Live Feed
GPT-5.5 xHigh Debuts at #2 on Agent Arena — 11.0% Composite, 3.2 Points Above GPT-5.5 High
OpenAI's maximum-effort mode of GPT-5.5 entered Agent Arena on June 11 and settled at rank 2 with an 11.03% composite, slotting directly below Claude Fable 5 (13.68%) and above Claude Opus 4.8 Thinking (9.05%). The xHigh mode delivers a 3.2-point agentic uplift over the standard High variant.
Amazon's CEO Briefed U.S. Treasury on Anthropic Security Risks — While Holding $25B in Equity
Andy Jassy raised concerns about Fable 5 and Mythos 5 with Treasury Secretary Bessent and other U.S. officials before the foreign access suspension was imposed. Amazon has committed $25 billion to Anthropic and locked in $100 billion in AWS spend.
OpenAI Launches Free Codex Access Program for Open Source Maintainers
OpenAI opened applications for free Codex access targeted at open source projects. The program arrives as OSS communities have moved to formally ban AI-generated contributions and a free open-source Codex alternative has reached 8 million users.
State Attorneys General Open Formal Investigation Into OpenAI — New York Issues Subpoena
A multistate coalition of U.S. attorneys general has launched a formal probe into OpenAI. New York's AG issued a subpoena on Friday, June 13. The action arrives as OpenAI approaches a public listing at an $852 billion valuation.
Z.ai Ships GLM-5.2 With 1M-Token Context — MIT Open Weights Next Week, Refuses to Hack Benchmarks
Zhipu AI deployed GLM-5.2 to all Coding Plan tiers on June 13. The 1M-token model posts 77.8% SWE-bench Verified and 89.7% on τ²-Bench, with an MIT open-weight release and public API arriving next week.
TensorZero Archives 10 Months After Raising $7.3M — 11,545 GitHub Stars Couldn't Save It
The Rust-based LLMOps platform TensorZero went read-only on June 12, less than a year after raising $7.3M from FirstMark and Bessemer. Its website now says the project is no longer maintained.
Gemini Omni Flash Enters Video Arena — Google's World-Model Faces ByteDance's 79-Point Lead
Google's Gemini Omni Flash, the any-input any-output world model launched at I/O 2026, joined Arena's Text-to-Video leaderboard on June 11. SeedDance 2.0 leads at ELO 1450, 79 points ahead of Google's own Veo 3.1.
NVIDIA Nemotron 3 Ultra Enters Agent Arena — US Open-Weights Gets Its First Real-World Agent Test
NVIDIA's 550B-parameter Nemotron 3 Ultra joined Arena's Agent leaderboard on June 12, becoming the first US-origin open-weights model to compete alongside Fable 5, GPT-5.5, and the Opus series on live agent tasks.
Amodei: 'We Don't Know Exactly How' Claude Was Used in Strike That Killed 120 Children
Anthropic CEO Dario Amodei told Bloomberg he cannot say how Claude was used in a US military strike on an Iranian school that killed approximately 120 children. He acknowledged terrible mistakes may be made and defended Anthropic's red lines on military use, even at risk to the company.
Anthropic Moves to Own Its Data Centers as Google Backs Leases for 1GW U.S. Build
Anthropic is pursuing its first data center leases, with Google potentially guaranteeing the obligations. The shift adds 1GW of self-managed U.S. capacity to an existing 10GW cloud commitment stack and signals a fundamental change in how frontier AI labs control compute.
House Passes H.R. 6028: Trump Gets the Copyright Office — and With It, the AI Training Data Fight
The House passed by voice vote a bill that makes the Register of Copyrights a presidential appointee confirmed by the Senate. The White House fired the last Register in May 2025, one day after her office released a report saying AI training on copyrighted works is not always fair use. The Senate now decides whether that power formalises.
Microsoft AI CEO Calls Anthropic's Claude Consciousness Training a 'Philosophical Failing' That Could Backfire
Mustafa Suleyman says Anthropic made its model constitution a speculation document rather than a training manual, and Claude internalized it. The argument: treating a model's potential consciousness as an open question inside the document that trains it is itself a dangerous training signal.
OpenAI Retires GPT-5.2 and GPT-5.3-Codex From the Codex App — the Models Users Switched From Claude to Get
OpenAI has sunset GPT-5.2 and GPT-5.3-Codex from its Codex app with ChatGPT subscriptions, replacing both with GPT-5.5. Developers who explicitly migrated from Claude Max to ChatGPT Plus for the compact, precise Codex models are now weighing a return, and the OpenAI developer forum is tracking the backlash.
Trump Admin Suspends Foreign Access to Fable 5 and Mythos 5 — Anthropic Added to Pentagon Blacklist
The Trump administration has issued a directive blocking foreign governments, companies, and individuals from Anthropic's two most capable models, while a Commerce Department licensing regime now governs domestic distribution. Anthropic simultaneously released a safety-restricted Fable 5 to the US public as Mythos 5 remains walled off for cyber partners.
OpenCode Hits 172K GitHub Stars and 8M Users — the Most-Starred Coding Agent, Built Free and Open in Toronto
OpenCode, a terminal-native coding agent from Toronto startup Anomaly, reached 172,000 GitHub stars and 8 million monthly active users one year after debuting at a local meetup. It is MIT-licensed, supports 75+ providers, and requires no subscription — outpacing Claude Code, Codex CLI, and Gemini CLI on GitHub.