Muse Wonder vs Reflection Beam: One Is a Rate Card, the Other Is a Promise

Seven days apart, from opposite directions. M​eta put Muse Wonder into general availability on October 12, 2026, on the Meta Model API and inside Muse Code, at $4.00 per million input tokens and $20.00 per million output tokens, with a 1,048,576-token context window and no download of any kind. Reflection announced Reflection Beam on October 5, 2026 — a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active per token — with the weights promised under Apache 2.0, a waitlisted preview behind a sign-up wall, and no published price on any of the company's own properties.

Which means the honest framing of this matchup is not model against model. It is a rate card against a promise. One of these you can call this afternoon and will never own; the other you may eventually own and cannot call at all today. If you are picking a default for a workload that ships this quarter, that distinction settles more than any benchmark line does, and it is worth being explicit that neither side of it is a knock on the other lab.

Seven days, two directions of travel

Muse Wonder arrived as a finished commercial product with unusually thin public documentation underneath it. The model is callable, the price is listed, the parameter surface is enumerated — a reasoning-effort ladder running from minimal to max with a medium default, an output ceiling of 943,718 tokens, every reasoning token billed at the output rate — and a second checkpoint, Muse Wonder Full, ships in the same launch at a higher score but only inside a Muse Code subscription and not sold per token at all. What M​eta did not ship is anything downloadable. There is no model card release, no checkpoint, no licence grant.

Reflection did the reverse. Beam's launch post is a research announcement with a serving promise attached: benchmark tables, a pre-training account that names 6,144 GB300 NVL72 systems and a four-week run over 23.8 trillion tokens, a reinforcement-learning stage on 10,500 GB300s with more than 100 million rollouts and roughly 1.3 billion sandboxes, and a specific licence commitment — the weights, technical report, model card and developer artefacts will follow under Apache 2.0, in the company's own words, later in the month. What did not ship on October 5 is a price, a serving context window, a maximum output figure, or a file.

Two labs, one week, and between them the entire spectrum of what "released" can mean in 2026.

What each one has actually published

Every figure below is the vendor's own, and where a number does not exist, that is the finding.

• Callability — Muse Wonder is generally available since 2026-10-12 on the Meta Model API under muse-wonder-1.0 and inside Muse Code; Reflection Beam's preview is a waitlist for "a limited set of users," with no self-serve path in

• Price — Muse Wonder $4.00 in / $0.20 cached in / $20.00 out per million tokens, published; Reflection Beam not published anywhere on Reflection's own properties, so any dollar figure you have seen attached to it is an estimate rather than a list price

• Context — Muse Wonder 1,048,576 tokens; Reflection Beam's API documentation lists a 256,000-token serving limit with a footnote warning the figure may change during the beta, while the launch post separately says midtraining extends the model's effective context to 1M tokens. Those are a training-side capability and a serving-side limit, not a contradiction — but only one of them is a number you can build against

• Maximum output — Muse Wonder 943,718 tokens; Reflection Beam not published

• Modality — Muse Wonder takes text, image, speech, video and files, and returns text; Reflection Beam is text in, text out

• Weights — Muse Wonder none, and none announced; Reflection Beam Apache 2.0 promised, not released as of today

• Independent score — Muse Wonder 52 on Artificial Analysis' Intelligence Index, ranked #9 of 219; Reflection Beam has no row on that board, or on any other independent evaluator we can find

• Architecture — Muse Wonder undisclosed; Reflection Beam 501B total / 23B active, a stated ratio of about 4.6%

The asymmetry in that list is the whole story. Muse Wonder has a price and a target and no artefact. Reflection Beam has an artefact promised and neither a price nor a target.

The efficiency claim is the one to hold at arm's length

Reflection's most interesting claim is not a benchmark score. It is that Beam pairs coding and agentic capability with "highly efficient reasoning," achieving scores on advanced reasoning benchmarks comparable to GLM-5.2 while using 3–4× less inference compute, with the gap widening against models in the 2T+ parameter family such as Qwen 3.8-Max. The chart behind that claim is a vendor chart, on a vendor's own axes, and it is unreproduced — but the underlying arithmetic is not exotic. Inference FLOPs scale roughly as two times the active parameter count times the number of tokens generated, so a 23B-active model that generates comparable output for a task really does cost less per task than a 40B- or 100B-active one. The architecture is the argument, and on architecture Beam's case is coherent.

What is missing is the denominator. Compute per task is a ratio, and the published material gives us the parameter side of it and nothing on the token side. Reflection publishes no output-token count for any of its benchmark runs. There is no verbosity figure, no tokens-per-task measurement, no independent run.

Muse Wonder, by contrast, has exactly that number and it cuts the other way. Artificial Analysis measured Muse Wonder (max) at 130 million output tokens consumed during its Intelligence Index evaluation, against a median of 85M across comparable models, and at about $4.52 per index task. The evaluator's own label for that behaviour is "somewhat verbose." Muse Wonder is a frontier model at a frontier price and it pays for its score in output tokens, because reasoning tokens bill at the output rate — which is the single most consequential line in M​eta's own pricing documentation.

So the two labs are making opposite bets and neither has published the numbers that would settle it. M​eta has a measurable verbosity and no disclosed parameter count, so you cannot compute its FLOPs per token. Reflection has a disclosed parameter count and no measured verbosity, so you cannot compute its tokens per task. Anyone who tells you Beam is three to four times cheaper to run than Muse Wonder is comparing a parameter ratio against a token count that does not exist.

Sparsity buys inference cost, not capability

There is a second trap in the 501B figure, and it is easy to fall into because the number is so large. Total parameters in a sparse mixture-of-experts model are a storage and pre-training statistic, not a per-token compute statistic. Every token Beam generates touches roughly 23 billion of its 501 billion parameters. The other 478 billion sit idle on that call. Beam is not a 501-billion-parameter model in the sense that a dense model of that size would be, and quoting it next to Muse Wonder's undisclosed size invites a comparison that means nothing.

Where the ratio does matter is deployment. A 501B checkpoint in BF16 is on the order of a terabyte of weights, and the active-parameter count does not shrink the memory footprint — only the compute. Serving Beam yourself means either a large multi-GPU node or a heavy quantisation, and quantisation is exactly the kind of thing you can only evaluate once weights exist. That is the second reason the Apache 2.0 release is the event that matters here rather than a footnote: it converts a promising architecture into something you can measure on your own hardware, at your own batch size, with your own prompts.

M​eta's position on the same axis is the opposite. Muse Wonder's architecture is undisclosed and its deployment story is entirely someone else's API. You will not quantise it, will not run it on-premises for a regulated workload, and will not fine-tune it. You rent it, at a listed rate, indefinitely.

Where the vendor numbers and the independent ones disagree

This is worth a short section because it is the discipline that keeps the rest of the comparison honest. M​eta's own harness puts Muse Wonder Full at 59.1 on its Personal Superintelligence Index and Muse Wonder at 56. Artificial Analysis, running its own index on Muse Wonder (max), gets 52. Same model family, same week, four points of daylight — and the independent figure is the lower one. Muse Wonder Full at 56 on the independent board is a genuine result, ranking #4 of 699 models in that class, and it is still four points below what M​eta's own table implies. None of that makes M​eta's numbers wrong; vendor harnesses are tuned for the vendor's deployment, and M​eta labels them as its own. It does mean the only figures in this article that can be laid against each other are the Artificial Analysis ones, and those exist for exactly one of the two models.

Reflection's tables carry a different version of the same problem. Beam is listed at 80.9 on SWE-Bench Verified, 80.1 on Terminal-Bench v2.1 and 44.4 on DeepSWE v1.1, all vendor-reported, all on the same page as comparisons against Inkling and Nemotron 3 Ultra. Two of those three are scores Muse Wonder also has a vendor number for, and comparing two sets of vendor numbers produced by two different harnesses is not a comparison. It is a coincidence of column headings.

If you have to choose this month

The decision reduces to which missing thing you can tolerate.

• Shipping this quarter, and you need callability, a price you can forecast, a 1M-token window and a 943,718-token output budget — Muse Wonder is the only one of the two that exists as a service today, and the choice makes itself

• Shipping this quarter, and Muse Wonder's $20-per-million output rate against a 130M-token evaluation profile is outside your budget — the useful move is the effort ladder rather than a different model. Reasoning is mandatory on Muse Wonder, but the ladder runs from minimal to max and every reasoning token bills as output; running extraction and classification at minimal and reserving high or max for the calls that earn them is the lever that moves the invoice

• Regulated, air-gapped or on-premises requirement — neither model solves it today. Muse Wonder will never meet it. Beam might, in a month, under a permissive licence, at a memory cost you have not yet measured

• Research or fine-tuning plans — wait for the weights. Do not start a project against a 501B architecture you cannot download

• Benchmarking against your own workload — Muse Wonder can be measured this week and Beam cannot be measured at all. Every Beam number in your evaluation spreadsheet is currently a vendor claim with no denominator

What we route, and what we do not

Being plain about this: neither model is in OrcaRouter's catalogue. There is no meta/muse-wonder entry and no Reflection entry of any kind, and we will not present a route that does not exist. What we carry from Meta Superintelligence Labs is the Muse Spark line —Muse Spark 1.2 and Muse Spark 1.1 — which is the closest thing to M​eta's reasoning family a team can call through us today and costs roughly a fifth of what Muse Wonder does per token. Among the models this matchup is priced against, Claude Opus 5.5 is on our list at A​nthropic's own rates, and so is D​eepSeek V4 Pro at a fraction of that.

That honesty has a practical consequence that a comparison like this one makes unusually clear. Automatic failover only ever fails over to something we actually route, and a waitlist with no endpoint is not a fallback. If your plan for Beam is "we will test it when the preview clears," the useful thing to hold in place until then is a primary you can measure and a second route behind it — which is a configuration change on one O​penAI-compatible endpoint. OrcaRouter serves 200-plus models behind one key and one bill, so when Beam's weights land and some provider starts serving it, reaching it is an ID in a config file rather than a new integration.

What would change the answer, and when to look

Three events, in order of how much they would move this.

The first is the weight release. Reflection has committed to Apache 2.0 and said the artefacts are coming later this month. That release is what turns a 501B claim into something downloadable, quantisable and benchmarkable, and it also happens to be the only path by which Beam ever appears on a routing platform. Until it happens, Beam is a hypothesis with a citation.

The second is an independent evaluation. A single Artificial Analysis row for Beam would do more to settle this matchup than any amount of vendor table-reading, because it would supply the tokens-per-task figure that the efficiency claim is missing. Watch that board rather than the launch blog.

The third is Muse Wonder's own configuration surface settling. Its $4.52 per-task figure was measured at max reasoning on a model that has been available for one day. As teams exercise the lower rungs of the effort ladder in production, the realistic cost of the model will fall — and that number, not the index score, is what Beam would have to beat.

The likeliest near-term answer is neither a clean win nor a clean loss: a team that needs a model now runs Muse Wonder, and a team that needs to own its weights ends up with whichever of the two it can actually download. Which is a strange way to describe a head-to-head, and an accurate one.