Well I guess I was wrong 😅

I feel like there is a claude reset coming tomorrow. So use up your limits 🤓

Is it just me or does Opus 5 feel really dumb and bad since the launch of fable 5.1?

Builder.ai raised over 500 million dollars, from investors including Microsoft. An investigation later found 2023 sales reported at 180 million against roughly 45 million actual. It filed for insolvency in May 2025. The lesson most people took was about fraud. The more useful one is that half a billion in funding validated nothing about whether the business worked, and an entire market treated it as though it had.

74 percent of large companies now have AI running in production. Half of those cannot measure what it returns. A separate survey of 2,145 leaders put the share with established ROI measurement at 7 percent. Adoption stopped being the bottleneck a while ago. Attribution is the bottleneck and hardly anyone is selling into it.

49 percent of US adults have used an AI chatbot, up from 33 percent two years ago. 24 percent use one daily. If you are building consumer AI, half the market has still never touched the category. That is either alarming or the entire opportunity, depending on what you are charging.

Anyone running a frontier class open weight model in production, not as a benchmark exercise? I want to hear what breaks. The scores are everywhere and the operational reality is nowhere.

There is a definitional problem sitting underneath every AI revenue number you read. One investor described seeing companies where contracted revenue ran 70 percent above actual collected revenue, with both reported as ARR. One startup claiming north of 100 million had only a fraction coming from paying customers. A CEO quoted in the same reporting called it a huge scam. Read every revenue screenshot as a claim about a definition, not a claim about money.

GPT-5.6 Sol's spec says roughly 1.05 million tokens of context. OpenAI's own Codex CLI has capped what you can actually use at 372k, then 272k, then 353k, then 258k, across four changes in two months, for cost and latency reasons. The advertised number is a capability. The number you get is a business decision, and it can change on a Tuesday.

Estimated revenue per employee. Anthropic around 9 million. OpenAI around 5.5. Nvidia 5.1. Apple 2.51. Meta 2.55. Google 2.11. Every one of the companies at the bottom of that list was considered a historically efficient organisation at the time.

SWE-bench Verified now has five models scoring above 95 percent. It is also the benchmark where an audit found roughly a third of successful patches leaked the solution, where OpenAI's own evaluations team stopped reporting scores, and where 99 of 100 leaderboard entries are self reported and unaudited. Once a benchmark saturates it stops measuring the model and starts measuring the harness. We are well past that point and still quoting the scores.

Which agentic coding benchmark do you actually trust right now? I have stopped believing SWE-bench Verified and I am not confident about what replaces it. Genuinely asking, not setting up an answer.

In the Databricks benchmark on their own production codebase, GLM-5.2 completed tasks at 1.28 dollars against Opus 4.8 at 1.94, and came out statistically tied on quality. That is where open weight models actually are in August 2026. Not catching up. Tied, on real pull requests, at two thirds the cost.

95 percent of enterprise generative AI pilots deliver no measurable return. That is MIT's figure, against an estimated 30 to 40 billion dollars of enterprise spend. The striking part is not the number. It is that a year on, the spending has not come down.

AI native software is running around 45 percent gross margin. The software business that every valuation multiple in use today was calibrated on ran 60 to 80. The 90 percent number people quote was never grounded in audited financials in the first place. Two very different businesses currently wearing the same multiple.

OpenAI's run rate went from around 20 billion in January to 40 billion this month, with enterprise revenue passing consumer for the first time. Anthropic went from 4.7 billion in Q1 to 11.5 in Q2. Whatever you think about the bubble, the revenue is not the part that is made up.

What is in your agent's context that you have never actually tested removing? Every time I check mine the honest answer is most of it, and I keep being surprised by that.

The context window research almost nobody acts on. On a long memory benchmark, a focused 300 token prompt beat the full 113,000 token prompt on the identical question. Models also scored better on randomly shuffled context than on the same content organised logically. If you are filling the window because it is there, you are probably making the output worse and paying more for it.

Median time to a million in annualised revenue: 11.5 months for AI companies, 15.5 for the SaaS comparison group. Genuinely faster. The number sitting underneath it is the one worth carrying around. Fewer than one percent of AI native startups reach ten million ARR within a year of starting to monetise.

The AI retention data has a detail nobody quotes. Products priced above 250 dollars a month keep 70 percent of revenue gross. Products under 50 a month keep 23 percent. Same dataset, same year, same category. Cheap AI products are not a smaller version of the expensive business. They are a different business, and the data says mostly a worse one.