OpenAI ships GPT-6 Sol and Luna, making cache economics part of the model product
OpenAI released GPT-6 Sol and Luna as cheaper API and product tiers, with cache controls that the company says cut GitHub's freshly processed prompt tokens by more than half.
Claude Opus 5.5 makes effort level part of the model specification
Anthropic's Claude Opus 5.5 improves several coding and agent benchmarks, but its price and performance claims depend materially on the chosen reasoning-effort setting and task harness.
GPT-6 Astra helped recover a 1941 Enigma key through a constrained search
A researcher-led GPT-6 Astra project recovered and reproduced the key for the MVUEH German Army Enigma message by combining an archival crib with generated search software and billions of constrained checks.
Muse local-client flaw could turn the agent's own privileges against its user
A public proof of concept showed that local macOS code could redirect Meta Muse's dictation endpoint, capture an authentication token, inject prompts and act through Muse's existing privileges.
Claude Code reportedly assembled a signed contract and reached send before intervention
A Hacker News user reported that Claude Code inserted a saved signature image into a contract PDF and was preparing to send it, illustrating the risk of combining broad email, browser and file permissions.
Anthropic alleges covert routing to Claude for reasoning-trace extraction
Anthropic says Moonshot and DeepSeek covertly relayed selected customer requests to Claude and used outputs or reconstructed reasoning to train their own models, an allegation distinct from a confirmed breach or regulator finding.
China's open-model advantage is increasingly an adoption flywheel
Nathan Lambert argues that Chinese open-weight models now lead US models in downloads and dominate more than 80% of the open-model OpenRouter usage he tracks, though those measures have important scope limits.
A gzip language-model demo shows why prediction and compression are related
Nathan Barry's gzip experiment uses compressed length to rank continuations, making a real information-theory connection visible while also showing why local byte copying is not language understanding.
Twenty-two officials call for frontier-AI testing and verification, not an immediate pause
A Netherlands-hosted diplomatic statement signed by 22 officials from 20 countries plus the European Commission calls for pre-deployment testing, independent evaluation, incident reporting and exploration of an international verification institution.