Muse Wonder vs GPT-6 vs Claude: Prices, Scores, Verdict

By ToolPilot Pro · Updated 2026-10-13 · Chatbots

#chatbots#muse wonder#meta muse wonder#muse wonder pricing#muse wonder vs gpt-6#muse wonder vs claude#muse wonder full#meta frontier ai model#muse wonder benchmarks#comparison

On Monday, October 12, 2026, Meta did something it had never done before: it released a frontier AI model. Not a research demo, not an open-weights community release in the Llama tradition — a closed, API-priced, benchmark-chasing frontier system called Muse Wonder, alongside a larger sibling called Muse Wonder Full. Mark Zuckerberg announced it personally, writing that Wonder "brings us to the frontier" a little over a year after Meta's Superintelligence Lab was formed under former Scale AI chief Alexandr Wang.

I spent the day of the launch reading Meta's announcement, the model card, the API docs, and the first independent numbers from Artificial Analysis, separating what is genuinely new from what is marketing gloss. Here is the honest version: what Wonder is, what it costs, how it actually scores against GPT-6 and Claude, and who should care.

What Actually Launched

Two models, two very different availability stories.

Muse Wonder is the one you can actually use. It is live on the Meta Model API at https://api.meta.ai/v1, on OpenRouter, and inside Muse Code. It speaks three API dialects out of the box — OpenAI Responses, OpenAI Chat Completions, and Anthropic Messages — which means existing code written for GPT or Claude can switch endpoints with minimal changes. That compatibility choice is deliberate: Meta wants migration to be a config change, not a rewrite.

Muse Wonder Full is the bigger model, and most of the impressive benchmark numbers belong to it. Right now you can only use it inside Muse Code on the Power Usage ($50/month) and Maximum Usage ($200/month) plans. There is no API for it yet, and Meta has not said when — or whether — one is coming.

In the Muse consumer app, Wonder is rolling out to paying subscribers first. Both models will also power the Muse agent going forward.

The specs that matter: roughly 1 million tokens of context (1,048,576), multimodal input (text, images, video, audio, PDF), text output, adjustable reasoning effort from minimal to max, and the full agent toolkit — tool calling, structured output, search grounding, and computer use. Rate limits are 3,000 requests and 4 million tokens per minute per team.

Active Memory and Active Focus, Explained

Meta is marketing two capabilities harder than anything else, and they are both about the same problem: today's AI assistants forget you between conversations and drift off-task during long ones.

Active Memory is Meta's name for carrying what the model knows about you across sessions — not just within one long chat, but over days, weeks, and months. "We put a tremendous amount of effort into making sure your model can remember what's important to you over days, weeks, months, and beyond," said research scientist Kavya Kopparapu in the launch video. In practice, this is the memory layer Meta has been building toward for its personal-assistant ambitions: the model builds and maintains a picture of your preferences, projects, and history instead of starting cold every time.

Active Focus is the other half: keeping the model working a hard goal through blockers instead of giving up or guessing. Meta's framing is that Wonder checks in and asks clarifying questions before consequential actions rather than charging ahead — which, given OpenAI just cancelled a model for doing the opposite, is a well-timed selling point.

The honest caveat: memory claims are the hardest thing in AI to verify from a launch post. Meta's own MemoryBench numbers put Wonder above Fable, Opus, and Astra — but the launch materials do not publish the raw per-benchmark numbers behind that claim, and it is company-supplied testing. Treat it as a direction, not a verdict, until independent testers have lived with the memory layer for a few weeks.

The Benchmarks, Honestly

Here is where Meta's own charts and the independent numbers diverge, and the divergence matters.

Meta's model card puts Wonder and Wonder Full in the same range as Fable 5.1 and GPT-6 Astra. Its own SWE-Marathon benchmark — a long-running software engineering test — scores Wonder Full at 57.5 against 46.3 for Claude Opus 5.5. That looks like a lead.

The independent read from Artificial Analysis is closer and less flattering. On the Intelligence Index: Muse Wonder Full 56, Muse Wonder 52, vs 58 for Claude Opus 5.5, 56 for Claude Sonnet 5.5, and 53 for GPT-6 Astra. Wonder Full lands top-three finishes on six of ten evaluations — genuinely strong — but two points behind Opus 5.5 overall, and the standard Wonder sits mid-pack among frontier models.

On specific tests: Terminal-Bench 4.0 jumped from 33% to 57% success for Wonder Full, and the Coding Agents Index hit 68, matching top Anthropic models. Coding is clearly the strong suit. Factual knowledge is the weak spot: just 30 points on the AA Omniscience test, lower than rivals.

And then there is cost per unit of intelligence. Artificial Analysis pegs Muse Wonder at $4.52 per benchmark task — nearly triple the $1.60 of Muse Spark 1.3 from six weeks ago. The frontier jump is real, but so is the price jump. Meta's "meaningfully below rivals" pricing line deserves the honest caveat: at $4/$20, Wonder costs roughly what GPT-6 Sol and Opus 5.5 cost. It is cheaper than Anthropic's top tier and pricier than GPT-6 Astra or Gemini 4 Argon per the independent analysis.

Bottom line on benchmarks: Wonder Full is a legitimate frontier model, roughly level with Sonnet 5.5 and a notch behind Opus 5.5. It is not the new king. It does not need to be — for Meta's first attempt, "in the conversation" is the story.

Why Meta Is Betting Everything on Memory

To understand why Meta leads with memory instead of raw benchmark scores, look at what Zuckerberg actually announced. His four-post thread on X framed Wonder as "the next step on our scaling ladder and a major leap toward personal superintelligence" — not the best coding model, not the cheapest API, but the model that knows you.

This is Meta playing to its structural advantage. Google owns search history. Microsoft owns the office. Meta owns the social graph — years of who you talk to, what you care about, how you spend your evenings. A model with genuine long-term memory, plugged into that graph, is a personal assistant no competitor can clone by training a bigger model. The bet: the winner of personal AI is not the smartest model but the one that remembers the most about you.

The training happened at Prometheus, Meta's gigawatt-scale data center in Ohio — the kind of infrastructure spend that tells you this is a decade-long bet, not a product launch. And the timing, one day after OpenAI's safety crisis dominated headlines, let Meta position Active Focus ("asks before consequential actions") as the responsible alternative to agents that freelance.

Whether the memory lives up to the marketing is still unproven — but the strategy is the clearest Meta has ever articulated. Watch whether the promised safety report appears, and whether independent testers confirm the memory claims. Those two things decide whether this is a platform shift or a press release.

What to Try First

If you want to evaluate Wonder hands-on this week, here is the highest-value sequence:

1. The memory stress test. Start a conversation, tell it five specific facts about an ongoing project, then open a new session tomorrow and ask what it remembers. Compare against ChatGPT's memory and Claude's. This is the single most important test because Active Memory is the entire pitch — if it holds up, Wonder earns a place in your stack.

2. The API swap. If you have code calling OpenAI or Anthropic, point it at https://api.meta.ai/v1 with muse-wonder-1.0 and run your existing evals. The compatible API formats make this a 15-minute experiment, and your own numbers beat any benchmark table.

3. The long agent run. Give it a genuinely multi-step task — research a topic across sources, build a small project, plan something complex — and watch whether Active Focus actually checks in before consequential steps or just claims to. This is where the OpenAI comparison gets interesting: the model that asks first is the one enterprises will buy.

Run all three before you believe the launch post. Frontier models are now launching monthly; your own tests are the only benchmark that matters.

What It Costs

The pricing is the real plot twist of this launch, because it ends an era.

API (Muse Wonder): $4.00 per million input tokens, $20.00 per million output tokens. That is a sharp rise from Muse Spark 1.3's $1.25/$4.25 — and it tells you Meta spent two years insisting it would stay open-source while Google and OpenAI charged for frontier access, then priced its frontier model like everyone else's. These numbers sit closer to GPT-4-class pricing than to anything Llama ever charged.

Muse Code: $50/month (Power) or $200/month (Maximum) — the only way to get Wonder Full right now.

Muse app: rolling out to paying subscribers first; reported consumer tiers at $20 and $100/month.

For comparison, our ChatGPT vs Claude vs Gemini breakdown covers where each lab's models win on price and capability — Wonder slots into the same premium tier as GPT-6 Sol and Opus 5.5, not the budget tier.

The End of the Open Era

This deserves its own section, because it is the biggest strategic signal in the launch.

Meta built its AI reputation on Llama: open weights, downloadable by anyone, the foundation of a thousand startups and research projects. Muse Wonder is closed. No weights, no downloads, no self-hosting — like Claude or GPT, not like Llama.

The company that spent two years arguing openness was the path to safe, broadly distributed AI just launched its most capable model as a closed API product. Whether that is a temporary posture (frontier capability first, openness later) or the new normal is the most important unanswered question about Meta's AI strategy. For developers who built on the assumption that Meta equals open, the assumption just broke.

There is also a transparency gap to note: Meta's safety report for these models has been promised but not yet published. A frontier launch without a published safety report, in the same month OpenAI cancelled a model over safety failures, is worth flagging. Watch for the report before trusting Wonder with anything consequential.

Wonder vs Wonder Full vs the Competition

Note: GPT-6 Astra as a product line is in a strange place — OpenAI scrapped the GPT-6.1 Astra release over failed safety tests, so "Astra" currently means the earlier GPT-6 Astra generation, not the cancelled successor.

How to Access It

Developers: the fastest path is the Meta Model API at https://api.meta.ai/v1 with model ID muse-wonder-1.0 (or meta/muse-wonder on OpenRouter). If your code already talks to OpenAI's or Anthropic's APIs, the compatible formats mean you can point your existing client at Meta's endpoint and change the model name — the intended migration path.

Muse Code users: Wonder is available now; Wonder Full requires the $50/month Power plan or $200/month Maximum plan.

Everyone else: the Muse app rollout starts with paying subscribers. Free-tier timing has not been announced.

Enterprises comparing agent platforms should also look at Google's Gemini agent, which takes the opposite approach — model-agnostic routing across Gemini and Claude with built-in governance — and at Claude's new Dashboards and Motion for the analytics side.

Honest Caveats

• No published safety report yet. Meta promised one; it is not out. In this market, that is a gap, not a footnote.
• MemoryBench is company-supplied. The "beats Fable/Opus/Astra on memory" claim has no published raw numbers. Independent verification is pending.
• Wonder Full is gated. The model behind the headline scores cannot be bought via API at any price right now.
• Weaker factual knowledge. 30 on AA Omniscience vs stronger rivals — do not use Wonder as your source of truth without verification.
• The price tripled in six weeks. $1.60 to $4.52 per benchmark task from Spark 1.3 to Wonder. Frontier capability, frontier pricing.
• Closed weights. If your stack depends on self-hosting or fine-tuning open models, Wonder is not for you — look at the open-weight releases like Mistral Large 4 or Aleph Alpha Kolibri instead.

Who Should Actually Switch

Try Wonder if: you are building agents and want to test whether Active Memory actually holds up over multi-day workflows; you want a third frontier option at Opus-class pricing to route between; or your code already uses OpenAI/Anthropic API formats and you can A/B test with a config change.

Wait if: you need the best raw scores (Opus 5.5 still leads independently); you need open weights; you need a published safety report before production use; or you are price-sensitive — GPT-6 Astra-class pricing and Gemini options undercut it.

Skip Wonder Full for now if: you are not already a Muse Code subscriber — there is no other way to get it, and the standard Wonder is the honest evaluation target.

The Bottom Line

Meta's first frontier model is genuinely good — not the best on the independent leaderboard, but firmly in the frontier conversation, with a real pricing and availability story and two interesting ideas in Active Memory and Active Focus. The bigger story is strategic: the open-weights era at Meta just ended, at least for the frontier, and the company that argued openness was the path forward now sells its best model the same way everyone else does.

For your purposes, the practical take is simple. Wonder is a legitimate third option at the premium tier, worth A/B testing if you already speak OpenAI or Anthropic API. Wonder Full's scores are not buyable at any API price yet. And until Meta publishes the safety report, keep it out of anything consequential. The frontier just got a fourth player — the leaderboard got more interesting, and so did the pricing pressure on everyone else.