Science & Analytics
The Scientist
Muse Wonder's $4.52 per task is the price of a 93% cache hit rate.
At 7.69M input tokens per task, the bill is almost entirely input, and Meta charges $4 uncached against $0.20 cached, so the quoted price is really a measure of how stable the prompt prefix stays. An agent loop that churns its prefix down to 50% hits pays closer to $17, which is the figure to test against if your workload rewrites context between steps. Gemini 4 Argon's launch figure has a shelf life too, since its price doubles after Nov 30.
In Play
Agent gates that fail open
A parser default let an incident agent's page_team action clear its permit gate with every test green, and CSO reports the same fail-open pattern in agent frameworks and MCP servers. Like every headline figure in this edition, a confidence value holds only under a hidden condition, here that the parser gets exactly the input it expected. Fix your gates this week (deep dive below), and scan for exposed DCGM Exporter endpoints (top quick hit).
Ask ClarityMuse Wonder clears two independent harnesses
Wonder Full scores 56 and Wonder 52 against Opus 5.5's 58 on Artificial Analysis's Intelligence Index v4.3, and Vals AI ranks Full #5 and Wonder #10. Only Wonder has an API price. The deep dive gives the one test worth running.
Ask ClarityCost per task runs on caching and promos
Launch-day cost per task for Muse Wonder and Gemini 4 Argon depends on your cache hit rate and on promotional pricing that ends Nov 30. The cost deep dive shows how far each one moves your bill.
Ask ClarityHuman video as robot pretraining
Runway reports that robot policies pretrained on web video reached 16.1 cm placement error, against 16.0 cm for teleoperation video, per Not Boring. Dyna claims 99.4% success over 24 hours of napkin folding from a human-video prior of more than 1M hours, per Latent Space. Both companies ran these tests themselves and reported no variance. Runway's 0.95 sim-to-real rank correlation covers only 8 policies, none of which were optimized against the simulator.
Ask Clarity
Deep Dives
- ●
A missing key turned 0.90 for one action into permission for another
When model output gates an action, the parser's default decides whether it is safe, and test fixtures built from normal cases will never show you which way that default fails.
The bug was one fallback. If the selected choice was absent from Jev's probability map, decisions.py used the largest probability of any option as its confidence. With page_team missing and inspect_logs at 0.90, page_team scored 0.90 and cleared the 0.80…
2 action items
- ●
Muse Wonder's gains hold up outside Meta, but only the weaker variant has a price
Two outside harnesses back Meta's agentic gains, but the variant you can actually price is the weaker one, and none of the deltas will carry over to your stack.
Meta's own leaderboard row is the weakest evidence, because Techpresso notes it rests on competitor runs the model card calls 'best-effort'. Two outside evaluators point the same way. On Terminal-Bench, Artificial Analysis reports Wonder up 19 points over Spark 1.3…
1 action item
- ●
Your cost per task is set by your cache hit rate and a Nov 30 price change
Launch-day cost figures assume an agent loop that keeps its prompt prefix stable and pays promotional rates; you need to check both assumptions before you route traffic.
The $4.52 figure looks like a property of the model. It is mostly a property of Artificial Analysis's agent loop. At Meta's list rates ($4 input, $0.20 cached input and $20 output per million tokens), Wonder's ~54k output tokens cost…
3 action items
The edition continues
Take the signal into the room.
Sign up or log in to read all 3 deep dives in full, plus the final take.
Read the full editionContinue with LinkedIn