Section
Artificial Intelligence
Frontier models, agentic systems, and the labs and silicon powering them.
Featured
OpenAI ships GPT-5.5 with a native agent runtime built in
The new model ships with a sandboxed code execution layer, persistent memory, and tool calling that no longer requires the Assistants API — a direct shot at Anthropic's Claude Code and Google's Gemini Agents.
Anthropic's Claude 4 Opus tops SWE-bench Verified at 78.4%
The frontier model now solves nearly four in five real GitHub issues unaided, a jump that pushes coding agents closer to autonomous junior-engineer territory.
Google DeepMind's Gemini 3 quietly rolls out to Workspace users
An A/B test surfaced the unannounced model in Gmail's Help Me Write last week, with response speeds roughly 40% faster than Gemini 2.5 Pro on long-context tasks.
Meta's Llama 4 leak suggests a 2T-parameter mixture of experts
Weights briefly appeared on a private HuggingFace fork before being pulled. Researchers say the architecture is a clear pivot away from dense scaling.
Perplexity hits $18B valuation in talks led by SoftBank
The answer-engine startup is now valued at more than three times the figure from its January round, even as Google and OpenAI build directly competing surfaces.
Mistral open-sources Magistral, a 70B reasoning model under Apache 2.0
The French lab's first dedicated reasoning model trades raw benchmark scores for permissive licensing — and immediate enterprise deployability.
xAI's Colossus 2 cluster goes live in Memphis with 550,000 H200s
Elon Musk says Grok 5 training has begun on the world's largest contiguous GPU supercluster, drawing 1.4 gigawatts from a mix of grid power and on-site gas turbines.
DeepSeek R2 matches o3 on AIME at one-twentieth the inference cost
Independent evaluations from Artificial Analysis confirm the Chinese lab's latest reasoning model delivers near-frontier math performance for $0.04 per million output tokens.
Microsoft Copilot rebuilds around a Mustafa Suleyman-led 'companion' mode
The redesign pushes proactive assistance, memory, and emotion-aware voice — and quietly drops the spreadsheet-first framing of the original Copilot launch.
Cohere shifts focus to sovereign AI deployments after $500M raise
The Canadian startup is now positioning itself as the default 'in-country' frontier model for governments worried about US and Chinese data jurisdiction.
Cerebras files for IPO with $1.4B in inference-only revenue
The wafer-scale chip startup's S-1 reveals that 87% of revenue now comes from Llama and DeepSeek inference, not training — a striking reversal from 2024.
Anthropic and Amazon strike $8B deal to anchor Indiana data center campus
The Trainium-based buildout, the largest single AWS infrastructure commitment to date, will power Claude inference across North America by mid-2027.