Daily Brief

DeepSeek released V4 Pro at 1.6T total / 49B active parameters, but Artificial Analysis benchmarks show it costs 15x more per token than V3.2 — more expensive than Gemini 3.1 Pro — while failing to reclaim benchmark leadership. Community analysis confirms decreased intelligence density: the model delivers less per token despite 2.5x scale increase over V3.2. The models are optimized for Huawei Ascend chips, signaling DeepSeek's infrastructure pivot away from NVIDIA dependency.

DeepSeek V4 Flash (284B-A13B) matches Gemini 3 Flash on performance benchmarks at 20% of the price, and hands-on testing confirms strong tool-use accuracy and multi-tool calling in large code editing tasks. The efficiency-tier model is where DeepSeek's value proposition holds — developers report it trades some reasoning depth for speed. Flash runs on Huawei Ascend alongside Pro.

Anthropic's engineering postmortem confirms the company reduced reasoning effort and pruned session memory in Claude Sonnet 4.6 and Opus 4.6 to improve latency, degrading model quality for deployed users. After user feedback surfaced the regression, Anthropic reverted the changes. The incident is notable for the transparency of the admission but equally notable for the fact that the tradeoff was made and shipped without user disclosure.

Google has committed $40B to Anthropic, dwarfing Amazon's $5B stake and making the combined external investment in Anthropic the largest single AI company backing on record. The deal sits alongside Google's direct competition with Anthropic via Gemini, raising questions about antitrust exposure and what Google receives in return beyond cloud compute commitments. Full deal terms remain undisclosed.

OpenAI's Romain Huet confirmed that Codex has been folded into the main GPT-5.4 model rather than maintained as a separate coding variant. GPT-5.5 shows measurable gains in agentic coding and computer use, with no standalone Codex product planned going forward. OpenAI also released a GPT-5.5 prompting guide covering multi-step task handling and perceived responsiveness optimization.

llama.cpp now supports NVFP4 (Nvidia E4M3 format) and ik_llama.cpp supports MXFP4 (Microsoft/AMD MX standard), bringing FP4 inference to local hardware. The two formats target different hardware backends. Combined with Qwen3.6-27B hitting ~80 tokens/sec at 218k context on a single RTX 5090 via vLLM 0.19 with NVFP4, FP4 is moving from research to practical local deployment.