Grok 4.6
xAI shipped Grok 4.6 on 2026-08-12, framed less as a raw capability jump and more as an agent-and-visual-work release — "stronger first passes on visual and interactive projects" and better self-testing on long-running tasks than Grok 4.5. Pricing holds at $2/MTok in, $6/MTok out (a faster variant at 2x price), with double usage in Grok Build and Cursor for the first week. Artificial Analysis has it tying GPT-5.6 Sol at 61 on their Intelligence Index (Grok 4.5 scored 56); xAI's own release notes claim CursorBench 3.2 69.9% (vs 4.5's 66.7%) and DeepSWE 1.1 65.9% (vs 4.5's 54%) — both xAI-selected comparisons, not yet independently reproduced.
Sources & depth
DeepSeek V4 Pro 0813
DeepSeek quietly moved its Pro-tier API to a new GA build, DeepSeek-V4-Pro-0813 — no blog post, just a docs update and an OpenRouter listing (both fetched this session). It's API-only for now: $0.435/MTok input (cache miss; $0.003625 on a cache hit), $0.87/MTok output, 1M-token context with up to 384K output tokens. DeepSeek's own docs warn "a significant increase" to overall pricing is coming "in the near future" — no open weights yet, unlike the Flash tier's 304B MIT release three weeks ago. Simon Willison flags the benchmark numbers circulating for it trace back through a WeChat group and a since-deleted Reddit post to an HN ASCII table — unverified, treat as rumor, not a claim.
Sources & depth
Qwen3.8-2.4T
Alibaba followed through on the open-weighting it promised three weeks ago: Qwen3.8-2.4T-A95B is now live on Hugging Face and ModelScope, in FP8 and full precision. It's 2.4T total / 95B active parameters across 92 layers, mixing Gated DeltaNet and Gated Attention with a 512-expert MoE (11 active per token) — a genuinely new architecture, not a scaled Qwen3.5 clone. Context is 262,144 tokens natively, extensible to just over a million. Qwen's own numbers: 92.6 on GPQA Diamond, 67.7 on SWE-bench Pro, 86.6 on Terminal Bench 2.1. It's "thinking-only" — every response reasons in <think> tags first, no non-reasoning mode. Licensed under the Qwen3.8-Max license, not a permissive OSI license.
Sources & depth
Ollama speeds up prefill for double-scale NVFP4 models
A small but concrete inference-serving win: Ollama's v0.32.10-rc0 fuses the float32 global-scale multiply and the dtype cast — previously two separate eager ops per projection — into a single compiled kernel. It only helps models whose ModelOpt checkpoints carry a global scale on top of per-group quantization (single-scale NVFP4, MXFP8, and affine checkpoints see no change). On an M5 Max, Ollama's own order-swapped A/B numbers show Qwen3.6:27b prefill going from 703 to 769 tok/s (+7.9%) and Muse-Glimmer:30b from 790 to 843 tok/s (+6.7%); speculative-decode throughput was unaffected on both.
Sources & depth