There is a pattern we keep noticing whenever Meta ships an AI product into consumer surfaces: the first wave of coverage frames it as uncanny, the second wave frames it as inevitable, and neither wave grapples with the release notes. The Verge's Muse write-up is the current example, and the "creepy" framing has done what these framings usually do — it has crowded out the questions that actually matter for anyone building on top of the frontier model market. Consider the release stack around it: Claude 4.7 Opus landed on 2026-04-15 at 82.4% on SWE-bench Verified, GPT-5.5 followed on 2026-04-22 at $5 in and $25 out per million tokens. That is the context Muse arrived into.
The Pattern of Misreading Meta's AI Releases
The pattern is this. A Meta AI product ships. The launch hits three or four reviewers who trade in vibes. The reviewers pick the affective register — creepy, cringe, warm, weird — and the frame calcifies before a single technical question gets asked. Two weeks later a second wave of coverage arrives arguing that the affective frame was overblown, that the product is actually inevitable, that opposition to it is a kind of nostalgia. Neither wave reads the model card. Neither wave benchmarks the outputs against anything. Neither wave asks what corpus the thing was trained on or what the product roadmap implies about the corpus it is about to be trained on next.
We have watched this cycle now across image models, video models, avatar systems, character chat, and whatever taxonomy Muse ends up occupying. It happens because the trade press covering Meta is structured around access — the reviewers who get the demo are the ones who write the frame, and the demo is designed to produce a single strong impression, not to survive a benchmark run. It happens because the ecosystem of independent evaluators who cover Anthropic, OpenAI, Google DeepMind and xAI does not really extend to Meta's consumer AI, which lives in a different reviewer bucket. It happens because Meta's public communications lean on capability screenshots rather than the leaderboard-style disclosure that Anthropic and OpenAI have been forced into by their B2B customer base.
The consequence is that a Meta AI release ends up covered the way a new phone gets covered. Verdicts about the feel of the object. Almost nothing about the substrate. When The Verge writes that Muse is creepy, that assessment is not wrong — the object probably is creepy in exactly the way described. It is just that "creepy" is the least load-bearing thing that could be said about a product whose training pipeline, deployment surface, and pricing structure will shape how billions of casual users first encounter multimodal AI. The affective register is where the discussion ends. It should have been where the discussion started.
What The Verge's "Creepy" Frame Actually Buries
The frame buries the training-data question, first. Every consumer-scale generative product raises a specific corpus question: what did this thing see during pre-training, what did it see during the fine-tune, and what is the feedback loop by which the deployed users' interactions get fed back in. For a chat-style product built on Llama-family weights — Meta's flagship open-weight line, which posted 88.6% on MMLU and 89.0% on HumanEval at the 405B tier — the pre-training corpus is at least partially public. The fine-tune is not. The RLHF-equivalent loop is not. What Muse "knows" is a function of choices made in rooms that no journalist walked into, and the creepy-feeling output is downstream of those choices, not upstream. Treating the output as the story is treating the symptom as the etiology.
The frame buries the product-intent question, second. Meta is not shipping a chat surface because chat is the endpoint. Chat is the acquisition wedge. The endpoint is a persistent AI companion embedded across the family of consumer apps, which is a product category with distinct engagement, retention, and revenue math from the API market where Anthropic and OpenAI operate. A companion product optimizes for time-on-surface. A frontier API optimizes for pass@1 on SWE-bench Verified. These are not the same product even when they share a transformer backbone. Reviewing Muse without naming the category it belongs to is like reviewing a hedge fund's prime brokerage relationship as though it were a retail app.
There is a subtler burial here too. If you actually pick up two Meta primary documents — the model card for the underlying weights and the developer-facing description of the consumer product's guardrails — they say partly contradictory things. The model card describes capability envelopes measured on public benchmarks. The product description describes behaviors that the guardrail layer will suppress or amplify regardless of the underlying capability. Both are operative. The Muse a user encounters is the intersection: a capability envelope minus a suppression layer plus an amplification layer trained on unknown reward signals. The "creepy" frame collapses all three of those layers into one aesthetic verdict.
The creepy verdict is the least useful thing that can be said about a product whose corpus, guardrails, and reward loop each deserve their own investigation.
The Pricing and Benchmark Context Nobody Anchored It To
Muse did not ship into a vacuum. It shipped into a frontier market that had reset three times in the previous eight weeks, and understanding that reset is prerequisite to understanding what the release means. On 2026-03-10 Anthropic shipped Claude 4.6 Sonnet at $3 in and $15 out per million tokens, holding the price flat while lifting SWE-bench Verified to 77.5%. On 2026-04-15 Claude 4.7 Opus followed at $15 in and $75 out per million, taking the SWE-bench mark to 82.4% and HumanEval to 94.0%. On 2026-04-18 Google DeepMind put Gemini 3.1 Pro at $2.5 in and $10 out per million tokens with a 2M context window, and it posted 94.3% on GPQA Diamond — the highest disclosed score on that benchmark from any provider we track. On 2026-04-22 OpenAI shipped GPT-5.5 at $5 in and $25 out per million, at 80.8% SWE-bench and 89.6% GPQA.
Two things about that pricing table matter for reading Muse. First, the frontier is bifurcating into a "premium reasoning" tier ($15-$75 per million out for Opus) and a "workhorse multimodal" tier that has been aggressively repriced downward — Gemini 3.1 Pro at $10 out is a genuinely disruptive number for a 2M-context model that beats every published GPQA score in the leaderboard. Second, that repricing collapses the economic argument for a companion product built on rented frontier compute. If Meta were paying $25 per million output tokens to serve every user, Muse would not be economically viable at consumer scale, which means Meta is either serving from Llama-family weights it owns end-to-end or is running a much smaller distilled companion model whose behavior is closer to Gemini 3 Flash territory ($0.30 in, $1.20 out) than to Opus territory.
That is the anchor The Verge should have provided. The right first question about Muse is not "does it feel creepy" — it is "what is it running on, and at what unit economics." Everything downstream, from response quality to guardrail latency to how aggressively the reward loop can be tuned, flows from that answer. The reason the frame matters is not academic. A companion product served from owned weights at effectively zero marginal cost per token has a very different product roadmap than one throttled by inference bills. It can afford to be present, always-on, low-friction. It can afford, in the coldest sense, to be more familiar with the user than a rented-compute product would ever be. That is where the actual uncanny valley sits, and it is not visible from a single demo session.
The Tool Stack a Working Builder Would Actually Assemble Around This
Let us concede something upfront. If you are a working builder trying to reason about products in the shape of Muse — persistent multimodal AI companions running on consumer surfaces — you are not going to reason your way there by reading Verge coverage. You are going to reason by building on the same substrate the labs use and watching how your own product feels when it is deployed against real users. That means an opinionated tool stack, and the one we run is not exotic.
Primary coding surface: Claude Code, $20/month bundled with Claude Pro, running against Claude 4.7 Opus for anything that touches architecture and Claude 4.6 Sonnet for high-volume mechanical edits. The reason is not brand loyalty. It is that 82.4% on SWE-bench Verified is currently the highest disclosed pass@1 on that benchmark and the terminal-native workflow lets us keep the entire repo in context without switching UIs. Secondary coding surface: Codex, $20/month, running GPT-5.5. It shipped on 2026-04-22 as OpenAI's push for the coding-tool top slot and it earns its second chair by being genuinely better at some spec-writing and code-review tasks even when its raw SWE-bench (80.8%) trails Opus.
Third surface, and this is the underrated one: Gemini CLI, free, open-sourced by Google in April 2026 and pointed at Gemini 3.1 Pro. We use it as the long-context researcher. When we need to load a whole standards document, three model cards, and a spec into one prompt and ask analytical questions about them, the 2M context window plus 94.3% GPQA Diamond makes it the correct tool. We are not paying per token because the CLI tier is free at reasonable volume. That is what "workhorse multimodal at ten dollars per million out" means at the substrate layer: it changes what analysis is affordable to do. Fourth: Aider, free, as our multi-model harness when we want to A/B a task across three providers without switching IDEs.
What we rejected. We rejected Cursor at $20/month, not because it is bad — it is genuinely well-built — but because the IDE wrapper adds a layer of state between us and the model that we cannot introspect. When Composer misbehaves we cannot tell whether the issue is the underlying model or the orchestration. We rejected Windsurf at $15/month for the same reason plus the fact that it defaults to Claude 4.6 Sonnet where we would rather be explicit about the model choice per task. We rejected GitHub Copilot at $10/month because its agent mode still lives inside the IDE plugin abstraction and we have found the CLI-native tools give us cleaner failure modes to debug against. None of these are indictments. They are trade-offs that fall differently depending on whether you value speed-of-first-edit or clarity-of-model-behavior. We value the second, so we pay slightly more for the second.
Watch three or four signals to update your view on where the Muse category lands. First, whether Meta discloses which weights family Muse actually runs on — if it is Llama 4 or a distilled variant, that is confirmable from a proper model card and it is the single most predictive fact about the product's economics. Second, whether the next frontier repricing round pushes the workhorse tier below Gemini 3 Flash's $0.30-in/$1.20-out anchor, because that is the threshold at which rented compute becomes competitive with owned weights for consumer companion products. Third, whether independent evaluators outside the Verge/Wired/9to5 axis start publishing benchmark suites for consumer AI companions the way the SWE-bench maintainers publish for coding models. Fourth, whether the guardrail-behavior gap between the model card and the deployed product widens or narrows over the next two release cycles. Those are the numbers to watch. The affective verdicts will keep coming; they are not the update signal.