AI Engineering Digest

AWS opens GPT-6 Astra UltraFast on Bedrock.

Thursday, October 1, 2026
Issue 29 · sourced briefing
Sources: AWS · Anthropic · llama.cpp · Hugging Face · Databricks
Lead Story

GPT-6 Astra UltraFast on Amazon Bedrock

On Tue Sep 30, AWS What’s New posted UltraFast mode for GPT-6 Astra on Amazon Bedrock — a premium speed tier for latency-sensitive apps (real-time coding assistants, interactive agents, customer-facing experiences). This is the Bedrock availability ship, distinct from Issue 28’s OpenAI API Ultrafast docs card (service_tier: ultrafast).

AWS cites OpenAI: up to 6× faster inference in the API and up to ~300 tokens/second — both vendor / OpenAI-attributed, not independent digest benches. Start via the Bedrock console or supported Bedrock APIs; regions, endpoints, inference profiles, and pricing live in Bedrock docs.

Why it matters after Issue 28: yesterday covered Astra Ultrafast as API docs; today is the AWS perimeter story — same speed tier, callable under Bedrock identity and billing. Do not claim Sol Ultrafast GA (OpenAI Ultrafast guide still treats Sol Ultrafast as preview / not broadly GA). Do not re-lead BMA, Sonnet 5.5, or DevDay/Sol.

Sources: AWS What’s New · OpenAI Ultrafast docs (context · Issue 28)

By the Numbers
Sep 30
UltraFast Bedrock · GPT-6 Astra
Nov 30
Claude Sonnet 4.5 retire · migrate 5.5
b11320
llama.cpp tip head · Metal b11316
5.18.0
Transformers · PEFT 0.21.2 compat

Quiet lanes

BMA / HyperPod / zero-trust Pt 2 / Sonnet 5.5 launch / Sol / DevDay / OpenShell not re-led · Sol Ultrafast not GA · Ollama still 0.35.1-rc0 · MLX 0.32.3 / mlx-vlm 0.7.4 Watch · Haiku 5.5 not invented · Tencent: no new OSS model


Commentary

Past 1–2 days. ASTGL / Fraser takes — sourced, not recap theater.

Bedrock UltraFast — Astra speed under the AWS perimeter

Issue 28 put Ultrafast on the OpenAI API docs shelf. Sep 30’s What’s New is the operator delta: the same Astra speed tier, now a Bedrock SKU you can wire into existing IAM, regions, and spend controls. Treat the “up to 6× / ~300 tok/s” line as OpenAI-attributed marketing until you measure your own path. And no — this is not Sol Ultrafast GA.

Sources: AWS WN · Ultrafast docs

Sonnet 4.5 deprecated Sep 30 — retire Nov 30

Anthropic marked Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) deprecated on Sep 30 with Claude API retirement Nov 30, 2026 — migrate to Sonnet 5.5 (Issue 27 launch, not today’s lead). Partner clouds set their own clocks: Bedrock and Google Cloud may retire on different dates. Monday action is Usage export + id inventory, not a re-read of the 5.5 launch post.

Sources: Platform RN · Deprecations

llama.cpp Oct 1 tip cascade — Metal mxfp4

Overnight tip churn after Issue 28’s GLM tip: head is b11320 (~7:20 AM ET Oct 1). The Apple Silicon story is b11316 — Metal bf16 math for mxfp4 mul-mat (#29770), with macOS arm64 assets. Tip ≠ GA v0.5.0. Smoke the Metal path if you already live on tips; keep the GA pin for rollback.

Sources: b11316 · b11320

Transformers 5.18.0 — ExecuTorch MLX in the RN

Hugging Face shipped Transformers 5.18.0 (Sep 30) with Nemotron 3 Diarization, NemotronH Omni, HyperCLOVAX Vision V2, and GTE encoders — plus an ExecuTorch MLX recipe and MPS/FP8 notes in the same release. PEFT 0.21.2 (Oct 1) is the compat patch so encoder-decoder stacks do not break under ≥5.18. Pair the bump; do not rehash 0.21.0’s new methods as today’s news.

Sources: Transformers 5.18.0 · PEFT 0.21.2

UK London — Opus 5.5 / Sonnet 5 in-region

AWS also put Claude Opus 5.5 and Claude Sonnet 5 on Bedrock in-region in UK (London) / eu-west-2 (Sep 30 WN). Residency story for FS / healthcare / public sector — AWS says processing stays in the Region you call. Do not conflate Sonnet 5 here with Sonnet 5.5 (Issue 27) or invent Haiku 5.5.

Source: AWS WN London


Today’s Stories
Local

llama.cpp tip b11316 Metal mxfp4 / tip head b11320

Oct 1: tip cascade after Issue 28’s b11279 GLM card. Apple-relevant b11316 (~4:58 AM ET): metal : use bf16 math for mxfp4 mul-mat (#29770); macOS arm64 assets. Tip head b11320 (~7:20 AM ET). Also same-day: b11313 re-enables -sm tensor for qwen4exp. Tip / pre-release — GA still v0.5.0. Do not rehash GLM-5.3-Flash as today’s lead.

Sources: b11316 · b11320

ML practice

Transformers 5.18.0 (+ PEFT 0.21.2)

Sep 30 GA: new models (Nemotron 3 Diarization, NemotronH Omni, HyperCLOVAX Vision V2, GTE) plus ExecuTorch MLX recipe (#48910), compressed-tensors FP8 on MPS (torch ≥2.15), PEFT×DTensor TP note, MPS maintainer. Several breaking changes in the RN. PEFT 0.21.2 (Oct 1) fixes encoder-decoder under Transformers ≥5.18 — bump both; do not rehash PEFT 0.21.0 methods.

Sources: v5.18.0 · PEFT 0.21.2

MLX

mlx-community K2-Horizon quants

Sep 29–30 Hub drops: dense 0.9B / 3.7B / 7B (bf16 / 8-bit / 4-bit) and MoVA 36B-A4B. Example 7B-4bit: mixed-precision 4-bit, Apache-2.0 from IFM base, 512K context claimed on card. Caveat: mlx-lm has no native k2_horizon arch yet (open request #1876) — ships k2_horizon.py via model_file + trust_remote_code. Tok/s / PPL on M4 Max are uploader-reported.

Source: K2-Horizon-7B-4bit

Frontier

Claude Sonnet 4.5 deprecated → Sonnet 5.5

Sep 30 Claude Platform RN: Claude Sonnet 4.5 deprecated; Claude API retirement Nov 30, 2026; migrate to Claude Sonnet 5.5. Deprecation dates apply to Anthropic-operated platforms (Claude API, Claude Platform on AWS, Microsoft Foundry); Bedrock / Google Cloud set their own schedules. Console Usage export remains the audit path. Do not re-lead the Sonnet 5.5 launch (Issue 27).

Sources: RN · Deprecations

Enterprise

Claude Opus 5.5 + Sonnet 5 — UK London in-region

Sep 30 What’s New: in-region support for Claude Opus 5.5 and Claude Sonnet 5 on Bedrock in UK (London) / eu-west-2. Aimed at FS / healthcare / public sector needing UK data processing; inference on bedrock-runtime in London; AWS states processing never leaves the Region called. Do not conflate Sonnet 5 with Sonnet 5.5.

Source: AWS WN

Enterprise

Databricks ai_decide (Beta)

Sep 30 blog: new AI Function ai_decide for fast structured decisions over governed data (routing, tagging, agent evals) — probability / named choice / ordered-scale score. Positioned for lower latency/cost than LLMs on decision (not generation) tasks; powered by a decision model (TypeSafe AI’s Jev mentioned). Available in Beta via SQL and REST. Latency/cost claims are vendor; Beta ≠ GA.

Source: Databricks blog


Other news in AI

None this window.


Signal Tables
Model / packageOrgNotes
GPT-6 Astra UltraFast
Sep 30 · Bedrock
AWS / OpenAI New Bedrock speed tier · vendor 6× / ~300 tok/s · Sol Ultrafast not GA
llama.cpp tip b11320
Oct 1 · Metal b11316
ggml Tip mxfp4 mul-mat Metal · GA still v0.5.0
Transformers 5.18.0
Sep 30
Hugging Face GA new models · ExecuTorch MLX recipe · MPS notes
PEFT 0.21.2
Oct 1
Hugging Face GA encoder-decoder compat · Transformers ≥5.18
K2-Horizon MLX quants
Sep 29–30 · Hub
mlx-community Watch trust_remote_code · no native mlx-lm arch
Ollama v0.35.1-rc0
unchanged · Issue 28
Ollama Watch still pre-release · no GA 0.35.1 tag
MLX / mlx-vlm
unchanged overnight
Apple ML / Blaizzy Watch still 0.32.3 / 0.7.4
ItemBeatStatus
Sonnet 4.5 deprecation
Sep 30 → retire Nov 30
Migrate to Sonnet 5.5 · partner clouds own schedules Watch
UK London Claude residency
Sep 30 · eu-west-2
Opus 5.5 + Sonnet 5 in-region on Bedrock Active
NIST NCCoE Agentic webinar
Oct 28 · comments Nov 9
DevSecOps / agentic AI Build 3 scoping Watch

Industry Shorts

Critical · tip vs pin · deprecation Nov 30

llama.cpp tip b11316/b11320 ≠ GA v0.5.0. Claude Sonnet 4.5 retires on the Claude API Nov 30 — inventory ids now; Bedrock/GCP may differ.

Watch · vendor UltraFast speed · Databricks Beta

Bedrock UltraFast 6× / ~300 tok/s is OpenAI-attributed. Databricks ai_decide is Beta — not a GA decision engine. Sol Ultrafast still not GA.

Opportunity · Metal mxfp4 tip + Sonnet 4.5 export

Smoke tip ≥b11316 on Mac for the mxfp4 Metal path. Export Claude Usage for Sonnet 4.5 traffic and stage the Sonnet 5.5 cutover before Nov 30.


Try This

Smoke llama.cpp tip ≥b11316 on Metal mxfp4

Build or grab tip b11316 or newer (head b11320). On Apple Silicon, load an mxfp4 (or related) GGUF and confirm the Metal mul-mat path runs. Keep GA v0.5.0 as the rollback pin. Tip ≠ production.

llama.cpp b11316 Metal

Grep Sonnet 4.5 traffic — or try Bedrock Astra UltraFast

Export Claude Usage / configs for claude-sonnet-4-5 ids and stage migrate to Sonnet 5.5 before Nov 30 (confirm Bedrock/GCP dates separately). Or open Bedrock console and smoke GPT-6 Astra UltraFast on a latency-sensitive path — measure your own tok/s; do not trust the vendor 6× line blind.

Sonnet 4.5 UltraFast Bedrock

Issue 29 · Thursday, October 1, 2026 · Not investment advice · Sources cited inline · Other news: none this window · Issue 28 BMA / HyperPod / zero-trust Pt 2 / Astra API Ultrafast docs / llama.cpp b11279 GLM / Ollama 0.35.1-rc0 / PEFT 0.21.0–0.21.1 not re-led · Issue 27 Sonnet 5.5 / GPT-6.1 Sol / DevDay not re-led · Issue 26 OpenShell not re-led · Do not invent Haiku 5.5 GA · Sol Ultrafast not GA · Listen omitted (episode not published)