ARK Invest
ARK Investtoday
Economics

What Is The Best AI Model In 2026? | The Brainstorm 140

28 min video5 key momentsWatch original
TL;DR

OpenAI's GPT-4.5 Turbo cuts costs in half versus Fable while matching or beating performance, but frontier models may be solving the wrong problem — most knowledge work doesn't actually need top-tier intelligence.

Key Insights

1

Half the cost, same performanceGPT-4.5 Turbo delivers equivalent intelligence to Fable at less than half the cost, signaling OpenAI's decisive edge in the efficiency race.

2

Chinese models leading usageChinese open-source models (Tencent, DeepSeek, Minimax) now dominate OpenRouter's weekly usage leaderboard, yet frontier labs still capture most enterprise revenue.

3

Agents are basically zero capabilityAI agents still fail basic 30-minute tasks with 70% consistency rates, meaning automated knowledge work remains sci-fi — frontier intelligence buys you nothing if agents can't actually execute independently.

4

All-you-can-eat biasAnthropic's enterprise customers paying per-token face budget pressure to switch models, while Claude Pro subscribers with unlimited plans think Fable is the best in the world because they never see the cost.

5

Free weights aren't free computeThe open-source cost advantage vanishes once you factor in cloud margins (50%), infrastructure, and team overhead — running 'free' models still costs nearly as much as paying OpenAI or Anthropic.

6

Anthropic 10x growth testAnthropic hitting $60–70B annualized run rate growing 10x yearly suggests frontier labs are winning despite open-source noise, but if cloud providers truly gain share, it signals the open-source bet is finally paying off.

Want this for every new video ARK Invest posts? Brevyd summarizes each upload automatically, the morning it drops.

Start free

Deep Dive

The Efficiency Squeeze: OpenAI Catches and Passes Anthropic

Brett opens by framing the week's announcements: GPT-4.5 Turbo costs half as much as GPT-5.5 and delivers the same performance as Anthropic's Fable at less than half Fable's price. Frank adds crucial context — we're in a month-and-a-half release blitz from all five top US AI companies. Google launched Gemini 3.5 Flash, Meta followed with a better model at lower cost, Grok 4.5 arrived competitive but cheaper, and Anthropic released Claude Co-work for mobile and web. The pattern is unmistakable: efficiency is the new arms race. Frank notes Anthropic hasn't disclosed whether their higher pricing reflects model inefficiency or pure margin capture — likely both. Beyond raw models, products are evolving fast. OpenAI shipped ChatGPT Work with Codex agents, Anthropic shipped Claude Co-work, and XAI iterated on Grock Build. The new ChatGPT voice capability is full-duplex (can talk and listen simultaneously) and delegates thinking to frontier models behind the scenes, eliminating the dumb or fabricated responses that plagued earlier versions.

The Open-Source Insurgency: Chinese Models Dominate Usage, Not Revenue

Nick pivots the conversation by asking the right question: not which model is best, but which one is 'good enough' and why it matters. He cites Satya Nadella and Alex Karp's recent warnings about the 'reverse information paradox' — every query you send to a frontier lab gives it proprietary data to strengthen future versions. His bomb: OpenRouter's weekly usage leaderboard is now dominated by Chinese models (Tencent at number one, DeepSeek at three, Minimax at four) until you hit number seven with Anthropic. Yet the revenue story contradicts this. Anthropic is running $60–70B annualized revenue at 10x growth, while open-source companies sit at hundreds of millions to low billions. Brett counters that OpenRouter users are self-selected switchers who actively route between APIs — they're not representative of the broader market. More importantly, token volume is a misleading metric. Burning cheap tokens on open models tells you nothing about whether frontier model economics are under pressure.

The Agentic Illusion: Frontier Intelligence Is Not Frontier Capability

Brett and Nick converge on the same uncomfortable truth: AI agents are still broken. Brett says he's skeptical of claims that agents can offload 12 hours of work at 70% success — because 30% failure rate kills you in real work. Current agents can assist and buttress, but can't actually execute independently. The median knowledge worker felt no material bump from recent model releases because they've already figured out how to use AI to complete their day-to-day tasks. Frank pushes back slightly, noting that for the top 1% of workers, frontier models unlock outsized gains. But he agrees the average knowledge worker is plateauing. Nick adds a sharper point: work done by average-ish frontier models instantly commoditizes in competitive landscapes. If you're a trader using pure AI output, you'll be destroyed by someone using AI to augment human judgment. There's a character difference between reading AI slop repeatedly versus reading work by a human augmented by AI — the seams are visible now, but won't be forever.

The Revenue Test: Who's Actually Winning?

Brett reframes the question away from benchmarks toward business reality. Anthropic is at $60–70B run rate growing 10x, OpenAI is in the low tens of billions. Open-source companies are hundreds of millions to low billions. Where the market spends money reveals the answer. Users consistently upgrade to smarter models over cheaper ones at equal cost. But Frank notes that most companies aren't technically skilled enough to run open-source models themselves — only the elite few. Everyone else pays cloud provider margins (roughly 50%), which nearly erases the cost advantage. Infrastructure, teams, and product maintenance aren't free just because model weights are open. So most customers still pay OpenAI or Anthropic. The true test: if open-source is winning, Google Cloud should be growing faster than Anthropic. Google Cloud is at 63% YoY growth on an $80B run rate. If that holds, Anthropic could exceed all of Google Cloud by year end — a sign frontier labs are still winning. Nick injects that there's narrative noise from actors like Palantir talking their book (they want you on open-source so they can sell closed-source solutions on top). Startup founders on Claude Pro unlimited plans think Fable is the world's best because they're not cost-sensitive per token — a deliberate strategy from Anthropic and OpenAI to hook them early.

Grok's Path: Performance-Per-Cost Frontier But Not Top-Tier Intelligence

When asked if Grok can reach frontier status in 18 months, Frank says no — they won't have a number-one ranked model. But they'll reach the Pareto frontier on intelligence-versus-cost, depending on their compute access. Nick agrees. On significant user growth, Frank says yes, token growth will exceed user growth because Grok now has competitive performance-per-cost. Nick is more skeptical, calling them strategically impaired. The implication: Grok can nibble at pricing-sensitive customers but won't dethrone OpenAI or Anthropic at the high end.

Takeaways

  • Stop optimizing for frontier model performance — most enterprise tasks work fine with mid-tier models at 40% of the cost, so audit your actual use cases before paying Anthropic's premium.
  • Track cloud provider growth (Google Cloud, AWS, Azure AI) versus frontier lab growth — if clouds grow faster than OpenAI/Anthropic, it signals open-source finally broke through.
  • Don't assume 'open-source' means cheap — factor in your own infrastructure cost or cloud margins before switching off enterprise APIs.
  • If your team is on unlimited Pro plans, you're benchmarking the wrong models — switch one person to pay-per-token and watch your model preferences change instantly.

Key moments

0:35GPT-4.5 Turbo's cost efficiency breakthrough

GPT4 5.6 is kind of one it's half the cost for the same performance of GPT 5.5 and two it offers I mean at least also half the cost for the same performance as Fable the Anthropics reputed best-in-class model

3:18Chinese models dominate OpenRouter weekly leaderboard

the top 10 leaderboard for weekly usage is dominated by Chinese models. So number one is Tencent, number two is Xiaomi, number three is Deepseek, number four is Miniax

11:00AI agents remain fundamentally broken at independent execution

How many tasks can you give to an AI agent today and say Do this in a half an hour and it does it perfectly the way that you would do it if you were doing a half an hour's worth of work against it. It's basically zero at least as far as I can tell

16:00Anthropic's 10x growth versus Google Cloud test

Anthropic could be bigger than all of Google Cloud by the end of this year. And that needs to change if open source models are winning. And that's not right now

20:54Unlimited plans create biased model preferences

there's a set of actors in the space who are on claude pro plans but not on enterprise where they have all you can eat plans where they're like why would you ever go away from fable. You know, it's basically like unlimited intelligence and they don't face the cost of the marginal token

You just read one. Brevyd does this for every upload.

Follow ARK Invest and every new video comes back as a summary like this, in your morning briefing. No watching required.

Start for free