totally agree for "general purpose" models. The valid exception would be the cost comparison of a general purpose frontier model to a smaller model fine tuned for a narrow use case. The latter can still yield material cost savings, for example in the Crowdstrike data below, but only if you have a stable and sufficiently scaled use case to justify the R&D investment (not one time but ongoing to keep up w/ the frontier).

How quickly things can change! Here's the updated frontier with Opus 5.5, GPT-6 Sol & Luna incorporated. Opus 5.5 has kicked Astra off the frontier, giving Anthropic more ownership of peak intelligence than before, while OpenAI still dominates the mid to low end of the frontier with GPT-6 Sol & Luna coming in smarter and 50% cheaper than their prior 5.6 generations. Xiaomi's MiMo takes the rotating open source spot for today, but otherwise the entire AA Pareto frontier is dominated by just Anthropic & OpenAI.

There is something clearly different in how Anthropic & OpenAI scale effort levels. Doesn't show up on every benchmark, but these results on FrontierCode make it really clear. Anthropic models peak at lower effort levels, where as OpenAI models start low and climb up fairly consistently. The result is a better score at a lower cost for Anthropic models, but an unintuitive experience where increasing effort does not increase scores and might actually degrade performance (on this benchmark at least).

The AI safety discussion exploded last week. We think AI leaders' intentions are in the right place, but the extrapolations are going too far. Focusing on product safety is a good thing, and "pacing" the frontier doesn't necessarily mean stagnation in infrastructure investment or AI lab revenue. More in this week's newsletter 👇

With the combination of Astra & 5.6 Luna, OpenAI dominates the pareto frontier on the Artificial Analysis Intelligence Index, owning 11 out of 15 positions. This is why they are gaining share.

As model capabilities improve, research inside frontier labs appears to be accelerating. OpenAI just gave us a peek behind the curtain. At the start of the year, the median OpenAI researcher wasn’t using coding agents. Today, that researcher spends more than $600 per day on tokens—roughly $12,000 per month. The top 10% consume an order of magnitude more: $7,000+ per day, or roughly $140,000 per month. Token consumption is growing faster in Research than in any other department, while the share of researchers using four or more concurrent agents has more than doubled. Lines of code shipped are up more than 6× versus the 2025 average, and average experiments per researcher have roughly doubled. Quality is improving, too. OpenAI measured agentic task success rates across different time horizons and found improvement since the beginning of the year.

Seven months after Big Ideas 2026, the pace of AI adoption and infrastructure buildout continues to surprise to the upside. In this video, we revisit the data across compute, power, consumer monetization, and agent efficiency.

OpenAI's active agent users continue to climb, now at 20 million. CNBC reporting that OpenAI's run rate revenue has grown 35% quarter to date, with enterprise revenue up over 50%. Agent users are only 2% penetrated against the ~1 billion active user base of ChatGPT.

Grok 4.6 lands in a great spot on the Pareto frontier per Artificial Analysis indexing. Same avg. score as GPT 5.6 Sol (max) and 32% cheaper. Contrary to popular assumptions, of the 16 models on the frontier, only 2 are open, meaning if you are opting for an open model to save money you are likely overpaying.

Stop thinking of “Frontier AI” as only the most intelligent models. Think instead of the Pareto Frontier. That’s the frontier that matters: any model below it is, by definition, overpriced for the intelligence it delivers. Today, OpenAI has the most models on the Pareto Frontier (3). Anthropic, Moonshot, xAI, Meta, and DeepSeek each have 1 frontier model. Data from @ArtificialAnlys

OpenAI is now up to 10 million active users for their agent products. Incredible progress, and yet still so much room to grow relative to a TAM of ~70 million knowledge workers in the US and ~1 billion globally.

To those worried about accelerated depreciation for GPUs, this is a great example of nearly 4 year old chips being rented out in a (likely) multi-year deal And the buyer is (likely) getting a great return too given these GPUs will power in-demand models at 60-70% GM to Anthropic