52-point model for more than they would pay elsewhere.
The capability gap is closed. The value gap is open, and Meta opened it.
Sources
- Artificial Analysis: Muse Wonder and Muse Wonder Full analysis (October 12, 2026)
- Artificial Analysis: Muse Wonder Full model page
- Artificial Analysis: Muse Wonder model page
- Artificial Analysis Intelligence Index v4.3 leaderboard, read October 13, 2026 (scores and cost per index task for all 27 configurations)
Columbus Day / Indigenous Peoples' Day Picks
long weekend sales
As an affiliate, we earn on qualifying purchases.
