Muse Wonder vs Claude Opus 5.5 vs GPT-6 Astra: Price, Context, Availability, and Coding Scores in One Table (With Every Number Labeled by Who Ran It)

Price and specs

Who can actually use each one

Coding scores, grouped by who ran them

This is the part the vendor cards flatten. Terminal-Bench 4.0 is the one benchmark every runner has touched, so it's the spine; the rest are shown where at least two of the three models were run by the same party. Rule for the table: a row is one runner, one harness. Vendor numbers are their own rows and never mixed with independent ones.

How to read it. On the independent runs, Opus 5.5 is first on every Terminal-Bench harness and on Vibe Code Bench, by one to seven points. Astra and Wonder Full trade second place: Artificial Analysis has Astra ahead by two, Vals has Astra ahead by one and a half, Vals' migration test has Astra ahead and its multimodal test has Wonder Full ahead by four. Wonder, the model you can rent, is a clear step behind — seven to ten points under its Full sibling on terminals. The one independent result where Meta leads outright is the Coding Agents Index, where Muse Code with Wonder Full ties Claude Code with Sonnet 5.5 at 68 and beats Claude Code with Opus at 66 — a result about the agent harness as much as the model, and one Astra wasn't in. Why Opus has four Terminal-Bench numbers: Anthropic's 66.4 is its own harness at xhigh 'with production safeguards on'; tbench.ai's 64.9 is Claude Code; Vals' 65.2 drops to 53.5 if you count safeguard fallbacks as failures; Artificial Analysis's 59.6 is the generic mini-swe-agent harness every model gets. Meta's card copied the AA number — so when Meta's table shows Opus at 59.6 and Wonder Full at 57.1, it's comparing on neutral ground, which is to Meta's credit, and also why Anthropic's own figure looks seven points better. The same spread exists for Astra (57.9 OpenAI, 58.2 Codex, 59.1–59.6 AA).

Speed, cost per task, and what they don't know

Artificial Analysis's framing of the Meta models is the fairest sentence written about them so far: 'Knowledge remains a relative weakness for both models, but this is mitigated by very low hallucination rates.' Wonder Full knows less than Opus or Astra (30 vs 46 and 44 on Omniscience) and makes things up less often (23 percent) — it abstains. For an agent that's a reasonable trade; for a research assistant it isn't. Wonder's speed is real: at max effort it starts answering in 19 seconds where Astra takes six minutes and Opus twelve, and it does so at a lower cost per task than Opus.

The skeptics, with the numbers

• 'Benchmaxxed.' The most-shared independent take (@bridgemindai, 335,000 views): 'Meta says it beats the competition on the key benchmarks. On the BridgeBench lava lamp test it built a red cylinder. No glass. No lava.' One task, one tester, and a fair illustration of the gap between a scorecard and a demo.
• 'Nowhere near the frontier.' Economist Kevin Bryan: 'Decent model but, like, everyone has private benchmarks at this point and it is obvious.' The independent numbers say 'near,' not 'nowhere near' — 56 vs 58 on the composite — but the point about private benchmarks lands on Meta's internal rows.
• 'Publishing intelligence without cost.' Charlie Marsh of Astral, on Wonder Full having no per-token price: 'I hope it doesn't become a trend.' It's the only frontier-class model you can't buy by the token.
• Methodology. VentureBeat: 'Meta's methodology deserves scrutiny.' Meta's own card concedes the mixed harnesses; its DeepSWE row cites results that aren't on the public board; its promised safety report is still 'soon.'
• A positive one, for balance: Vals, which runs its own harness, put Wonder Full fifth on its overall index — 'Meta just surpassed every OpenAI model on the Vals Index' — and Kingy.ai's small API test had Wonder ahead of GPT-6.1 Sol at medium effort while noting it 'does not establish a general coding winner.'

Which one, for what

• Best coding model you can actually run today: Opus 5.5. First on every independent terminal harness, $4/$20 flat, 1M context, in Claude Code, Cursor, Copilot, and three clouds, on a $20 plan. Its costs are time and tokens — the slowest to first output and the hungriest at max effort.
• Astra, if you're already in Codex or ChatGPT Pro. Within two points of Opus on most independent runs, OpenAI's computer-use flagship, and the only one of the three with a Critical cyber designation — but 2.5× the price, a surcharge past 272K, and absent from Cursor.
• Wonder Full, only if Muse Code is your tool. A genuine frontier-class model with the best agent-harness score in the set, available solely through a $50 or $200 subscription to a CLI that was weak a month ago on Spark. Worth a test month if you're a Meta Model API shop; not worth switching stacks for yet.
• Wonder, the sleeper for agents. Same price as Opus, three times faster to first token, multimodal input including video and audio, flat 1M context — and a tier below on code. If your agent loop is speed- and cost-bound and the tasks aren't hard engineering, it's the one to benchmark; if they are, Opus.
• For the Muse app, nothing changes today. Free users stay on Spark; subscribers get Wonder when it arrives; our Muse review scores what's in the app.

Our read

The table says Meta reached the frontier and did not take it. Wonder Full is the real thing — within two points of Opus 5.5 on the composite, a step behind Astra or a step ahead depending on whose harness, and the fastest on its own agent benchmark — but it's locked in a subscription CLI with no API, and the model Meta will actually sell by the token is a tier below on code. The honest comparison for most readers is the simpler one: Wonder versus Opus 5.5 at the same $4/$20, where Opus wins on code and knowledge and Wonder wins on speed and modalities. What Meta has done in fourteen months is remarkable; what it has shipped to a developer today is a very fast, well-priced second-tier coder and a frontier model you have to rent a terminal to meet. Our Muse Wonder explainer covers the launch, and the code editors ranking now lists Muse Code as also-considered pending a Wonder Full test.