Claude Drives 26% of Anthropic R&D; Meta Open-Sources 816B Model; Gemini 4 Argon Returns

Meta Description

A week in AI: Claude guides 26% of Anthropic's R&D direction, Meta releases Muse Wonder and the 816B open-source model Muse Spark Open, Gemini 4 Argon returns to #1 on Arena, and Anthropic posts $16.8B in Q3 revenue.

Article Content

AI labs just wrapped another packed release week. Meta stormed back to the frontier with the Muse series, Google finally shipped Gemini 4 Argon, and Anthropic dropped four announcements plus one rumor in one go. But the fact that surprised AI engineers the most is this: Claude is now deeply involved in Anthropic's own R&D work.

Claude Starts Steering Anthropic's R&D Direction

One page in the State of AI Report 2026 is worth reading over and over: Claude currently guides about 26% of the research direction in Anthropic's core model development, up from less than 1% back in February 2026.

Nathan Benaich shared a more concrete experiment on a podcast: a research path forks, the model continues down one branch, a human takes the other, and humans judge the results. Older models won only about 10% of the time, while Mythos-level models won roughly 60%. He calls this capability "scientific taste."

Behind this is the accelerating RSI (recursive self-improvement) narrative—AI is now participating in building AI itself, and the share is climbing fast.

RSI (recursive self-improvement) is a core concept in discussions of AI safety and capability evolution. It refers to an AI system's ability to participate in improving its own architecture, training pipeline, or research direction, forming a positive feedback loop of "better capability → better improvement ability → further capability gains." Once such a loop starts, capability growth may far outpace linear extrapolation, which is why it is often seen as one of the key paths to superintelligence—and a critical threshold that the AI safety field watches closely. The data cited in the article is striking because it turns RSI from a thought experiment into an observable engineering reality: Claude's share of guiding research direction jumped from under 1% to about 26% within months, meaning Anthropic's model development is now substantially shaped by AI itself. Nathan Benaich's branching experiment provides evidence at the capability level—in human-judged controlled experiments, the new model's win rate for "choosing the right research direction" reached about 60%, far above the older model's 10%.

Meta Returns to the Frontier: Muse Wonder and an 816B Open-Source Model

According to analysis from Bilibili creators and YouTube channels, Meta's move this time amounts to a "two-front offensive."

Muse Wonder: The Closed-Source Model Ranked #2 on Arena

Muse Wonder ranks #2 on Arena Text with a score of 1513, just 12 points behind Gemini 4 Argon. API pricing is $4 input / $20 output per million tokens, essentially matching Opus 5.5. The stronger Muse Wonder Full is only available to Muse Code subscribers, at $50 or $200 per month.

Meta's marketing cleverness was also pointed out by Peter Gostev: everyone else calls it "Max," but Meta calls it "Full," and together that makes "Wonder Full."

But Peter also spotted a concern—ranking #2 on Arena Text shows people enjoy talking to it, but it only ranks #12 on Agent Arena. He described it as "pretty bad" when he used Muse Wonder to build a website. His suspicion: what Meta is promoting and what the API actually delivers may be different models.

He also identified a pattern: "The half-life of a benchmark in launch materials is about one release. Cite it once because it looks good, then never mention it again."

Muse Spark Open: 816B Open Weights

In the final minutes of the show, LDJ brought breaking news: Meta had just released Muse Spark Open. 816B total parameters, 42B activated, 1M context, multimodal input and text output. Meta calls it "the strongest open-weight model from the US to date."

For comparison, it is about 3x smaller than Kimi K3, and a fairer comparison is GLM 5.3 (753B, 40B activated). Running it requires roughly 700GB of VRAM (NVFP4 format), or as Zuckerberg put it, two new M5 Ultra Mac Studios.

There is one detail about coding capability worth noting: Terminal-Bench 2.1 score of 76.4, below GLM 5.3's 83.9 and Kimi K3's 85. It is stronger than other US open-source models, but has not yet reached the level of China's top open-source models.

The license requires displaying "built with Muse Spark Open" in products, with an annual revenue cap of $20M. Nisten's assessment: excellent for generating data, and "possibly one of the best computer use models."

Follow-up analysis found that Spark Open is closer to Spark 1.1/1.2 than 1.3 in coding, meaning it is not the latest Spark version with directly opened weights.

Total parameters and activated parameters are two key metrics for understanding the scale of MoE (mixture-of-experts) models. "Total parameters" refers to the full number of weights in the model, determining VRAM usage and deployment cost; "activated parameters" refers to the subset of weights actually involved in a single forward inference pass, directly tied to compute consumption and speed. Muse Spark Open's 816B/42B means that although it is enormous (requiring about 700GB of VRAM to run in NVFP4 format), each inference only activates expert paths totaling roughly 42B parameters, so per-call latency and compute overhead are far lower than a dense model of the same total parameter count. This is also the mainstream approach for today's large open-source models: using sparsity to balance "knowledge capacity" and "inference cost." Compared with GLM 5.3's 753B/40B and Kimi K3, these models are clearly in the same tier, and the competitive focus has shifted to training data quality, post-training, and performance on specific tasks rather than pure parameter scale.

Google Returns: Gemini 4 Argon

Peter noted that Google had not released a full Pro-level model for 8 to 10 months. Gemini 4 Argon is fully available across the Gemini API, AI Studio, Gemini Pro/Ultra apps, and Antigravity for Ultra, ranking #1 on Arena Text and also #1 in Roboflow's vision evaluation.

On pricing, it is $2 input / $10 output until November 30, then $4 / $20 afterward. Artificial Analysis's intelligence index shows Argon tied with GPT-6 Astra at 53 points, at about 60% of Astra's cost per task, while Opus 5.5 still leads at 58 points.

Google DeepMind's Geneviève Huskens showed off Desk Racer, a racing game she built with Argon based on her real desk. Her workflow is worth learning from: provide reference images, let the model "interrogate" you with planning questions, then set one big goal: generate sub-agents and let them run as long as needed. The key is an "adversarial sub-agent" that scores the output out of 10 and sends it back for rework if it is not high enough. For the 3D parts, first render a rough version, let Nano Banana "make it look real," then hand the image back to Argon and say "make it like this."

Geneviève called this their largest training run yet, and Argon, as an early checkpoint, is already helping train the next-generation model.

Anthropic's Big Week: Defense, Observability, Subscriptions, and Earnings

Distillation Defenses Cut Data Harvesting by 90-99%

Wolfram noticed a subtle timing: after calling for regulation of distillation, Anthropic took action itself. By locking down Claude's reasoning traces, classifiers, and tens of thousands of banned accounts, the data available to distillers dropped by 90-99%.

The side effect: users frequently encountered Claude disconnections requiring re-login. Nisten also found that if you mention "reasoning" while writing LLM code, Claude Code now interrupts directly.

Distillation refers to using the inputs and outputs of a strong model as training data to train a new model, allowing it to approximate the teacher model's capabilities at lower cost. It is an important path for today's open-source models to rapidly catch up to the frontier. The controversy around it lies in this: capabilities that model vendors spent enormous compute and data to train may be reproduced by third parties through large-scale API calls to "harvest" them, a behavior that often violates terms of service and can even trigger regulatory discussion. Anthropic's approach—locking down reasoning traces, deploying classifiers to detect harvesting behavior, and banning accounts in bulk—is a proactive defense that sharply reduces extractable high-quality signal by 90-99%. The side effect is that normal users are also affected, for example forced interruption of login sessions or disconnections triggered by specific keywords in code. This reflects the core tension in the distillation offense-defense game: the stricter the defense, the higher the chance of harming legitimate use, and balancing the two is an engineering problem vendors must continuously tune.

OpenTelemetry Tracing

Claude's Messages API, Managed Agents, and SDK can now send agent turns, tool calls, MCP calls, latency, and token data to any OpenTelemetry stack (public beta). Ramp said this reduced token usage by 10-20% for some agents stuck in loops.

Agent SDK Restores Subscription Usage

After multiple changes to the help page within a single day, it was finally confirmed: Agent SDK, claude -p, and third-party apps still use subscription quota, and Max and Team users now receive monthly API credits.

$16.8 Billion Quarterly Revenue

At a closed-door pre-IPO investor day, Anthropic disclosed Q3 revenue of about $16.8B, up about 45% quarter over quarter, profitable on an adjusted basis for the second consecutive quarter, and cash-flow positive. The IPO is reportedly in November.

Open-Source Ecosystem and AI Engineer NYC

On the show, NVIDIA released Nemotron 3.5 Super VL: 120B total parameters, 12B activated, supporting images, documents, and video, aimed at long-running agents. NVIDIA claims throughput up to 5x that of open-source models of the same scale.

OpenBMB's MiniCPM-V 4.7 can run a 35B visual MoE on an iPhone using about 3GB. Liquid released LFM2.5-24B-A2B, its largest model for local agents.

At AI Engineer NYC, Capital One's Max Kanat-Alexander offered a widely cited view: the importance of code quality is proportional to the longevity of software. AI will make "throwaway software" even more common—a calculator for yourself, a planning tool discarded after the party ends, as long as it works. But durable software needs more rigor than ever, because "the difficulty of understanding does not grow linearly, but compounds over time."

Nathan Benaich also pointed out a huge usage gap: about 90% of people at most use models for email, while Ramp data shows average monthly spending per person of about $12, versus about $7,500 for the top 1% of users—a 580x gap.

Background Notes

Arena rankings and the Artificial Analysis intelligence index are two different paradigms for measuring model capability, and readers need to understand their differences when interpreting the data. Arena (LMArena) is based on blind user votes and is essentially an aggregation of "human preference," so it better reflects conversational experience, fluency, and likability; it is authoritative on text tasks but easily influenced by stylized outputs and may not correlate with real engineering ability. The Artificial Analysis intelligence index, by contrast, is a composite score of standardized benchmarks, emphasizing reproducible task performance and reporting cost per task, making it convenient for cost-performance trade-offs. In the article, the gap between Muse Wonder's "#2 in text, #12 in Agent Arena" and Gemini 4 Argon's "tied at 53 on the intelligence index, costing only 60% of Astra" are typical examples of how these two systems complement each other: high preference rankings do not mean strong agentic task performance, and identical intelligence index scores can correspond to radically different cost structures.

One-line Summary

Claude now guides about 26% of Anthropic's R&D direction, Meta and Google released new models in the same week, and the RSI narrative is accelerating.

Paragraph Summary

AI labs released a dense wave of announcements this week: Claude now guides about 26% of research direction in Anthropic's core model development, up sharply from under 1% at the start of the year, clearly accelerating the RSI narrative. Meta attacked on two fronts, launching Muse Wonder, ranked #2 on Arena Text, and the open-weight model Muse Spark Open with 816B total parameters / 42B activated, though its coding ability still trails GLM 5.3 and Kimi K3. Gemini 4 Argon became fully available and topped the Arena text leaderboard, with discounted pricing through November 30. Anthropic's distillation defenses cut data harvesting by 90-99%, while Q3 revenue reached $16.8B, up 45% quarter over quarter and profitable. In the open-source ecosystem, NVIDIA, OpenBMB, and Liquid all made new moves, and AI Engineer NYC's topics shifted toward how to manage hundreds of agents.