Gloves off, knives out

Anthropic cuts Windsurf off... and then lets them back in?

Photo by Anna Saveleva on Unsplash

I wrote the initial part of this article in June… things move fast in the AI world!

Recently, I heard the news that Anthropic had removed first-party access to Windsurf. Anthropic’s Claude 3.5+ models have been the ones that I have mostly used in Windsurf and Cursor, so this news is somewhat disconcerting. If I wanted to use Claude in proactive AI coding¹ workflows, I’ll have to use another tool like Cursor, Void or nao.

I’ve had such good results using these Claude models for this workflow that I haven’t felt the need to try Gemini 2.5 Pro yet, but it looks like I will end up doing that sooner rather than later.

A few weeks ago, I wrote about how I thought that Windsurf joining OpenAI could open a whole raft of possible advantages in terms of better model access, specialised model creation for different Windsurf flow states, and just raw GPU muscle for the Windsurf team to use to build new models of their own - which it turns out they were doing long before the acquisition.

This development reveals the darker and more troubling side of the situation. Despite many thousands or even hundreds of thousands of Windsurf users enjoying Anthropic models in Windsurf, who have no additional affiliation with OpenAI prior to the acquisition, Anthropic (clearly for competitive reasons) has decided to revoke access.

This isn’t unexpected; Windsurf essentially serves as both a training data generator and a tool for coding. It continuously generates context for AI-generated code, including the code that was actually generated, as well as various approximate and direct confirmations of quality. It’s very easy to generate a huge training set where Anthropic’s models have generated the code, and it has been confirmed to be good. This training set is then used to build Windsurf’s own models. Clearly, this was tolerable to Anthropic while Windsurf was independent, but now, faced with the possibility of enhancing their largest competitor’s models, too… a bridge too far.

While the advantages I discussed above and in my previous post are clear for Windsurf within OpenAI, there are also clear disadvantages. If Cursor remains independent, it is possible that they will continue to enjoy being able to use any model they like, whereas Windsurf is being restricted to fewer. If GCP decide also to remove access to Gemini models, then Windsurf becomes very restricted indeed.

It would seem very unlikely that OpenAI would go tit for tat and restrict use of its models in Cursor in order to counter the disadvantage now felt by Windsurf, as Cursor’s use of OpenAI models generates a very valuable training dataset for OpenAI.

As I said above, I wrote the first part of this post in June and then published some posts related to other ideas instead. Some unexpected events occurred after this!

The OpenAI deal suddenly falls apart, and the first we hear about it is that the founders and core AI engineers at Windsurf are moving to GCP, seemingly leaving the rest of the company behind. The GCP deal is a HALO agreement, where they have a non-exclusive license for Windsurf’s IP, for which they paid $ 2.4 B.

You’d imagine that with this much money going round, the investors in Windsurf and the remaining staff are made whole. After all, many members of the Windsurf team would have incurred significant opportunity costs to join a high-growth organisation with no reason to believe it would soon cease to be an independent company. However, we don’t know the details, and we probably never will. By the time we might have the information, our attention will have moved on, and we won’t take the time to find it.

The reason the OpenAI deal fell apart seems to be that OpenAI didn’t want the intellectual property acquired with Windsurf to be accessible to Microsoft, but they couldn’t get this to work legally so they pulled out of the deal. This is a worrying sign for OpenAI; clearly, they felt it was worth a lot of value to acquire Windsurf, and they had looked at Cursor before, but the price was too steep.

AI-powered IDEs are one of the clearest winning use cases for LLMs, which also have a flywheel effect in generating high-quality training data to then improve LLMs’ ability to serve the use case. OpenAI took the time to announce the deal to great fanfare and then had to pull out, which is frankly pretty embarrassing. I’m sure there will be other tales about what happened in order to soften the impact, but fundamentally, if you’re going to announce something like an acquisition, it doesn’t look good if you’re forced to back out.

I think this effect of the investment from Microsoft and OpenAI’s legal structure could make OpenAI less able to compete against its rivals who don’t have the same encumbrance. If Microsoft had access to the IP, then OpenAI would strengthen Microsoft through their acquisition and wouldn’t gain any IP that was uniquely theirs, eliminating the point of acquiring it in the first place. The more I think about it, the more this feels like a noose.

Then as a final twist, the remainder of Windsurf gets acquired by Cognition, who are the makers of Devin. I really think this is a great fit². There is a continuum of how AI can be involved with software engineering, with Windsurf on the left (pretty close to how SWE is done now, but with copilot and IDE agent) and Devin on the right (about as far away from how SWE is done now as is possible³). The activity on Windsurf produces the perfect training data and subagents to then improve Devin.

You can envisage land and expand happening, initially with engineers at a company having access to Windsurf as the first product offering in a deal with Cognition, and then gradually transferring more and more of the process until they are using Devin, with the engineers supervising and designing instead of actually building.

The first steps would be using Devin in CI to perform linting, document changes, adding test coverage, and implementing telemetry coverage. This would also involve enforcing code standards, adhering to package update policies, and handling DevOps tasks, including penetration tests, minor bug fixes, and investigating and resolving medium to major bugs. Then, once Devin was trusted to do these things, you’d start to use it with Jira to complete tasks of a small t-shirt size, which are well-documented and described by a product manager. The product managers begin to see the point in finally writing good PRDs with clear requirements, and writing bigger epochs, with the assistance of AI, that get split into many smaller tasks that can be worked on in parallel by many agents. I think where you go beyond this point, you’re in Loveable and Bolt territory.

Working with Claude Code in recent days has made me feel like the progression I described above is possible and even probable, when I didn’t feel like that before. It’s made me that much more capable than I was with Windsurf (in fact, I’ve begun to use Claude Code in Windsurf, and it’s a pretty great experience, especially where you know the space very well).

Now that Windsurf are part of Cognition, they have been allowed to offer all Anthropic models again, so we’re back to where we started at the beginning of this post… but a lot has changed, too.

1. As opposed to a copilot in the text editor of the IDE, which I consider reactive. I haven’t used copilots very much recently and have mostly talked to Cascade and watched what it has done in the text editor, and then fed back changes I wanted here, too.

2. Beyond my thoughts on why it’s a good fit, the now lack of core engineering at Windsurf, with very strong GTM remaining, and the opposite being true at Cognition, also means the acquisition makes great sense from a talent perspective.

3. Except if you consider tools like Loveable and Bolt, where SWE is abstracted away from product development entirely.