OpenAI Launches GPT-5.4 as Microsoft Defies Pentagon’s Blacklist of Anthropic

Frontier Models & Updates

OpenAI Launches GPT-5.4 and GPT-5.4-Pro: OpenAI has officially released GPT-5.4, a new frontier model optimized for complex reasoning, agent workflows, and computer-use tasks. The release features a 1M-token context window and a revised pricing structure, while the accompanying system card introduces new safety guidelines regarding user emotional reliance on AI.

• BREAKING: OpenAI just drppped GPT-5.4
  • https://openai.com/index/introducing-gpt-5-4
• Difference Between GPT 5.2 and GPT 5.4 on MineBench
  • https://minebench.ai
  • https://github.com/Ammaar-Alam/minebench
• Who the hell is going to pay the 5.4-Pro API prices?
• GPT-5.4'S SYSTEM CARD: OpenAI put "emotional reliance" in the same category as self-harm
• Difference Between GPT 5.2 and GPT 5.4 on MineBench
  • https://i.imgur.com/SPhg3DQ.png

AI Services & Platforms

Perplexity AI Updates Model Lineup with GPT-5.4: Perplexity AI has integrated the new GPT-5.4 Thinking model into its platform for advanced reasoning tasks. Concurrently, the service has removed Grok and Gemini Flash from its available model list, sparking discussions among users regarding the impact on speed and task reliability.

• GPT-5.4 Thinking Available now
• They removed Grok and Gemini Flash?
• Why would you kill Gemini 3 flash?

Open Source & Local LLMs

Qwen 3.5 Performance Benchmarks and Local Deployments: Recent updates to the Qwen 3.5 series include final optimized GGUF releases from Unsloth and successful "agentic" tests on consumer hardware like the M1 Pro. While the models show strong improvements in tool-calling and natural curve generation in 3D environments, users noted a performance gap in complex reasoning compared to top-tier proprietary models.

• Final Qwen3.5 Unsloth GGUF Update!
  • https://huggingface.co/unsloth/Qwen3.5-122B-A10B-GGUF
  • https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF
• Ran Qwen 3.5 9B on M1 Pro (16GB) as an actual agent, not just a chat demo. Honest results.
  • https://thoughts.jock.pl/p/local-llm-macbook-iphone-qwen-experiment
• Qwen3.5B VS the SOTA same size models from 2 years ago.

Industry & Corporate News

Microsoft Commits to Anthropic Despite Pentagon Blacklist: Microsoft has announced it will keep Anthropic's AI models available to its customers, defying the Pentagon's recent designation of the startup as a supply chain risk. Anthropic is planning a legal challenge against the blacklist, while Microsoft's legal team has cleared the models for continued integration into its products.

• Microsoft says Anthropic’s products remain available to customers after Pentagon blacklist

Audio & Transcription Research

Vexa Mitigates Whisper Silence Hallucinations: Developers at Vexa have identified 135 specific phrases that OpenAI's Whisper model commonly hallucinates during periods of silence, such as "Thank you for watching." To solve this, they implemented a multi-layered fix involving Voice Activity Detection (VAD) gating, output blocklists, and greedy decoding.

• We collected 135 phrases Whisper hallucinates during silence — here's what it says when nobody's talking and how we stopped it
  • https://github.com/Vexa-ai/vexa