Muse Wonder storms the index. The graphs grade the hype

Muse Wonder in five graphs, before a single verdict#

Meta released Muse Wonder and Muse Wonder Full on 12 October 2026, two models aimed at coding, memory and longer-running agentic work. Both are available to Muse Code subscribers, and the standard Muse Wonder is also on the Meta Model API, while Meta has set no date for bringing it to every Muse user. Meta says the pair deliver frontier-level performance against GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1. The independent scoreboard partly agrees: at maximum reasoning effort, Muse Wonder Full posts 56 on the Artificial Analysis Intelligence Index, the second highest rating on the board, behind only Claude Opus 5.5.

Read the ladder top down and the story is a patient prowl, not a rout. Claude Opus 5.5 holds 58. Muse Wonder Full sits at 56, a good three points clear of GPT-6 Astra and Gemini 4 Argon, which leaves vendor and independent readings in rare agreement. The standard Muse Wonder trails just behind Astra and Gemini with no published figure, so treat its exact rung as unwritten. Below them the field packs tight, and the rest of the frontier trails off toward the mid forties. For readers arriving from opus 5.5 benchmarks coverage, the shape will feel familiar: one clear leader, one close hunter, then a thicket.

This board is the one to watch when the Muse Wonder Terminal-Bench run lands, because the terminal is where coding agents earn their keep. Vals AI grades long, realistic jobs in a sandboxed terminal, and even the leader fails about a third of the tasks. Mind the footnote that follows the numbers everywhere: five Anthropic rows include provider-side fallback, and counting those attempts as failures drops Opus 5.5 from 65.15% to 58.08%, behind Sonnet 5.5 and Astra. Same model, same week, a different referee, a different order. That wobble is precisely why this piece charts the referees separately instead of crowning a champion.

Verbosity is a billing event wearing a benchmark costume. Muse Wonder Full generated 140M output tokens across the index against a median of 88M, while Opus 5.5 produced 260M. Tokens are the token trail that leads straight to the invoice, so a model that thinks out loud at max effort can still lose the value race it wins on points. The practical rule is older than any of these models: set effort per task, and keep the sprint for the open field.

Do not skip past this narrow chart, because it carries the most useful humility in the release. On Anthropic's own table, Astra still takes business workflows at 41.4% against 40.0%, and it also keeps agentic scientific research at 64.6% against 58.7%. Meta's claim that Muse Wonder beats Astra on automation tests is therefore the claim to watch most closely when independent automation scores land. A narrow gap is a scent, not a verdict.

The price of admission: Muse Wonder rate card and routes#

The rate card is public, precise and quietly aggressive. The Meta Model API lists muse-wonder-1.0 at $4 input and $20 output per million tokens, with cached input at $0.20 and no long-context premium whatever the window holds. Contributor mode, the discounted tier that lets Meta train on your traffic, is not available for Muse Wonder. Those list prices match Claude Opus 5.5 to the dollar, sit well under GPT-6 Astra at $10 and $50, and tower over Muse Spark 1.3 at $1.25 and $4.25. Price the finished pull request, never the prompt.

The field reads like a steady scent getting stronger toward the top left. Muse Wonder Full parks beside Opus 5.5 at twenty dollars of output against an index in the high fifties, while Astra and Fable charge fifty for 53. But list price is a starting bid. Anthropic prices Opus 5.5 at about a fifth of Astra's cost per task on FrontierCode at default effort, and Artificial Analysis puts Opus at $5.98 per index task. The Muse Wonder per-task cost is unlisted, and until it prints, the cheaper token is a promise, not a saving. Our own Grok 4.7 scorecard made the same point from the other side: a cheap token is a starting bid, not a final bill.

The subscription ladder sorts itself neatly. Muse agent plans run $20 and $100 a month, with Muse Wonder reaching the agent for subscribers first. Muse Code plans run Everyday at $5, High at $15 and Power at $50, and Muse Wonder Full is reserved for Code plans at $50 or $200 a month. Note what is missing: Meta has named no per-token price for Wonder Full and no date for free-tier Muse users, so the top of this ladder is still private property.

Two quiet lines in the docs deserve a slow read. First, there is no long-context premium, so a full 1M window costs the same rate as an empty one, which matters for agents that carry whole codebases in memory. Second, web search grounding bills at $2.50 per thousand queries on top of tokens, the meter that runs whenever an agent checks the live web. Rate limits sit at 3,000 requests and 4M tokens a minute per team. None of this is the headline, and all of it is the bill.

The Artificial Analysis Intelligence Index, and what a composite is good for#

The Artificial Analysis Intelligence Index is a composite of ten evaluations, from AA-Briefcase and GDPval-AA through Terminal-Bench 4.0 and SciCode to Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience and AA-LCR. Artificial Analysis describes it as a measure across reasoning, coding, knowledge, instruction following, scientific reasoning and multi-step tasks. A composite is a weather vane, not a diagnosis: it tells you which way the frontier wind blows, never which model heals your particular workflow. It is also the first report card on Meta's rebuilt Superintelligence Labs after a disappointing Llama 4 launch and an aggressive hiring overhaul, and the frontier capability Meta claims for the Full model finally gives that rebuild a number.

What the index is good for is shortlisting. When ten harnesses agree that two models sit two points apart, that gap is sturdy enough to plan around, and the 56 next to Muse Wonder Full is measured the same way as the 58 next to Opus 5.5. What it is not good for is selecting. Effort dials move scores, verbosity moves bills, and no composite has ever written your team's pull requests. For readers tracking muse wonder benchmarks as the numbers settle, the index is the page to bookmark and the wrong page to worship.

Hold the apparent contradiction gently: beating Opus on most of Meta's internal benchmarks while trailing it on the independent composite is normal, because the two sentences describe two different exams. Vendor suites flatter vendor harnesses. The composite exists to stop any single harness from grading its own homework. When the independent per-test ai agent benchmarks land for Muse Wonder, this table gets rewritten. Until then it stays a table, not a chart.

Terminal-Bench 4.0, and what terminal tests are good for#

Terminal-Bench 4.0 measures how well an agent completes long, realistic pieces of work in a sandboxed terminal, across software, science, machine learning, operations, hardware, security and media. Version 4.0 is a fresh set of 66 community tasks, none carried over from 2.1. What terminal tests are good for is predicting agent stamina: planning, iterating, recovering from errors and coordinating tools, which mirrors the Active Focus behaviour Meta advertises. A model that scores here can be trusted with longer leashes. One that cannot should stay on shorter ones. Start any coding-agent evaluation with a measured task, the way our coding agent cost checks prescribe, and never with a leaderboard.

Anthropic's rows fill out the tribe. FrontierCode and CursorBench are its agentic coding measures, where Opus 5.5 posts 54.4% and 57.8% against 53.3% and no score for Astra, while Terminal-Bench-Science tests research workflows in the terminal, the one coding row Astra clearly keeps at 64.6% against 58.7%. Note the footnotes travelling with these figures: Opus 5.5 ran at xhigh effort against Astra at high, and the Astra figures are OpenAI's own reports rather than Anthropic reruns. Effort settings are the quiet quarry hiding inside every coding leaderboard.

Where Muse Wonder stands in the coding tribe#

Meta claims frontier coding performance and wins over Astra, Fable and Opus on software-building tests, but its per-test coding figures sit inside the research blog graphics, and no independent coding board lists Muse Wonder yet. That is a gap in the record, not a verdict on the model. The Vals board in section one is the field to beat: Opus 5.5 at 65.15%, Astra at 59.60%, Gemini 4 Argon at 57.58%. When the Muse Wonder harness run appears, compare harness with harness, or compare nothing at all.

GDPval and AutomationBench, and what work tests are good for#

GDPval tasks span 44 occupations and nine major industries and demand real deliverables: documents, spreadsheets, slide decks, diagrams and reports. On GDPval-AA, Opus 5.5 scores 1846 Elo against 1735 for Fable 5.1, 1708 for Opus 5 and 1542 for Astra. What knowledge-work tests are good for is predicting professional agent quality in law, finance and advisory firms, which is precisely the territory Meta claims for Muse Wonder. Early enterprise reads point the same way: CoCounsel reports better expert evaluations and internal benchmarks with gains in speed and token efficiency. Legal work is the quarry both labs are now circling.

AutomationBench grades business workflows end to end, and here the margins are mouse-quiet: Astra 41.4%, Opus 5.5 40.0%, Fable 5.1 31.4%. What workflow tests are good for is predicting ops automation, the unglamorous middle where agents earn their retainers. Humanity's Last Exam with tools tests multidisciplinary reasoning at the hard end, Opus 5.5 at 67.7% against 57.2% for Astra, while Chartography checks visual chart recognition at 89.0%. Each test is a specialist instrument. None of them plays the whole orchestra.

For agencies, the legal row matters twice. First, clients in regulated sectors buy on evidence, and a legal-document win is a door opener. Second, Meta's own example tasks, navigating an insurance claim, making sense of a lease, reading a client's finances, are paralegal-shaped. What legal benchmarks are good for is shortlisting models for document-heavy practices. Until the Muse Wonder legal scores print independently, treat the claim as a flag planted on a hill nobody has surveyed.

Memory, computer use and safety: what agent tests are good for#

Active Memory carries what matters across conversations and retrieves the detail relevant to the task. Active Focus keeps working toward the goal when tasks get hard, tries other approaches, recovers from errors, asks for clarification and checks in before anything important. Meta built both into the Muse Wonder pair for longer-running work, which Meta says builds on the groundwork laid with Muse Spark. What memory tests are good for is predicting drift: agents that forget the brief mid-task compound errors the way gossip compounds in a small town. No independent memory score exists for Muse Wonder yet, so the long-context retrieval boards are the ones to watch.

Computer use is the sibling test. OSWorld measures whether a model can operate real computer environments on its own, and Anthropic reports 81.8% partial credit for Opus 5.5. What computer-use tests are good for is predicting desktop agents, the file-moving, form-filling kind. Meta's version of that future runs through a dedicated Muse Secure VM that acts across connected applications, sending email and booking travel. The VM is the den wall: useful work happens inside it, and the permission boundary is the whole point.

On safety, Meta claims its strongest safeguards to date against CBRN and cyber misuse while preserving frontier cyber capability for defensive use, with a full report of additional evaluations still to come. Note the Opus precedent travelling alongside: safeguard fallbacks can quietly reroute work to older models mid-workflow, which Vals AI caught 22 times on one Terminal-Bench run. What safety tests are good for is auditing that rerouting, a question our interactive story on AI agent rumours puts at centre stage. Log the effective model on every call. It costs one line and saves one very awkward client call.

What an agency should do with Muse Wonder this week#

Start with the wallet, not the leaderboard. At $4 and $20 with cache at $0.20 and no Contributor discount, Muse Wonder prices exactly like Opus 5.5 and far under Astra, so any bake-off should run those three at matched efforts on ten real tasks from last month. Score the outputs blind, count tokens, minutes and edits, and price the finished work. Our AI consultancy runs exactly this comparison for clients, and the published rates say what it costs.

Spend the saving on evidence, not volume. Clearer, cheaper drafting tempts teams to publish more than they should, and Google's guidance treats scaled, valueless content as a spam breach. Put the budget into research, first-party data and expert review, which is the work our content marketing service and SEO and GEO practice are built around. Volume is the cheap prey. Evidence is the quarry worth the prowl. For the threat side of the same ledger, our Astra security notes track what frontier cyber capability costs builders.

Run your own ai model comparison before you switch#

Treat any ai model comparison you read, this one included, as a hypothesis about your own work. Pick ten real tasks from last month, run them on your current model and on Muse Wonder at matched efforts, and score the outputs blind. Count the tokens, the minutes and the edits. The launch lantern is lit. Measure it, then decide how far to follow its glow. More field notes are gathering in the folkfox newsroom.

Frequently asked questions#