Frontier Large Models Lose Favor: Usage Share of Top Models on OpenRouter Drops from 55% to 14%

TechFlow Summary: Meta releases Muse Wonder, with companies claiming to stand at the "frontier," but actual usage is flowing toward cheaper models that are "good enough"—the share of usage among the top 10 high-quality models on OpenRouter has dropped from 55% to 14% over the past year, which is a direct blow to the narrative betting on top-tier model pricing power. Later, Bumble negotiates privatization at an approximate $1 billion valuation, as the discount exit of the 2021 IPO generation continues.

Hello everyone!

If today is Monday, it must be time for another AI model release. This time around, it’s Meta Platforms’ turn. It released Muse Wonder, claiming the model can "deliver frontier-level performance" on certain tasks. For this company, it’s a moment of vindication: its AI business had previously slumped, prompting a massive hiring spree over the past roughly 14 months. Now it claims to be "pushing the frontier."

This would normally be noteworthy, except that every company is making the exact same claim. Two weeks ago, Google released its latest model, Gemini 4 Argon, stating it "delivers frontier performance in complex workflows" across multiple disciplines (you wouldn’t want it to only deliver frontier performance in simple workflows, after all). In early September, OpenAI released GPT-6 Astra, claiming it "is at the cutting edge in computer operations, browsing, software engineering, cybersecurity, science, and professional work." Also in September, Anthropic launched Claude Fable 5.1 and Mythos 5.1, calling them "the world’s most advanced programming and knowledge work models."

We get it—AI executives want to hype every advancement, hoping it will attract more customers, and sometimes investors. But you have to wonder if anyone can keep up with this blizzard-like release cadence, let alone pay close attention to each one. As we’ve reported, businesses are increasingly opting for less advanced models to handle most of their compute needs to save costs, reserving the most advanced (and expensive) models for highly specific tasks.

In fact, we reported today that only 14% of AI model usage routed through OpenRouter in the third quarter went to the top 10 highest-quality models, down from 55% in Q1 of last year. This is because less advanced models are becoming cheaper, and they are already "good enough" for most tasks.

Consumers are the target audience for Meta’s Muse personal agent, and they certainly don’t need frontier-level AI models for every single task. However, Meta says its latest Muse Wonder model will "soon power experiences within Muse." The only impact of doing so may fall on the company’s costs. At some point, AI companies will realize that being at the frontier is a lonely place.

Bumble’s Stumble

The 2021 class of IPOs is gradually being pulled off public markets, often at steep discounts relative to their listing prices. The latest is dating app Bumble. The Wall Street Journal reported on Monday that the company is in talks to go private at a $1 billion valuation.

While this means the acquisition price exceeds Bumble’s recent stock trading levels—the stock surged 50%, rising from $1.26 to $3.80—it also confirms the company’s massive value shrinkage since going public. Bumble priced its IPO shares at $43 per share, with the stock nearly doubling in its first few days of trading, temporarily valuing the company at $9.5 billion. Unfortunately, in a crowded dating app market, its growth has slowed.

If the deal closes, Bumble will present a challenge to former Sequoia partner Ravi Gupta and former Meta Chief Revenue Officer John Hegeman. The two have launched a new firm actively hunting for acquisitions—starting with Bumble. At least they can pick it up at a reasonable price.

Other News

• According to reports from Bloomberg and others, Meta has banned ByteDance from advertising on its platforms in the U.S. and several other markets.

• According to Bloomberg, Bain Capital’s Bridge Data Centres is negotiating with banks to increase its loan facility for its Malaysia operations from $2.8 billion to approximately $6.6 billion. This financing will rank among the largest debt deals in Asia’s data center industry.

• Microsoft unveiled its latest generation in-house AI chip, Maia 300, on Monday, announcing plans to begin "multi-gigawatt-scale deployments" starting next year.

Join the official Coincamps community:

X: https://x.com/coincamps

Telegram: https://t.me/coin_camps