Does Muse Wonder write a better CV than Muse Spark?

What changed when Meta launched Muse Wonder?

Meta released Muse Wonder on 12 October 2026, its first frontier-class model and a clear step up from Muse Spark. Where Spark was built to be small and fast, the Wonder family is bigger and aimed at complex, multi-step work — coding, long documents, agent tasks that run for a while. There are two versions. Muse Wonder is the one you can actually reach: it's served through the Meta Model API and rolling out to paid Muse app subscribers. Muse Wonder Full, which carries most of the headline benchmark numbers, is locked to Muse Code subscription plans with no API at all. Neither ships downloadable weights, which breaks sharply with the Llama era Meta spent years building its reputation on.

The numbers are real, if less dramatic than the launch posts suggest. On one widely tracked intelligence index, Muse Wonder Full scores 56 and Muse Wonder 52, against 48 for Muse Spark 1.3 — Claude Opus 5.5 sits at 58. The sharpest gain is agentic: Terminal-Bench 4.0 performance jumped from 33% to 57%. Meta also leans hard on two trained-in behaviours, Active Memory and Active Focus. Active Memory carries details about you between conversations and resurfaces them when they're relevant; Active Focus keeps the model chasing a goal instead of quitting at the first blocker. On Meta's own MemoryBench, both Wonder versions land around 85.5, ahead of Opus 5.5 at 81.5 and Fable 5.1 at 79.

Here's the catch for job seekers: none of those are writing benchmarks, let alone CV benchmarks. A model can top a legal-reasoning leaderboard and still produce a two-page CV that a parser chokes on. And access is the other wrinkle — the free Muse tier still runs Spark, so the model everyone is talking about isn't the one most people will type into this month. If you've been following what Meta actually opened up with Muse, the pattern is familiar: loud launch, paywalled capability, real but narrow gains. The useful question isn't which model wins a chart. It's which one writes bullets a recruiter believes.

Muse Wonder costs $4 per million input tokens and $20 per million output, with cached input at $0.20 — the same rate card as Opus 5.5, despite Meta briefing that it would come in meaningfully cheaper than OpenAI and Anthropic. The context window is roughly one million tokens, and the API takes text, images, video, audio and PDFs. For CV work that last detail matters more than the benchmarks: you can hand it your actual PDF rather than retyping your history into a chat box. Reasoning effort is adjustable from minimal to max, which is the lever most people ignore and the one that changes CV output quality most.

Does Muse Wonder actually write a better CV than Muse Spark?

Yes — but the gain shows up in structure, not sentences. Muse Wonder is better at the one CV task that genuinely needs reasoning: reading a 900-word job ad alongside your full history and deciding which four bullets belong under each role. Spark tends to rewrite line by line, so you get prettier versions of the same badly ordered CV. Wonder reorders. Ask it to map your last three jobs against a specific job description and it will reliably spot that your "stakeholder reporting" bullet is the one the ad cares about and your "general administration" bullet is dead weight. That editorial judgement is where the extra model capacity earns its keep.

The million-token window changes the workflow more than the prose. You can paste a long job ad, the company's careers page copy, your old CV and two previous versions in one go, and Wonder will hold all of it. Spark starts dropping threads on inputs that size — you'll see it reuse a title from the wrong role, or attribute a 2019 project to your current employer. Wonder's MemoryBench result is essentially a test of this: facts buried in 64,000 to 100,000 tokens, with distractors and mid-conversation contradictions thrown in. It's long-context recall within one session, not a diary of your career, and that distinction matters when you're pasting sensitive salary history.

Sentence-level writing is roughly a wash, and that's the honest verdict. Both models reach for the same tired constructions — "spearheaded," "leveraged cross-functional synergies," "demonstrated a proven track record." Wonder uses marginally fewer of them and holds a requested tone more consistently across a full page, but neither sounds like you unless you force it to. The fix is unglamorous: give it three of your own real bullets as a style sample and tell it to match the register, no superlatives. Models follow examples far better than adjectives. If you want a cleaner starting point altogether, an AI CV builder that keeps the formatting ATS-safe removes the formatting fight from the equation.

Where does Muse Wonder still invent experience on your CV?

Muse Wonder still fabricates, and it fabricates more convincingly than Spark did. The pattern is consistent: ask it to tailor your CV to a job ad, and it will quietly import the ad's requirements into your history. The posting wants Kubernetes, so a line about container orchestration appears under a role where you touched Docker once. It wants "managed a team of five," so your mentoring of two juniors becomes line management. Bigger models are better at making this sound plausible, which makes it more dangerous, not less. Spark's inventions were clumsy enough to spot. Wonder's read like something you might genuinely have written and forgotten.

Numbers are the worst offender. Tell it to "add metrics" and it will — "reduced processing time by 40%," "grew pipeline by $2.3M," "improved retention by 18 points." Every one of those is a guess dressed as a fact. Recruiters ask about them in interviews, reference checks surface them, and in regulated sectors a fabricated figure on a CV is a dismissable offence. My rule is blunt: if you can't name the dashboard, report or invoice the number came from, delete it. Replace it with scope you can defend — team size, budget, number of accounts, ticket volume, the systems you owned. Verifiable scope beats invented percentages with every hiring manager who has been doing this a while.

Language is the second blind spot. Meta's own model card concedes the model hasn't been evaluated across all the languages it was pre-trained on, and that's not a footnote if you're applying in Israel with a Hebrew or bilingual CV. Expect the Hebrew to be grammatically decent and idiomatically off — calqued American phrasing, English-order sentences, job titles nobody in Israeli hi-tech actually uses. It also has no concept of right-to-left file structure, so whatever it writes still has to survive export. Read Hebrew output as a native speaker before it goes near AllJobs or Drushim, and never let a model choose your Hebrew job titles for you.

Is Muse Wonder worth paying for just for a job search?

For most job seekers, no — not on its own. Muse Wonder sits behind a paid Muse tier at $20 or $100 a month in the app, while the free tier still runs Spark, and Wonder Full is restricted to Muse Code plans aimed at developers. Paying twenty dollars a month for marginally better bullet restructuring is poor value when you're spending maybe six focused hours on CV work during a whole search. The exception is if you're already subscribed for other reasons, or you're a developer with API access, in which case the one-million-token window genuinely speeds up tailoring across a dozen applications.

There's a sharper way to spend that money. The expensive part of a job search isn't drafting — it's the silence after you apply, and that silence usually traces back to formatting, missing keywords or unquantified impact rather than prose quality. A general-purpose model can't tell you your two-column layout is destroying your employment dates in extraction, because it never sees the parsed version of your file. It sees the text you pasted. That gap is why pasting a CV into any frontier model and calling it done is the most common unforced error I see.

Compared with its peers, Wonder is competent rather than distinctive for writing work. If you've looked at how GPT-6 Astra handles job-search tasks or what to check in anything Mistral Large 4 drafts, you'll recognise the shape: all of them restructure well, all of them invent numbers, all of them write formatting a parser dislikes unless you specify otherwise. Wonder's real differentiators — agentic persistence, cross-session memory, terminal work — are built for automating workflows, not for writing one page about your career. Pick your model on availability and cost, then spend your effort on verification.

How should you check a Muse Wonder CV rewrite before you send it?

Run three checks in order: a fact audit, a parse check, and a scoring pass against ATS criteria. The fact audit comes first because everything downstream is wasted if the content is fiction. Read every bullet Wonder produced and ask one question of each — did I do this, and can I prove it? Strike anything that fails. Then look specifically for requirements that migrated from the job ad into your experience; that's the failure mode bigger models are best at hiding. Expect to cut two or three lines from a tailored rewrite. If you cut nothing, you probably didn't read closely enough.

The parse check is the step almost everyone skips. Save the file, upload it somewhere that reads it the way screening software does, and look at what comes back out. Columns collapse. Icons vanish. Text boxes disappear entirely, taking your skills section with them. Dates in headers get stripped. A model that writes beautiful bullets into a template that breaks extraction has made your situation worse, not better. HRLens scores the file you actually send — a score out of 100, five category scores covering experience, tech stack, impact and ownership, clarity and structure, and ATS compatibility, plus a visual layout analysis, free.

Then close the loop with targeting. Paste the job description into the analysis — paste the text rather than a link, because LinkedIn and Indeed block fetching — and you'll get the missing skills named, plus ATS-friendly rewritten versions in six templates that mirror the ad's keywords. That's the piece no chat model gives you: an independent read of the document rather than a continuation of the conversation that produced it. Worth checking too whether the draft reads as machine-written at all, since recruiters have grown fast at spotting it — a quick check on whether your CV sounds AI-generated takes a minute and saves an awkward interview.

Frequently asked questions