Meta Muse Wonder vs Claude Opus 5.5 vs GPT-6 Astra: Benchmarks, Pricing, and Cost Per Task (October 2026)

Meta Superintelligence Labs introduced two new models on October 12, 2026, Muse Wonder and Muse Wonder Full, describing them as the company's most capable models to date. The split release is the part worth paying attention to before the benchmarks: the stronger model of the two is not generally available, and the one you can actually call from the API is the weaker one, by a meaningful margin on several of Meta's own evaluations.

What the two models are, and the gap between them

Both models build on the Muse Spark line and target agentic work specifically: Meta highlights expert domain coverage in law, finance, and health, frontier coding performance for tasks like building games or business applications, and two named capabilities called Active Memory, which carries relevant details across a conversation, and Active Focus, which keeps a session working toward a goal through failed attempts, tries alternate approaches, and asks for clarification before taking consequential actions rather than guessing. Both accept text, image, audio, and video input and produce text output, with a context window of roughly one million tokens, and the model ID for the available one is muse-wonder-1.0, built to be compatible with existing OpenAI and Anthropic development tool interfaces rather than requiring a bespoke integration.

On Meta's own benchmark table, run at max reasoning effort, Wonder Full leads the pack on Harvey LAB-AA, a legal benchmark, at 13.1 against Claude Opus 5.5's 4.2; on AutomationBench-AA, a business workflow benchmark, at 45.4 against Opus's 41.2; on DeepSWE 1.1, a software engineering benchmark, at 79.0 against Opus's 74.2; and on SWE-Marathon, a long running coding benchmark, at 57.5 against Opus's 46.3 and GPT-6 Astra's 42.5. Opus 5.5 pulls back ahead on the aggregate Artificial Analysis Intelligence Index, 58 against Wonder Full's 56, and leads on JobBench, Terminal-Bench 4.0 AA, and HealthBench Pro, where GPT-6 Astra also edges out Wonder Full. The standard Wonder model, the one actually available through the API, trails Wonder Full on every one of those benchmarks, scoring 52 on the aggregate index, a point behind GPT-6 Astra's 53.

Worth flagging directly: this table mixes third party and Meta run evaluations, some of the coding comparisons use different tooling across vendors, and none of it is independently controlled the way a single neutral benchmark harness would be. Treat the relative ordering as directionally informative, not as a precise ranking you'd stake a procurement decision on without your own evaluation against your actual workload.

Pricing and the cost per task number that actually matters

Muse Wonder's API pricing is $4 per million input tokens and $20 per million output tokens, with cached input at $0.20 per million and an extra $2.50 per thousand web search queries, at a flat rate regardless of how much of the context window a given request uses. That matches Claude Opus 5.5's per token rate exactly on the ratio, though Opus's absolute price is higher at $10 and $50. GPT-6 Astra and Claude Fable 5.1 are both listed at $10 and $50 in the same comparison. Per token price is a weak proxy for what a task actually costs, though, and the more useful figure is cost per average Artificial Analysis task: Wonder comes in at $4.52, Opus 5.5 at $5.98, and GPT-6 Astra at $3.26, meaning Astra is the cheapest per completed task of the three despite a similar headline per token rate to Opus. Muse Wonder Full has no announced pay as you go price at all. The only way to reach it is a Muse Code subscription, either the $50 per month Power Usage tier at 20 times the Everyday plan's usage allowance, or the $200 per month Maximum Usage tier at 80 times that allowance, both of which buy capacity inside Meta's own coding assistant rather than unrestricted developer API access.

There is also a data use distinction buried in the access tiers that matters for any team with a client confidentiality obligation: Wonder's API is available in Standard mode only, where Meta states submitted content is not used for training, while Muse Code subscriptions default to Contributor mode, which permits Meta to use submitted content to improve future models. Meta's zero data retention option is limited and, per reporting on the release, does not support server side conversation state, which has a direct implication for Active Memory, since that feature depends on retaining context across turns.

Why the stronger model isn't the one you can actually use

Meta has not published its reasoning for gating Wonder Full behind a subscription rather than an API, but the facts available point to a plausible read: Wonder Full's strongest results cluster in legal, business workflow, and long running coding tasks, which is exactly the profile of an internal agent Meta wants running inside its own coding assistant and subscription product rather than being resold at commodity API margins to every competitor building an agent platform on top of it. For a buyer evaluating these models against Claude Opus 5.5 for a client facing or internal agentic deployment, the practical choice right now is not actually between Wonder and Wonder Full, it is between the generally available Wonder, GPT-6 Astra, and Opus 5.5, with Wonder Full sitting as a capability signal of where Meta's agent roadmap is headed rather than something you can provision today.

When this applies to your stack

If you are evaluating frontier models for an agent deployment that needs software engineering or long running task capability specifically, and software engineering or long context agentic reliability is the deciding factor rather than aggregate intelligence score, Wonder's benchmark profile versus GPT-6 Astra's lower cost per task is a real tradeoff worth running your own evaluation against, not something to decide from a vendor comparison table alone. For AI integration and agent engineering clients choosing between Claude, OpenAI, and Meta's model families for a production deployment, we run exactly this kind of workload specific evaluation rather than defaulting to whichever model tops the aggregate leaderboard that week.

FAQ

Can I actually use Muse Wonder Full today? Only through a Muse Code subscription, at $50 or $200 per month depending on tier. There is no pay as you go API price for it, unlike the standard Muse Wonder model, which is available now through the Meta Model API.

Is Muse Wonder better than Claude Opus 5.5? Not on the aggregate Artificial Analysis Intelligence Index, where Opus 5.5 scores 58 against Wonder's 52. Wonder Full, the subscription only variant, scores 56 and leads on several specific benchmarks including legal and long running coding tasks, but it is not the model available through the API.

Which model is cheapest per completed task? Among the three with published task level cost figures, GPT-6 Astra at $3.26 per average Artificial Analysis task, versus Muse Wonder at $4.52 and Claude Opus 5.5 at $5.98. Per token list price does not predict this ordering on its own.

Does Meta use my data to train future models if I use Muse Wonder? Through the API in Standard mode, no, per Meta's stated policy. Through a Muse Code subscription, the default is Contributor mode, which does permit Meta to use submitted content for future model improvement, and the zero data retention option available does not fully support server side conversation state.