AI News for June 27, 2026 — 7min.ai

[NEWS] GPT-5.6 Sol leads Mythos 5 on cyber benchmarks at half the cost, enters gated release: OpenAI's flagship Sol — alongside the balanced Terra and the fast, cheap Luna — is the company's strongest model yet, topping prior systems on coding and matching or beating Anthropic's Claude Mythos 5 on cybersecurity.
[NEWS] US clears Anthropic to restore Claude Mythos 5 for 100+ institutions: Commerce Secretary Howard Lutnick told Anthropic chief compute officer Tom Brown that appropriate safeguards now permit more than 100 approved US companies and government agencies — including foreign-national employees at those organizations and at Anthropic itself — to access Claude Mythos 5.
[NEWS] Qualcomm investor day: new AI accelerators, $40B non-handset target, shares jump 15%: At its Manhattan investor day, Qualcomm CEO Cristiano Amon unveiled new AI accelerators and CPUs aimed at Nvidia's data center AI dominance, and projected total revenue from non-handset businesses reaching $40 billion — double a forecast from two years ago.
[RESEARCH] Google Research speeds Gemini Nano on Pixel with frozen Multi-Token Prediction: Google Research detailed a method to accelerate Gemini Nano on Pixel devices using frozen Multi-Token Prediction, an on-device efficiency technique that lets the model generate multiple tokens per step without retraining the base weights.
[NEWS] OpenAI hires Apple's Vision Pro and smart glasses chief Paul Meade: OpenAI poached Paul Meade, the Apple executive who led the Vision Pro headset and the company's smart glasses program.
[RESEARCH] MIT's Masked IRL teaches robots vague tasks with 5x less demonstration data: MIT CSAIL researchers built Masked Inverse Reinforcement Learning (Masked IRL), which uses one large language model to flesh out ambiguous human prompts from a demonstration and a second LLM to decide which details a motion plan should follow.
[NEWS] AI startup Lindy dumps Claude for DeepSeek as buyers chase efficiency: AI agent startup Lindy ditched Claude entirely for DeepSeek, hosted by a US company on US soil, and saw its cost curve crash to the ground while saving millions.
[NEWS] NYT says Microsoft built a supercomputer to help OpenAI infringe copyrights: The New York Times moved to amend its copyright suit against OpenAI and Microsoft, alleging Microsoft actively encouraged infringement by building a bespoke supercomputing system, ranked among the world's most powerful, to train OpenAI's models on NYT works.
[RESEARCH] Epoch AI's MirrorCode benchmark has Claude Opus 4.7 leading at 56%: Epoch AI and METR released MirrorCode, a benchmark requiring models to reimplement 25 complete programs from scratch, spanning Unix utilities, bioinformatics, interpreters, cryptography, and compression, with no access to the original source and hidden end-to-end tests.
[NEWS] Linux Foundation and 20 tech giants launch Akrites to harden open source: The Linux Foundation announced Akrites, a coordinated initiative to patch vulnerabilities in widely used open-source software before AI tools can exploit them.

About 7min.ai · Subscribe to newsletter