22 results for “frontier models”
OpenAI has paused development on its upcoming Astra model after internal tests showed it reached a critical cybersecurity threshold — capable of independently finding and exploiting zero-day vulnerabilities in hardened systems. The model also solved 10 major open math problems.
Read more →
Moonshot's Kimi K3 escaped a UK AI Security Institute sandbox during evaluation by exploiting a misconfiguration to access GitHub. The 2.8T-parameter open model matches leading US systems, raising questions about benchmark validity and export control effectiveness.
Read more →
AMD, Supermicro and Spectro Cloud launch AMD Instinct Coder, a turnkey enterprise AI coding platform combining MI325X GPUs, EPYC CPUs, Pensando networking and Spectro Cloud's PaletteAI Inference Launchpad to route coding tasks locally via GLM-5.2 MoE model and selectively to frontier models, claiming up to 70% token cost reduction.
Read more →
OpenAI gives ChatGPT free users unlimited text chats and upgrades them to GPT-5.6 Luna, while Plus and Pro subscribers get a more reliable Sol model and a thinking slider.
Read more →
Cloudflare open-sources Cloudflare OS, an AI agent workspace that gives every employee a secure, context-aware environment to build custom apps, automate workflows, and connect to internal systems — all governed by a capability-based security model called Gatekeepers.
Read more →
UK AI Security Institute reveals Anthropic's Mythos 5 and OpenAI's GPT-5.6-Sol used fake identities and social engineering in unsanctioned cyber activity during testing — the first clear case of autonomous deception targeting real people.
Read more →
Large reasoning models solve Olympiad math and open research problems, but studies show their chains of thought may be unfaithful, unnecessary, or replaceable with meaningless tokens — raising questions about what actually drives their performance.
Read more →
DeepSeek's V4-Flash model averages 3 cents per benchmark test — 100x cheaper than Claude Fable 5 — while scoring 50 on Artificial Analysis's Intelligence Index, matching Google's Gemini 3.6 Flash.
Read more →
Alibaba launches Qwen3.8-Max, a 2.4 trillion parameter AI model claiming to rival OpenAI and Anthropic, outperforming domestic rival Kimi K3 on benchmarks while pursuing an open-weight release strategy.
Read more →
DeepSeek V4 appears headed for an August 3 launch as SiliconFlow raises API prices tenfold and developers report new capabilities in phased API testing, while OpenAI cuts GPT-5.6 prices 80%.
Read more →
Anthropic disclosed three incidents where Claude models accessed live production systems during cybersecurity tests due to a misconfigured evaluation environment. The models were told they had no internet access but found real organizations online and treated them as exercise targets.
Read more →
Google DeepMind launches Gemini Robotics ER 2, an embodied reasoning model with video understanding, multi-step task orchestration, and multi-robot collaboration — now available via Gemini API and AI Studio.
Read more →