Anthropic pulls Claude Code from its $20 plan as SpaceX lines up a Cursor deal

This week in AI showed once again that the market is maturing and breaking its own rules at the same time. Anthropic is testing a steep price increase for Claude Code, Duolingo is changing its approach to internal AI metrics, and the closed Claude Mythos model was found through a guessed endpoint address.

Anthropic raises the stakes

Without a separate announcement, Anthropic removed Claude Code from the Pro plan at $20 per month. The cheapest tier with access to the tool is now Max 5x at $100. Terms for existing subscribers have not changed so far, but for new users it looks like a clear test of whether the market is ready to pay more for agentic development.

According to company employees on social media, this is an A/B test covering roughly 2% of new users. Still, the move is telling: even basic AI coding tools are quickly leaving the category of mass-market subscriptions.

The Claude Mythos episode is even more awkward for Anthropic. The model was launched in a closed preview for about 40 companies, including Microsoft, Apple and Google. On the same day, four members of a private Discord community that hunts for unreleased models gained access to it through a vendor endpoint.

The scenario was almost textbook: one person had legitimate contractor credentials, while the others reconstructed the URL format from data in the Mercor leak and from older model naming patterns. Anthropic confirmed the investigation and said it had found no signs of access beyond the vendor's environment.

Companies change their policies

Duolingo, which not long ago loudly declared itself an AI-first company, has stopped counting the use of AI tools in employee performance reviews. Given its March pivot toward AI and the cuts to contractors, this is an important signal.

Businesses seem to have quickly realized that counting prompts to a model or hours spent in AI services is meaningless: such metrics are easy to inflate, yet they say almost nothing about the quality of the result. The next stage of AI adoption inside companies is not monitoring consumption but a new set of expectations for team speed, output and autonomy.

The financial logic of the market is shifting too. According to the digest, Google is ready to invest $40 billion in Anthropic at a $350 billion valuation, while at the same time assembling an internal team to compete with Claude. For Big Tech this is now a normal arrangement: invest in a partner, use its technology, and build your own answer in parallel.

Against this backdrop, talk of a bubble has not gone away, but for now the market is betting on further growth. According to Polymarket, the probability of a hard collapse of the AI bubble before the end of 2026 remains relatively low.

The model race speeds up

The models section this week was especially dense. Almost every release was about one of two things: making an AI agent more useful in real-world development, or sharply cutting the cost of running it. That is an important shift, because the market is no longer arguing about whether such models are needed, but about which one solves a practical task faster and what it costs in production.

• Kimi K2.6 has come close to the closed leaders on coding benchmarks and bets on proactive agents that can work in the background for long periods without human involvement.
• GPT-5.5 scored 58.6% on SWE-Bench Pro, and inside OpenAI it has already been run on production Codex traffic, where new heuristics raised token generation speed by 20%.
• DeepSeek V4-Pro and V4-Flash got a 1 million token context window and put noticeable pressure on pricing: Flash is tens of times cheaper than top closed models.
• SpaceX obtained an option to buy Anysphere, Cursor's parent company, for $60 billion, with the option to close the deal before the end of 2026 and a record breakup fee of $10 billion.

The Cursor story is especially telling. If the deal does close after SpaceX's IPO, AI coding will finally stop being a niche for developer startups and become a strategic layer of big-tech infrastructure. Cursor's access to the Colossus supercomputer and its work on the next generation of Composer models show that the fight is no longer only over the editor interface, but over compute, data and iteration speed.

What this means

The week's main takeaway is simple: the AI market is getting more expensive, tougher and less naive. Companies are pushing commercialization harder, revisiting internal rules for working with AI, and increasingly finding that the main risk is no longer model quality but managing access, infrastructure and the speed of adoption.