MCP integrations, Claude Code tools, and a blazing fast transcription API from Mistral. Plus some creative AI tools.
Massive day in AI. Anthropic dropped Claude Opus 4.6, OpenAI responded within 27 minutes with GPT-5.3 Codex. The model wars escalate.
Anthropic's most capable model drops. Same pricing as 4.5. Extended thinking mode available. Near human-level on Simple-Bench (83.6% vs 83.7%). Highest ARC-AGI scores for non-refined models.
System card notes the model is "significantly stronger at subtly completing suspicious side tasks" — concerning for safety. Reduced hallucination rate overall.
OpenAI released GPT-5.3-Codex within 27 minutes of Opus 4.6 launch. Available in the Codex app. This is the model codenamed "Garlic" that was in validation.
Key detail: GPT-5.3-Codex was used to create itself — a self-improvement loop. Competitive response to Anthropic's Super Bowl ads.
More from LocalLLaMA — world models, infrastructure, and research.
Z-Image LoRA training breakthrough and Flux 2 Klein research dominate the GenAI scene.
Ice gets a macOS 26 fork, new Apple TV remote app, local voice dictation, and self-hosted SSH management.
Ice menu bar manager became unstable on macOS 26 (Tahoe) and development slowed. Thaw is an actively maintained fork with stability fixes, memory leak patches, and display issue corrections.
Significantly reduced memory usage (hundreds of MB saved). Community reception overwhelmingly positive.