Cognition Raises $1B, More Than Doubling Its Valuation in Just 8 Months

Cognition, the AI coding company behind Devin, has raised $1 billion at a $25 billion pre-money valuation. The company's annual recurring revenue (ARR) has reached $492 million, and its valuation has more than doubled in just 8 months since its previous funding round. The raise is drawing attention as a major round that symbolizes the rapid growth of the AI coding space.

The company's flagship product, Devin, is an AI software engineer that autonomously writes, tests, and debugs code. Backed by enterprises' growing demand for improved development efficiency, revenue has been expanding rapidly. In the AI coding market, OpenAI's Codex and Anthropic's Claude Code are also locked in fierce competition, and the market as a whole is growing at a rapid pace.

Latest SWE-rebench Update: GPT-5.5, Opus 4.7, and Kimi K2.6 Face Off

The March–May 2026 edition of SWE-rebench, a benchmark that measures software engineering capability, has been released, and comparisons of leading models' performance are generating buzz. GPT-5.5, Claude Opus 4.7, Kimi K2.6, and Cursor (Composer 2.5) all rank near the top, with the gaps in the leading group now razor-thin.

Meanwhile, MSN coverage suggests that "Codex has knocked Claude off the AI programming throne." Benchmark scores don't necessarily match how models perform in real-world development, making model selection harder than ever.

ReAligned-Qwen3.5 Released

ReAligned-Qwen3.5 has been released in the open source community. A realigned version of Qwen 3.5, the model aims to deliver improved alignment and greater practical utility. It is attracting attention in the LocalLLaMA community, adding yet another high-performance option for running models locally.

Google's AI-Powered Fitbit App Gets Review-Bombed

Google has redesigned the Fitbit app as a new version packed with AI features, and it is facing fierce backlash from users. According to Kotaku, the new app has been described as "terrible AI garbage," drawing a flood of negative reviews. Users criticize the loss of the old, simple, easy-to-use interface, underscoring consumer resistance to forcing AI onto existing products.

PostHog Makes AI Training on User Data Opt-In by Default

PostHog, a product analytics platform, has announced a policy of using user data to train AI models. The problem is that the setting is opt-in by default — consent is assumed unless users change it. The move has sparked a major debate on Hacker News, with criticism pouring in from a data privacy perspective. It once again brings to the surface an industry-wide challenge: SaaS companies using user data for AI training.

What 1,000+ Experiments Reveal About the Reality of "Self-Improving Agents"

A study gaining attention in the MachineLearning community reports the results of more than 1,000 harness experiments testing whether AI agents can improve their own performance. While agents did show significant gains at one point by improving the "harness" (the test environment) themselves, the results also suggest there are limits to continuous self-improvement. It stands out as an important data point for thinking through the feasibility of self-improving AI.

---