TL;DR
- Meta Superintelligence Labs released Muse Wonder, Muse Wonder Full, and Muse Spark Open — three models in four days, the biggest US lab debut since GPT-6.
- Muse Wonder matches Claude Opus 5.5 on independent benchmarks (AA Intelligence Index 52 vs 58) and on price ($4/$20 per MTok), but all coding leads are from Meta-run evaluations.
- Gemini 4 Argon went from gated to public at $2/$10 per MTok — that rate doubles on 1 December 2026.
Who should care: Teams deciding which frontier API to standardise on; self-hosters tracking the open-weight ceiling; anyone on Gemini 4 Argon’s gated program who has not yet budgeted for the December price change.
Verdict: Test Muse Wonder against your real tasks before committing — vendor benchmarks look good, independent scores are more cautious. Lock in Gemini 4 Argon at the intro rate before December.
The Week That Made Meta a Frontier Competitor
On 12 October, Meta Superintelligence Labs announced Muse Wonder and Muse Wonder Full — its first closed models aimed directly at Claude Opus 5.5 and GPT-6 Astra. Three days later it released Muse Spark Open, an 816-billion-parameter open-weight mixture-of-experts model on Hugging Face. One week, three models, and zero independent benchmark results at launch.
Gemini 4 Argon, available to Fairwind defenders since late September, also opened to general API access on 13 October with a clear two-tier pricing schedule. The introductory rate makes the next six weeks the cheapest window to evaluate it.
This Week’s Rankings
| Rank | Model | Vendor | What changed | Key number (source) | Verdict | Best for |
|---|---|---|---|---|---|---|
| 1 | Meta Muse Wonder | Meta Superintelligence Labs | First public frontier API model from Meta; launched 12 Oct at $4/$20 per MTok with 1M-token context | AA Intelligence Index 52 (Artificial Analysis, independent) | Test it | Long-horizon coding and agentic work — once independent evals confirm the vendor claims |
| 2 | Gemini 4 Argon | Moves from gated Fairwind program to general availability 13 Oct; $2/$10 intro rate doubles to $4/$20 on 1 Dec 2026 | AA Intelligence Index 52.6 v4.3.2 (Artificial Analysis, independent) | Use it | Multimodal synthesis, long-output tasks; budget-conscious teams before December | |
| 3 | Meta Muse Spark Open | Meta Superintelligence Labs | 816B total / 42B active open-weight MoE released 15 Oct on Hugging Face; hosted at $1.25/$4.25 per MTok on OpenRouter | MCP-Atlas (tool use) 89.2% (Meta-run); Terminal-Bench 2.1 76.4% (Meta-run; Kimi K3 85.0) | Test it | Enterprise self-hosters; agentic tool-use pipelines; teams that can provision 700 GB+ VRAM |
| 4 | Meta Muse Wonder Full | Meta Superintelligence Labs | Stronger variant; no public API — only via Muse Code ($50/$200/month plans); consumer Muse app US/Canada only | AA Intelligence Index 56 (Artificial Analysis, independent) | Watch | Muse Code subscribers doing long coding or legal-agent tasks; everyone else waits for API access |
Benchmark Caveats
All coding scores from Meta in this scorecard were produced by Meta at maximum reasoning effort. Artificial Analysis ran the Intelligence Index independently — those are the only cross-model numbers evaluated on a consistent methodology. Here is where the independent view diverges from the vendor view.
The AA Intelligence Index puts Opus 5.5 at 58, Wonder Full at 56, Wonder at 52, GPT-6 Astra at 53, and Gemini 4 Argon at 52.6 — a tight cluster near the frontier. Meta’s own comparison table is more favourable to Muse Wonder. The sharpest discrepancy is ProgramBench (coding from scratch): Meta scores Wonder at 50.0% and Opus 5.5 at 70.5% in the same table — a 20-point deficit that the DeepSWE headline does not show. Long-horizon completion is a different skill from starting from a blank file.
Meta’s stated methodology is “best-effort” for rival models, which is a lower bound on competition performance. Read the model card before trusting any row in its comparison table.
What Does It Cost to Review 200 Files?
A standard pre-merge code review across a 200-file PR runs roughly 200,000 input tokens and 40,000 output tokens. Here is what each model charges at current API rates, with the math shown:
| Model | Rate (input/output per MTok) | Input cost | Output cost | Total | Rate source |
|---|---|---|---|---|---|
| Gemini 4 Argon (now) | $2 / $10 | $0.40 | $0.40 | $0.80 | Google Gemini API pricing page; valid until 30 Nov 2026 |
| Muse Spark Open (hosted) | $1.25 / $4.25 | $0.25 | $0.17 | $0.42 | OpenRouter launch pricing, 15 Oct 2026 |
| Muse Wonder | $4 / $20 | $0.80 | $0.80 | $1.60 | Meta Model API pricing page |
| Claude Opus 5.5 | $4 / $20 | $0.80 | $0.80 | $1.60 | Anthropic API pricing page |
| Gemini 4 Argon (from 1 Dec) | $4 / $20 | $0.80 | $0.80 | $1.60 | Google Gemini API pricing page; rate effective 1 Dec 2026 |
| GPT-6 Astra | $10 / $50 | $2.00 | $2.00 | $4.00 | OpenAI API pricing page |
Muse Spark Open at $0.42 per review is the cheapest option on this list if you use a hosted API rather than self-hosting. Gemini 4 Argon at $0.80 is the cheapest frontier model right now, and that halves the cost of Muse Wonder or Opus 5.5. After 1 December, Argon costs the same as both. For a team running 500 code reviews per week, the difference between Argon at current rates and GPT-6 Astra is $400 per week — $20,800 per year.
For cost math on sub-cent token options, see our earlier comparison of GPT-6 Luna and Muse Spark 1.3.
For Swiss & EU teams
Muse Wonder’s consumer app (Muse) is available only in the United States and Canada; no European launch date has been announced, according to Notebookcheck. The Meta Model API’s EU availability is unspecified — Swiss and EU developers should confirm regional access and data-processing terms with Meta before building production pipelines on it.
Muse Spark Open is a different situation. The weights are public on Hugging Face and the Muse Model License permits use by most teams (the $20 million revenue clause applies only to companies offering it as a hosted service to others). EU organisations can self-host on SGLang, vLLM, or llama.cpp — on-premises, no US data routing — provided they have the hardware (at minimum 700 GB VRAM in NVFP4 quantization, or two M5 Ultra Mac Studios at roughly $22,000 combined). Hosted APIs from OpenRouter and Together AI are listed at launch, but those route through US infrastructure unless the provider offers EU-region endpoints.
Model Notes
1 · Meta Muse Wonder — Test it
Meta’s first frontier closed API model lands at $4 input / $20 output per million tokens — the same as Claude Opus 5.5. The independent Artificial Analysis Intelligence Index score is 52, versus 53 for GPT-6 Astra and 58 for Opus 5.5. The API model id is muse-wonder-1.0; context window is 1 million tokens. Meta’s own table shows leads on long-running coding tasks: DeepSWE 1.1 (78.8% vs Opus 5.5 74.2%) and SWE-Marathon (53.1% vs 46.3%). On ProgramBench — coding from a clean prompt — Wonder scores 50.0% while Opus 5.5 scores 70.5% in Meta’s own table. The model is better at completing tasks it starts iteratively than at generating solutions from scratch. Source: felloai.com; VentureBeat.
2 · Gemini 4 Argon — Use it
Argon moved from Google’s Fairwind defender program to full general availability on 13 October. Pricing is $2 per million input tokens and $10 per million output, with cached input 95% off, until 30 November 2026. On 1 December those rates double to $4/$20. The Artificial Analysis Intelligence Index score of 52.6 (v4.3.2, independent) is level with GPT-6 Astra. On AA-Omniscience, Argon records a 15% hallucination rate — the lowest Artificial Analysis has reported — but answers only 50% of those questions (Claude Fable 5.1 answers 67%). The model supports up to 1 million output tokens per response, which matters for document synthesis and large-output agentic tasks. Full pricing and benchmark breakdown: vortx.ch.
3 · Meta Muse Spark Open — Test it
The open-weight model has 816 billion total parameters and 42 billion active per forward pass (62 layers, 256 MoE experts, 8 active per token). BF16 weights are about 1.6 TB; NVFP4 quantization (experts only) is ~530 GB, requiring at minimum 700 GB VRAM in a data centre configuration or two M5 Ultra Mac Studios linked via Thunderbolt 5 with RDMA. Benchmarks are mostly vendor-run. On MCP-Atlas (tool use), it scores 89.2% — ahead of Qwen3.8 (84.5%), GLM 5.3 (84.2%), and Kimi K3 (82.3%), all per Meta’s model card. On Terminal-Bench 2.1 (long coding), it trails those same three Chinese open models by 6–9 points. The Artificial Analysis Intelligence Index for Muse Spark Open is 39 — well below the frontier group but ahead of smaller open models. Source: MarkTechPost; implicator.ai.
4 · Meta Muse Wonder Full — Watch
Wonder Full scores 56 on the Artificial Analysis Intelligence Index — two points above Wonder, two below Opus 5.5. Meta’s own DeepSWE 1.1 score is 79.0% (Wonder: 78.8%, Opus 5.5: 74.2%) and SWE-Marathon is 57.5% (Wonder: 53.1%, Opus 5.5: 46.3%). The practical problem: there is no pay-as-you-go API. Wonder Full is only accessible through Muse Code subscriptions at $50/month (Power) or $200/month (Maximum), with usage caps on both. Until Meta adds API access, production deployment at scale is not possible. Track it; test it if you subscribe to Muse Code. Source: VentureBeat.
What to Watch Next Week
Mistral’s preview of Large 4 “Le Chonk” promised open weights by end of October — no date confirmed. If the weights land next week, they will be the largest open-weight model from a European lab. Meta Watermelon and GPT-6.1 Astra are also expected, though GPT-6.1 Astra was reported delayed by Reuters due to safety reviews. Independent benchmark results for any of the three Muse models would significantly change the picture above. For how to run your own eval before switching, see our evaluation guide.
Further Reading
- Muse Wonder: Benchmarks, Price and the Catch — Clearest single-page summary of pricing and benchmark caveats, including the ProgramBench gap.
- Meta’s 816B Muse Spark Open Trails Chinese Rivals at Coding — Honest breakdown of where Muse Spark Open leads and where it does not.
- VentureBeat: Muse Wonder and Wonder Full achieve frontier performance — but only the weaker one is via API — Full pricing terms and benchmark table with context on methodology.
Your turn: Has your team already run Muse Wonder or Gemini 4 Argon on a real task this week — and how do the numbers hold up against the benchmarks? Reply to our newsletter or send us a note — we feature the best answers in next week’s Scorecard.
How this article was made: AI researched and wrote this article under the Adrian persona, using the sources linked above, and it was published automatically without a human edit. Editorial guidelines are set by Adi. Spotted an error? Tell us and we will correct it.