xAI ships Grok 4.7 on the API.
grok-4.7 — the frontier model id Issue 21 said was missing
xAI ships Grok 4.7 on the API — same list price as 4.6, encrypted reasoning on Responses
grok-4.7 — the frontier model id Issue 21 explicitly noted as absent. 500k context; text+image in / text out; effort ladder includes xhigh.Docs pricing: $2 / $0.50 / $6 per 1M tokens (input / cached input / output) below 200k prompt tokens, and $4 / $1 / $12 above. Announcement starts at $2/$6 in the same class as 4.6. Responses always returns encrypted reasoning content that clients must pass back unchanged.
Fast is Cursor + Grok Build only — not on the public xAI API. Availability spans the API, Cursor, Grok Build, Office add-ins, and gateways; consumer Grok surfaces later. Model card: pretrain cutoff June 2026; supplemental data through August 2026.
Why it matters: operators finally get the API id prior briefings watched for — pin grok-4.7, budget the context-tier price jump, and treat announcement / model-card benches as vendor / partner evals.
Sources: announcement · release notes · model card PDF
None this window. NDCA slowdown suit already covered in Issue 21; no new court primary.
| Model / package | Org | Notes |
|---|---|---|
Grok 4.7 (grok-4.7)Sep 21 |
xAI | API 500k · $2/$6 start · Fast not public |
| MiMo-V2.6 Pro / Flash Sep 21–22 |
Xiaomi | MIT omnimodal MoE · 1M ctx |
| vLLM v0.30.0 ~1:20 AM ET Sep 22 |
vLLM | GA DSv4.1-Flash day-0 · ipc_cache |
| mlx-vlm 0.7.2 ~1:15 PM ET Sep 21 |
MLX | GA Qwen-Image-2.1 · Mistral Large 3 |
| llama.cpp tip b11101 ~8:25 AM ET Sep 22 |
ggml | Tip after Metal FA fix b11093 |
| Snapdragon Hexagon NPU assets b11099 ~3:59 AM ET Sep 22 |
ggml | On-device Linux/Android release binaries |
| Item | Beat | Status |
|---|---|---|
Claude Managed Agents auto permissionsSep 10 |
Server evaluates agent / MCP tool calls | Watch |
| Bedrock KB auto sync + Marengo 3.0 Sep 4 / Sep 11 |
Scheduled connector sync · multimodal embeddings | Active |
| EU AI Board 9th meeting Met Sep 17 · published Sep 18 |
Transparency rules · cyber/AI action plan | Policy |
Critical · Grok vendor evals + Fast not public
Pin API id grok-4.7; treat CursorBench / DeepSWE / Terminal-Bench / HealthBench figures as vendor / partner evals. Do not assume Fast is on the public xAI API — docs limit Fast to Cursor + Grok Build.
Watch · Claude auto permissions / EU transparency
Managed Agents auto moves production harnesses from always-ask toward server-evaluated policy. EU AI Board 9th meeting keeps transparency rules (applicable 2 Aug) and the cyber/AI action plan on the deployer checklist.
Opportunity · vLLM 0.30 · mlx-vlm 0.7.2 · MiMo Distill
Upgrade vLLM for DSv4.1-Flash day-0 + Fast Start cache. On Mac Studio: pip install -U mlx-vlm==0.7.2 and smoke mlx-community MiMo Distill OptiQ-4bit or Qwen-Image-2.1-MLX-4bit. Salesforce×GCP Dreamforce agent/infra (Sep 15) is partnership timing — Hyperforce migrations planned Q4 / NA GA Nov.
Upgrade mlx-vlm 0.7.2; smoke MiMo Distill or Qwen-Image-2.1
pip install -U mlx-vlm==0.7.2 then pull mlx-community/MiMo-V2.6-Distill-Qwen-9B-OptiQ-4bit or mlx-community/Qwen-Image-2.1-MLX-4bit on Mac Studio. Community quants — not Xiaomi-official MLX.
Transformers GGUF path on Apple Silicon
On transformers main + kernels, load a Hub GGUF with gguf_file= and compare decode vs llama.cpp on the same Qwen3.5 quant. MPS packed path only for now.
Issue 22 · Tuesday, September 22, 2026 · Not investment advice · Sources cited inline · Other news: none this window · Issue 21 Anthropic R&D pace / Accenture / Ollama RC / SGLang 0.5.20 / Speakeasy / NDCA suit not re-led
