Cursor $0.50 Model, OpenAI’s IPO Filing, and the 8,000 Meta Employees Who Trained Their Replacements

Participate in the quiz based on this newsletter and the lucky five winners will get a chance to win a coffee mug!

Take the Quiz

Cursor Just Shipped a Coding Model That Matches GPT-5.5 – For One-Tenth the Price

Cursor launched Composer 2.5 on May 18th. The headline claim is that it matches Claude Opus 4.7 and GPT-5.5 on coding benchmarks. That alone would have been worth a mention. But the part that actually changes the conversation is what it costs: $0.50 per million input tokens. Opus 4.7 charges $5. GPT-5.5 is in the same range. That is not a small gap that is one-tenth the price for comparable output on the tasks developers actually run every day.

Here is what is under the hood. Composer 2.5 is built on Moonshot AI’s open-source Kimi K2.5 base  a one-trillion-parameter Mixture-of-Experts model from a Chinese lab that is quickly becoming one of the most widely used open-weight foundations in the industry. But the base model is not really the story here. Cursor spent 85% of the total compute budget on their own post-training pipeline: reinforcement learning, 25x more synthetic coding tasks than Composer 2, and a new targeted feedback technique that teaches the model how to correct specific mistakes in the middle of a task rather than only learning from how the full session ends. That last detail matters more than it sounds. Long agentic coding sessions fail in subtle ways  a bad tool call here, a wrong file edit there  and a model that can course-correct mid-session is meaningfully more reliable than one that only optimises for the final outcome.

The benchmark numbers: 79.8% on SWE-Bench Multilingual, 63.2% on CursorBench v3.1. Both are within range of Opus 4.7. Terminal-Bench, which tests real command-line task execution, is where GPT-5.5 still holds an edge worth knowing before you switch anything critical.

For this first week, Cursor is doubling the included usage at no extra cost. That is a real window to test it against your actual codebase before making any decisions.

Why this matters :

The pricing gap between frontier models and specialized coding agents is closing faster than most people expected. If you are currently running heavy Claude Code or GPT-5.5 agentic sessions and burning through budget, Composer 2.5 is worth a serious eval this week  not because the benchmark numbers say so, but because the cost structure is genuinely different and the real-world coding task performance appears to back it up.

🔗 Read the full story

OpenAI Filed for an IPO. Here Is What That Actually Means for Builders.

This one has been rumoured for a while. But this week it moved from speculation to something concrete.

OpenAI is reportedly preparing to file a confidential S-1 with the SEC  the first formal step toward an IPO. A confidential filing means the document stays private until OpenAI chooses to make it public, which gives them room to work through the numbers without the scrutiny of a live filing. But make no mistake: you do not file a confidential S-1 as a dress rehearsal. This is the real process starting.

The timing is worth sitting with for a moment. OpenAI launched in 2015 as a nonprofit. It added a “capped profit” structure in 2019. It has been raising private capital at increasingly staggering valuations ever since. And now, in 2026, the same week that a California jury unanimously rejected Elon Musk’s lawsuit arguing the company abandoned its nonprofit mission  OpenAI is moving toward public markets. The irony is almost too clean.

Here is what an IPO actually means for the industry. When a foundational AI company goes public, the rules of the game change. Quarterly earnings calls. Shareholder pressure. Revenue growth expectations that do not leave much room for “we decided to slow down because safety concerns came up.” OpenAI will have to explain, in plain numbers, how it turns the most expensive model training process in history into a sustainable business. The pressure to monetize harder more enterprise contracts, more API pricing, more upselling  will be real and visible in a way it simply is not when you are a private company with patient capital.

This also sets a clock for everyone else. Anthropic just reported its first quarterly operating profit and is closing in on a $900 billion valuation. Google DeepMind is embedded inside one of the most profitable companies on Earth. The moment OpenAI files publicly, every major AI lab’s financials will be under a new kind of scrutiny  not from journalists, but from institutional investors comparing balance sheets.

Why this matters:

If you are building on top of OpenAI’s API, an IPO changes your vendor relationship over a multi-year horizon. Public companies optimise for different things than private ones. Pricing, model deprecation timelines, and enterprise-first feature decisions will increasingly reflect what moves the stock. That is not a reason to panic or switch today  but it is a reason to pay attention to how OpenAI’s product decisions start to shift once public market pressure is real, not theoretical.

🔗 Read the full story

Meta Used Its Own Employees to Train AI. Then Fired 8,000 of Them on the Same Day.

There have been a lot of uncomfortable AI-and-jobs stories over the past few years. This one is different. This one is going to stay with people.

A leaked audio recording from a Meta all-hands meeting on April 30th surfaced publicly on May 19th the exact same day approximately 8,000 Meta employees received layoff notices. In the recording, Mark Zuckerberg explains what Meta calls the “Model Capability Initiative” an internal program that tracked employee activity across Gmail, Google Chat, the internal assistant Metamate, and VS Code to train Meta’s AI models on how highly skilled people actually work.

Zuckerberg’s framing in the audio was careful: the data was anonymised, no humans were watching, it was purely used to improve model capability. “The AI models learn from watching really smart people do things,” he says. He described it as feeding a large amount of behavioural content into the model  how experts write code, structure communication, approach problems.

Here is what made this land so hard. It was not just that Meta had been collecting this data, though the lack of explicit informed consent is a genuine ethical problem on its own. It was the timing. The people whose working patterns were studied to train the model that would make certain roles more automatable  those same people were being handed termination notices in Singapore at 4am local time, then rolling through European and US time zones, hours after the audio dropped.

Employees inside Meta responded immediately. Flyers appeared in meeting rooms. A petition against the initiative circulated internally. On social media, the phrase “train your replacement” showed up in hundreds of posts. Someone called it a real-life Black Mirror episode. That description spread faster than any official statement Meta released.

Meta has committed more than $125 billion to AI infrastructure and data centres in 2026. The internal logic is clear: smaller, AI-assisted teams doing more. The human cost of that equation became very visible this week.

Why this matters:

This story is not really about Meta. It is about a pattern that is going to repeat at other companies probably already is  just with less dramatic timing. The question of whether it is ethical for an employer to harvest the work patterns of its own employees to train systems that reduce headcount is not settled. It is barely even being asked at most organisations. What happened at Meta this week is the version of this story that broke through because the timing was so stark. But the underlying dynamic is not unique to Meta. It is worth asking whether it is happening where you work, and whether anyone has thought carefully about the consent and disclosure questions involved.

🔗 Read the full story

.