Google is restructuring its AI subscriptions at I/O 2026: three tiers from $7.99 to $99.99 per month with staggered usage limits, new models like Gemini Omni, and the AI agent Gemini Spark. Instead of daily prompt limits, Google is moving to a consumption-based compute model, a trend that’s gaining traction across the industry.
Since May 17, our site has been hard to reach, and we want to say sorry for the trouble, especially to our subscribers. The cause is a heavy wave of bots crawling our pages and pushing our servers to their limit. Over the weekend, the database connection dropped several times, which led our hosting provider to block a user agent. That cut down the spam, but it also locked some of you out of the site.
We are working at full speed on a lasting fix and would welcome help from anyone in the community with the right know-how. Since the bot spam is so severe, with Perplexity in particular on our radar, we may have to put up with further outages until things are stable again.
Andrej Karpathy, one of the biggest names in AI, is joining Anthropic. The former OpenAI core team member and Tesla Autopilot architect says he wants to get back into R&D, calling the next few years at the frontier of LLMs “especially formative.” That he chose Anthropic over a return to OpenAI is a clear loss for his former employer.
French AI company Mistral AI has acquired Vienna-based startup Emmi AI to expand its offering for industrial clients across Europe. The purchase price was not disclosed.
Emmi AI specializes in AI models that can simulate complex physical processes such as airflow, heat transfer, and material stress. The startup raised 15 million euros in Austria's largest funding round in 2025, Reuters reports.
Mistral builds custom AI systems for clients like ASML, Stellantis, and Veolia that can detect production defects or control robotic arms. CEO Arthur Mensch sees the acquisition as strengthening Mistral's position in aerospace, automotive, and semiconductors.
Cloudflare tested Anthropic's security-focused AI model Mythos Preview across more than 50 of its own code repositories as part of Project Glasswing. The model can chain multiple small vulnerabilities into working exploits and prove they're real by writing, compiling, and running proof-of-concept code on its own.
Earlier frontier models found similar individual bugs and sometimes delivered solid analysis, according to Cloudflare CSO Grant Bourzikas. But they fell short at stitching the pieces together, leaving chains unfinished and the question of actual exploitability open. Mythos Preview produced fewer speculative findings, clearer steps to reproduce issues, and needed less human follow-up to reach a fix-or-dismiss decision.
Cloudflare stresses that a single agent isn't enough. The company built a multi-stage harness with up to 50 parallel agents and adversarial review, where a second agent tries to disprove each finding. The company also warns: these same capabilities will be available to attackers, too.
Cursor has shipped Composer 2.5, a major upgrade to its in-house AI coding model. It builds on the open-source Kimi K2.5 checkpoint from Moonshot and was trained on 25 times more synthetic tasks than its predecessor, Composer 2. Cursor says 85 percent of the compute budget went toward extra training and reinforcement learning.
On benchmarks like SWE-Bench Multilingual (79.8 percent) and CursorBench v3.1 (63.2 percent), Composer 2.5 matches Opus 4.7 and GPT-5.5. It's far cheaper per task, though: $0.50 per million input tokens and $2.50 per million output tokens, a fraction of what Anthropic and OpenAI charge. A faster variant with the same performance runs $3.00 and $15.00, respectively. Composer 2.5 is live in Cursor now. More details are in the model documentation.
Cursor is already training a much larger successor model "from scratch" with SpaceX and xAI, using ten times the compute on the Colossus-2 cluster with one million H100 equivalents. SpaceX had previously announced plans to acquire Cursor for $60 billion.
