The AI Model Tier List

August 24, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
August 24, 2026

HEADLINES

Hugging Face Weighs a Sale

Hugging Face is reportedly fielding acquisition offers worth at least $13 billion — nearly triple the $4.5 billion valuation from its 2023 round, which included Google, Amazon, Nvidia, Intel, and Salesforce. The platform has grown into critical infrastructure for the open-model economy, now hosting more than two million models, 1.5 million datasets, and 1.5 million AI apps; a buyer would be acquiring the distribution and discovery layer that helps developers find, vet, and ship open-weight releases. Speculation has centered on Nvidia as a natural acquirer, with one commentator suggesting an Nvidia-owned Hugging Face could "easily drop an open weight model that beats China before the end of the year." Both Stripe's purchase of OpenRouter and this new interest in Hugging Face look like bets on persistent model fragmentation as the industry's future.

NVIDIA's $6 Billion Poolside Deal

Nvidia's model-training ambitions kept surfacing this week: a $6 billion non-exclusive licensing deal plus a $1 billion equity investment gives it access to Poolside's technology and over 100 of its engineers, who'll work on Nvidia's Nemotron line, while Poolside's founders remain in place pursuing separate research. The deal — explicitly framed as neither an acquisition nor an acqui-hire — came together in a hurry after Poolside lost a 40,000-GPU cluster it couldn't finance in time. The stated goal is a frontier open coding model that can rival Chinese labs like DeepSeek and Moonshot. The same week brought reports that Nvidia is also backing fresh funding rounds for data-labeling startup Mercor and for Perplexity, underscoring how seriously the company is now investing across research, talent, training data, and the app layer.

NVIDIA Raises Prices as Much as 17%

Nvidia has begun notifying customers that prices for its top-end Grace Blackwell and Vera Rubin chips will rise by as much as 17%, even on orders already placed for delivery next year. A full 72-chip Vera Rubin rack is expected to hit $8 million, adding roughly $5 billion to the cost of building a gigawatt of compute. Bloomberg ties the increase to spiraling memory costs, and cloud providers are expected to pass the hikes on to their own customers rather than absorb them. It reads as further confirmation that the industry is positioning for a memory shortage that stretches well into next year.

Alibaba Raises $10 Billion

Alibaba raised $10 billion in a record-breaking secondary share sale on the Hong Kong exchange, the largest offering of its kind — though shares dropped as much as 10% on the news, their steepest intraday fall since last April. The move marks a departure from Alibaba's usually tight management of share supply, prompting Union Bancaire's Vey-Sern Ling to wonder whether the company needs more capital than expected or is racing to get ahead of rivals. Michael Burry weighed in too, calling Alibaba's low-cost LLM push "impressive as a disruptive force" while warning that share issuances mean its return on invested capital "will continue to fall."

Unitree's 460% IPO Pop

Unitree Robotics went public on the Shanghai Stock Exchange, raising $900 million at a $9 billion market cap — then surged more than 460% on its first day of trading, a debut Bloomberg Intelligence's Ian Ma called a signal of "strong appetite for China's embodied AI sector." It's now the fourth Chinese IPO this year to pop more than 400% on day one, a pattern regulatory guardrails on selling help produce mechanically, though the scale of the pops still marks a real difference from the more muted debuts typical in the US.

Dr. Dre and Jimmy Iovine Are Pro-AI

Dr. Dre and longtime producer Jimmy Iovine used a New York Times profile to make the case that AI is good for music, not a threat to it. "In the studio, when gifted people have A.I., they're going to make better records," Iovine said, while also acknowledging that AI companies have "the worst public relations in the history of the world." Dre compared AI skepticism to past resistance toward drum machines and synthesizers, arguing only "people who have trouble creating" see it as a threat — and revealed that producers like Timbaland are already using the tools quietly. "They're using it," he said, "they just don't want to admit it."

MAIN STORY

The AI Model Tier List

A viral weekend tier list from AI creator Theo — Fable 5 alone in S tier, GPT-5.6 Sol a notch below in A, and a crowded, jagged middle underneath — became the occasion for a bigger argument: the industry has moved past asking which model is best, into a period where model stacks, routers, and task-specific tradeoffs matter as much as frontier capability. That shift already has its own business model, from router companies like OpenRouter (just acquired by Stripe for $7 billion) to a widely misread chart on Fable 5's enterprise adoption, AT&T's push to shift AI queries onto open models, and gateway data showing open-weight tokens overtaking closed ones.

WHICH MODEL IS BEST?

Theo's Tier List
S tier: Fable 5, alone. Everything else is jagged.
The full ranking put GPT-5.6 Sol alone in A; Kimi K3, DeepSeek V4 Flash, and GPT-5.6 Luna in B; Grok 4.6 and MuSpark 1.2 in C; Opus 5, Sonnet 5, GPT-5.6 Terra, and Composer 2.5 in D; DeepSeek V4 Pro in F; and Gemini 3.7 Flash and Gemini 3.1 Pro in their own tier below F, reserved for Google alone. Theo's own caveat: a tier list undersells how many separate axes — cost, speed, token efficiency, task fit — now separate these models.

Fable 5 and GPT-5.6 Sol
The genius nobody wants to work with, but nobody wants to fire.
Fable is the only S tier model — "the model that writes code I want to merge," the one trusted to double-check other models' work — but it still trips over things, takes shortcuts, and loses track of tasks. Sol, ranked a full tier below, is described as the default anyway: "a slightly dumber robot that does exactly what you tell it," and the model that would be missed more if it disappeared.

The Ramp Data Fight
The 30-day retention policy the chart missed.
A widely shared FT chart showing Fable 5's enterprise adoption plateauing was read across social media as proof that businesses have decided the model isn't worth its price. Missing from that read: Fable's 30-day data retention policy — a holdover safeguard from its government-mandated relaunch — is enough on its own to rule the model out for security-conscious enterprises, and the underlying Ramp data comes from a cost-control product, skewing the sample toward price-sensitive users to begin with.

AT&T
Holding frontier spend flat, routing everything else to open models.
AT&T plans to grow open models' share of its roughly 100,000 employees' AI queries from 40% today to 60-70%, while keeping spend with OpenAI and Anthropic flat — reserving frontier models for advanced coding and leaning on Nemotron, Llama, and Gemma for simpler tasks like PR summaries. A model router has already cut its AI coding costs by 56%, with only a 2% hit to quality.

The Open/Closed Split
Open models are winning tokens. Frontier is still winning the economics.
Data from Vercel's AI gateway shows open-weight token share jumping from 28% to 62% over two months, even as OpenAI and Anthropic's combined growth accelerated over the same period — a split MIT's Christian Catalini frames as three separate markets: cheap open-weight generalists, closed frontier generalists, and an emerging middle tier of enterprise-tuned "SOTA specialists" pairing open weights with proprietary context, the bet behind products like Microsoft Foundry.