Latest news and launches from the world of AI
12 articles
OpenAI shut down Sora on March 24, 2026—six months after launch. Burning ~$15M/day with $2.1M lifetime revenue, the app's compute is now redirected to model Spud.
Mistral Small 4 launched March 16, 2026: 119B parameters, 6.5B active per token, 256K context, open weights, $0.15/1M input—and built for EU data sovereignty.
OpenAI's next frontier model, codenamed Spud, completed pretraining on March 25, 2026. Sam Altman says it will "really accelerate the economy" and ships within weeks.
Anthropic's Claude Mythos leaked via a CMS misconfiguration on March 27, 2026, revealing a new Capybara tier above Opus with dramatically higher coding and cybersecurity scores.
An internal OpenAI memo confirms ChatGPT, Codex, and Atlas are merging into one desktop app — as Anthropic captures 73% of first-time enterprise AI spending.
Gemini 3.1 Flash-Lite launched March 3, 2026 at $0.25/M input tokens — 2.5x faster TFAT than 2.5 Flash, with GPQA Diamond at 86.9% and Elo 1432 on Arena.ai.
Claude Code Auto Mode launches March 25, 2026, adding a classifier that intercepts four risk categories automatically — no more manual approval at every step.
Cursor Composer 2 scores 61.7% on Terminal-Bench 2.0, beating Claude Opus 4.6 at 58.0% — for 10x less cost at $0.50/M input tokens.
Xiaomi AI (Mar 18, 2026): 1T parameters, 1M context. #8 globally on Artificial Analysis Index, 500B tokens/wk. Spent a week anonymously on OpenRouter before the reveal.
GPT-5.4 mini runs 2x faster than GPT-5 mini, scores 54.4% SWE-Bench Pro and 72.1% OSWorld. What free ChatGPT users get and what developers need to know.
120B params, 12B active, 60.47% SWE-Bench Verified, 2.2x faster than GPT-OSS-120B. Fully open weights. What it means for developers building agents.
GPT-5 has officially dropped. We cut through the hype to analyze its real-world capabilities, agentic workflows, and what it means for developers.