Every AI launch now comes with the same picture.

A big table. A column of numbers. The new model's column shaded blue, with its wins in bold. It looks like a report card, and it is designed to be screenshotted.

Today it was Meta's turn. The company released Muse Wonder, its first frontier model, with a 14-row chart comparing it against Claude Opus 5.5, Claude Fable 5.1 and GPT-6 Astra.

So instead of just telling you who "won," this issue reads that chart with you, row by row. By the end you'll know where Meta actually stands, and you'll have five questions you can use on every launch chart from now on, including the next one, which could land this week.

Two models from Meta Superintelligence Labs, the team Alexandr Wang has run since Meta rebuilt its AI effort last year.

Muse Wonder is the one you can use. It's on Meta's API at $4 per million input tokens and $20 output, with a 1M token context window.

Muse Wonder Full is the stronger version. It isn't on the API at all. You only get it inside Muse Code, Meta's coding agent, on the $50 and $200 monthly plans.

Both were trained at Meta's Prometheus data center in Ohio, and Meta is pitching two traits most launches don't lead with:

Active Memory: keeping what matters about you across conversations, over days and weeks, not just inside one chat.

Active Focus: sticking with a hard task, trying a different approach when it gets stuck, and checking in before important actions. Wang used a dinner reservation as the example: if the first attempt hits a wall, it tries another way instead of giving up.

Wonder will also power the Muse app, starting with paid subscribers. Mark Zuckerberg called it "a major leap towards delivering personal superintelligence for everyone."

Count the rows and here's the scoreboard:

Meta: 7 wins (6 for Wonder Full, 1 for Wonder)

Claude Opus 5.5: 6 wins

GPT-6 Astra: 1 win

Claude Fable 5.1: 0 wins

Where Meta's lead is real and big:

Where Claude still leads:

On first read, it looks like a split decision with Meta slightly ahead. Now let's read it properly.

This is the part worth saving. It takes about a minute, and it works on every launch.

1. Who wrote the test?

Two of Meta's seven wins come from tests Meta built itself: the "Personal Superintelligence Index" and MemoryBench. That doesn't make them wrong. It makes them the homework the student also graded. Count them, but count them lighter.

2. Did everyone run under the same rules?

Almost every number on this chart is at maximum effort, the most expensive setting, which isn't how most people will run it. The exception is the most eye-catching row: SWE-Marathon ran Wonder Full at "xhigh," with each rival inside its own agent setup. Meta's own model card says that row compares "complete systems, with model and harness effects intertwined." In plain English, it's not a like-for-like test. As one early reviewer put it, "Nobody has run Wonder at a normal effort level on a normal budget yet."

3. Which column can you actually buy?

The headline column is Wonder Full, and you can't call it through the API. The version developers can actually buy is the second column, plain Wonder, and it tops exactly one row: MemoryBench. Always read the column you'd actually be paying for.

4. Who's missing?

The chart compares against Opus 5.5, Fable 5.1 and Astra. It leaves out Claude Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon. That matters, because on the independent index below, Sonnet 5.5 ties Wonder Full, and Sonnet costs half as much as even regular Wonder on the API. And Artificial Analysis estimates GPT-6.1 Sol costs $0.72 to run each of its benchmark tasks, against $4.52 for Wonder.

5. What does the independent row say?

Look at the bottom of Meta's chart. The last row is the Artificial Analysis Intelligence Index, which Meta didn't run:

Credit where it's due: Meta put a row it doesn't win at the bottom of its own chart. Plenty of labs don't.

Within hours of the launch, a screenshot started circulating: a Minecraft-style game reportedly built by Muse Wonder on its "xhigh" setting.

The progression checklist in the corner is incredibly ambitious. Punch a tree, craft planks, mine diamonds, build a portal, all the way down to "enter THE NETHER!" The actual world is a few trees, some floating sand blocks and one mysterious pink cube.

Cue the word of the day: benchmaxxed, meaning a model that has been tuned to score well on tests more than to do well at real work.

There's a real point hiding in the joke. Building something from nothing is what ProgramBench measures, and it's Wonder Full's weakest row against Opus, 59.0 to 70.5. Meta's chart already told you this.

But apply the same standard both ways. One screenshot from one prompt isn't a benchmark either. We don't know the prompt, the settings or how many tries it took. The fair reading is that the screenshot matches what the chart says, not that it proves anything new.

When we judge a rumor or a claim, one of our three tests is simple: has this company done this before?

With Meta and benchmarks, the answer is yes. In April 2025, Llama 4 Maverick shot near the top of the popular LMArena leaderboard using a special "experimental chat version." When the leaderboard tested the version Meta actually released, it ranked 32nd.

So people are squinting for a reason. But this launch is different in ways that count:

There's a model card on day one, with methodology caveats written down instead of left out.

The independent numbers line up. Artificial Analysis's results match what Meta published.

Meta says it tested its own benchmaxxing, comparing tasks it trained for against tasks it didn't.

That's a company that learned something from Llama 4.

Meta is back at the frontier. On a neutral index, Wonder Full sits just behind Opus 5.5, level with Sonnet 5.5, and ahead of GPT-6 Astra and Gemini 4 Argon. A year and a half after Llama 4, that's a real comeback.

Where it shines: legal and finance agent work (that 13.1 vs 4.2 legal score is the standout), long-running coding projects, and memory.

Where it trails: building things from scratch by a wide margin, with smaller gaps on computer use and terminal work. And on price.

On price, Wang had said Wonder would cost "meaningfully below" leading OpenAI and Anthropic models. At $4 and $20, it's below GPT-6 Astra, but it's the same as Opus 5.5 and twice the price of Sonnet 5.5 and GPT-6.1 Sol.

A launch chart is a sales pitch with decimals. Every lab makes one, and that isn't a scandal. It's just the genre.

The useful skill isn't distrusting charts. It's reading them in order. The blue column is the pitch. The bottom row is the second opinion. And twenty of your own real tasks, run through the model you'd actually pay for, is the verdict.

There's a bigger picture too. For a while, the very top of AI was a three-lab race. Meta just made it four, and arrived at the same price as the leader. More competition at the frontier is good news for everyone who builds with these models, as long as you can tell which claims are worth believing.

The stories moving fast beyond today's headlines:

🟦 Meta Model API is where to try Muse Wonder yourself. Run your own prompts before you believe anyone's chart, including Meta's.

📏 Artificial Analysis is the independent index from the bottom row. Bookmark it and check it every time a new model launches.

🧪 promptfoo lets you build the eval that matters most: twenty of your own real tasks, run side by side across Wonder, Opus 5.5 and Sonnet 5.5.

Meta released Muse Wonder, its first frontier model, and its own chart shows it winning 7 of 14 rows against Opus 5.5, Fable 5.1 and GPT-6 Astra. The independent row at the bottom tells a more grounded story: Wonder Full lands just behind Opus 5.5, level with Sonnet 5.5, and ahead of Astra and Gemini 4 Argon. That's a real return to the frontier, just not the knockout the blue column suggests. The thing to keep is the habit, not the score: next time a launch chart crosses your feed, ask who wrote each test, whether everyone ran under the same rules, which column you can actually buy, who's missing, and what the independent row says. Then spend twenty minutes with your own tasks, because that's the only benchmark built for your work.

Stay building. 🤖