This Year's Model
The $11,000,000,000,000 in capex allocated between 2024 and 2029 has one objective: the training and inferencing of large language models. Training costs huge amounts of money, while inferencing prints it in even greater amounts. For now. Noops Research has seen one clear trend emerge: the model is the only layer of the 'sand to software' stack that's deflating, while everything under it inflates. That'll be a big topic at our AI Investor event.
Closed
The frontier is a crowded band of closed-source models, per Artificial Analysis, led by Anthropic's Claude Opus 5.5, Meta's Muse Wonder Full, and OpenAI's GPT-6 Astra.
Google DeepMind was at the frontier, and has announced its return with Gemini 4 Argon, but hasn't yet rolled it out. Problems with rollout of Gemini 3.5 Pro Google to release a series of point-revision rollouts of their Flash models, which score decently well on intelligence and look to beat the frontier on cost-competitiveness. Can Google reach frontier status again? Does it matter for Google's business?
There are many other labs training closed source models; the best of them have only ever nipped at the heels of the frontier labs. Instead, they've triggered a price war over 'good enough' AI.
'Tokenmaxxing' in April led to cost minimising by June, as the bills rolled in. Fable and Astra, at the frontier, can get for frontier prices for their tokens. But not from everyone, and certainly not all the time: Anthropic's Fable was only getting 11.4% of customer spend and 6% of token volume in August, and had been overtaken by Opus 5 in enterprise spend.
As of mid-October, OpenAI GPT-6.1 Sol offers the best overall pricing among the near-frontier closed models; behind them are the DeepSeek and MiMo (Xiaomi) open models, offering near-frontier capabilities at less than 1/3rd the price per task of GPT-6.1 Sol.
Open(ish)
Facebook AI Resarch (FAIR) pioneered the open source large language model with the release of Llama in February 2023, progressing rapidly through Llama 3 in 2024, then stumbling badly with Llama 4, withdrawing and organising into Meta Superintelligence Labs. MSL has released one open model since then, Muse Glimmer, and promised to release the weights for its near-frontier Muse Spark 1.3, though that promise has yet to be fulfilled.
A raft of Chinese labs followed suit, led by Alibaba's Qwen, DeepSeek, Moonshot (Kimi), Z.ai (GLM). Each of these labs defines "open" somewhat differently, with Moonshot most closely resembling a middle-of-the road offer: large commercial users pay commercial rates while researchers and individuals (who can afford the kit to run them) can use them freely.
What's most interesting is the published research that has flowed out of these labs; fundamental innovations in attention, in context, and in efficiency. DeepSeek r1, the first open-source reasoning model, performed nearly as well as the frontier GPT-o1 model. That established a precedent: the frontier models would establish a new standard for performance that would be available in the open source models within six months. (Mozilla AI's State of open source AI report recently measured that gap at just 4.4 months.)
There are several notable open-source models from the United States, led by @Thinking Machines' Inkling Big, Poolside 's Laguna S2.1, and Reflection AI's Beam. None of them have as much raw capacity as the open models from the Chinese labs, but we expect that gap to narrow.
Small
Meta's Llama conclusively proved that small models could be powerful, though there remained a deep bias toward 'bigger is better' within the field. The intelligence of a model was held to be a direct correlate of the number of parameters within that model.
That belief was conclusively disproven on the 14th of August, 2026, when Alibaba's Qwen released Qwen3.8-27B. A "dense" model with just 27 billion parameters - approximately one-hundredth the size of its flagship model, Qwen3.8 Max - it performed at the rough equivalent of the very best frontier model (Opus 4.6) of just six months earlier.
Size, it would appear, isn't everything.
The next generation of models - possibly built atop Weibo's "VibeThinker", a tiny but very powerful reasoning model - will be able to run comfortably on a top-of-the-range smartphone. Inferencing will happen everywhere: in a pocket, on a desktop, and in the data centre.
Noops Research asserts that the demand for inferencing will grow by a thousand-fold over the next decade, as agents become common and then necessary tools to operate successfully in both commercial and private spheres. On device inferencing will increase the aggregate demand for inferencing.
USA versus China
Much of the race toward the frontier has now been characterised as a geopolitical battle between the USA and China - and, sotto voce, between capitalism and communism. This framing is both incorrect and unhelpful. It frames normal technological competition in strategic terms, defining a winner-take-all scenario as an inevitable outcome.
Since neither side would be prepared to let the other have unrestricted (and presumably sole) access to superintelligence (whatever that is, and should it ever arrive), neither side is willing to engage the other constructively around issues that confront all operators at this level of the stack. The fact that we even have to use the language of 'sides' - as though this were some front in an undeclared war for the future - illustrates exactly how far this narrative has penetrated into the consciousness of politicians, the public, and the labs themselves.
Noops Research contends that openness, transparency and collegiality represent the optimal path forward for all parties.
"Too Good" to be good?
In April 2026, Mark Pesce wrote a paper titled "How much Does a Token Cost?", predicting costs of 'alignment' would rise in lock-step with model capability. In September, that prediction became a reality, as OpenAI and Anthropic made a call to 'pace the frontier', which can be read both as a way to establish a legal cartel on machine intelligence, and as technique to provide the necessary breathing room to develop better alignment and containment techniques for tokens that have suddenly become "too good".
Where the cost of policing a model exceeds its economic value an economic dead-end has been reached. In this scenario - which seems to be the shape of the present moment - "good enough" tokens provided by commodity suppliers, requests routed through OpenRouter, take up the vast majority of token volume, with the frontier labs providing a stream of 'luxury goods' purchased only in cases where their benefits outweigh their costs. That OpenAI pulled back their public release of GPT-6.1 Astra because of alignment concerns only reinforces the sense that beyond a certain point, intelligence brings as many burdens as benefits.
The implication is clear: frontier labs have worked very hard to create something they might not be able to sell - and will be priced accordingly.
Is this a business?
Do the labs have an enduring business model? Deflation goes hand in hand with intense competition: arguably the most important business metric, 'cost per task', has been dropping continuously across 2026.
Opus 4.8 cost $1.80 per task in May, GPT-5.6 Sol cost $1.04 in June, Kimi K3 $0.94 in July, with Grok 4.6 at $0.84 in August. followed by DeepSeek V4.1 Flash at $0.27 in September, and Claude Haiku 5.5 at $0.21 in October. That's nearly a 9x decrease in cost for equivalent an outcome over the space of 3 months.
Given this price deflation toward very-low-cost tasks, can a closed-source lab offer enough value, relative to their open source and commodity peers, to make multi-trillion dollar valuations viable?