4 Dollars In, 20 Dollars Out: Pricing Muse Wonder Against Sol and Opus 5.5

Pricing is usually the most concrete fact in an AI model launch, and Muse Wonder's is no exception: through the Meta Model API, Wonder costs $4 per million input tokens and $20 per million output tokens. Meta's own framing, echoed by TechCrunch's coverage of the release, places that roughly on par with GPT-6 Sol and Opus 5.5. That's worth unpacking, because the choice of exactly those two comparisons is doing more work than it looks like.

Why those two models, specifically

It helps to be precise about what each comparison model actually is, since both names sound similar to their respective flagships but describe meaningfully different products. GPT-6 Sol sits below GPT-6 Astra in OpenAI's lineup — a separate, cheaper model rather than a discounted version of the same model, announced at the same DevDay cycle that introduced ChatGPT's Dots feature. Opus 5.5 occupies the middle tier of Anthropic's current model family, below whatever Anthropic currently positions as its top-end option. Neither is a flagship. Both are the tier each lab has built specifically for cost-sensitive, high-volume use — exactly the profile of a personal agent that might be calling a model dozens of times across a single day's worth of tasks, rather than a handful of carefully-constructed, high-value queries.

Meta didn't price Wonder against the most expensive frontier offerings on the market; it priced Wonder against the cost-conscious tier each competitor offers for exactly that kind of everyday, high-volume usage. That's a deliberate positioning choice, not a coincidence of where the market happened to land — Meta is telling developers "this costs what you'd pay for a budget-tier model from either major lab, but it benchmarks at frontier tier," which is a genuinely competitive pitch if the benchmark claims hold up.

What the benchmark context adds

That pitch is more credible than usual because Wonder isn't just claiming to be cheap — Artificial Analysis's Intelligence Index, an independent measure, scores Wonder Full at 56 and the base Wonder model at 52, placing Wonder (Max) at #9 out of 214 models tracked. We cover what that external verification does and doesn't confirm in Muse Wonder and Muse Wonder Full: Meta's First Frontier Model. Combine that outside number with the $4/$20 pricing and Meta's argument becomes: competitive capability at a budget price, which is exactly what a company trying to win developer attention away from two entrenched competitors wants to be able to say.

What the input/output split tells you

The shape of the price itself is worth reading, not just the headline numbers. A 5x gap between input cost ($4/M tokens) and output cost ($20/M tokens) is a common pattern across the industry — output tokens are more expensive to generate because they require a full forward pass through the model for every single token produced, while input tokens can be processed in parallel as a batch. Meta's specific ratio here isn't unusual, but it does tell you something about how Wonder is likely to get expensive in practice: workloads that involve Wonder reading a lot of context — a long document, an extended conversation history pulled in via Active Memory — stay comparatively cheap, while workloads that ask Wonder to write a lot — long-form legal or financial analysis, extended code generation — cost proportionally more. For a model Meta is explicitly marketing on expert-domain reasoning in law, finance, and health, all of which tend to produce long, carefully-worded output rather than short answers, that output-heavy pricing is worth budgeting for rather than treating the headline "$4 in" as the number that matters most.

The asymmetry worth flagging

What this pricing comparison doesn't cover is Wonder Full. The more capable model — the one Meta says beats Fable, Opus, and Astra on its MemoryBench claim, covered in MemoryBench and Active Memory — has no published per-token API price at all, because it isn't sold through the API. It's bundled into a Muse Code subscription instead, a structural choice we examine in Why Wonder Full Skips the API. That means the $4/$20 figure, however competitive, describes only the lesser of the two models Meta announced. Anyone comparing "Wonder" to Sol or Opus 5.5 on cost is, by definition, comparing the weaker half of what Meta shipped.

A separate number worth not confusing with this one

This per-token API pricing also has nothing to do with what a Muse consumer pays. The Muse app's own subscription tiers run $20 or $100 a month, a flat consumer price point entirely separate from the $4/$20-per-million-tokens developer pricing covered here — two different products, two different customers, two different pricing logics, and conflating them produces a meaningless comparison. We look at the consumer pricing specifically, and what it implies about who Meta expects to pay for Muse directly versus through transaction fees, in 20 Dollars or 100 Dollars a Month.

Pricing a new model to match two well-known reference points isn't subtle, and it isn't meant to be. It's Meta telling developers exactly where to put Wonder on their existing mental map, which only works as a pitch if the benchmark numbers backing it up continue to hold under independent testing rather than just Meta's own model card.

The caveat that applies to every number in this article

None of the pricing figures here come with Meta's own promise of stability. xAI, for comparison, has repriced Grok Bot's access tiers multiple times over the course of 2026 as usage patterns and competitive pressure shifted, and there's no reason to assume Meta's API pricing is immune to the same kind of adjustment once Wonder has been live longer and Meta has real usage data to price against. A $4/$20 split announced at launch is a snapshot, not a contract. Anyone building a product around Wonder's API pricing today should treat this figure the way you'd treat any cloud-compute price list — accurate as of October 2026, and worth rechecking before it becomes load-bearing in a business plan.

Part 34 of 46 in the Meta Muse series. Previously: MemoryBench and Active Memory: What Muse Wonder Claims to Do Better. Next: No Open Weights: What Changed Since Meta's Llama Era. Browse more builder insights or explore AI skills at aiskill.market.