Bedrock prompt caching for open-weight models
Open-weight models on Amazon Bedrock now support explicit prompt caching, with cache reads billed at a lower rate than standard input.
The Deploy Log | Week by week | Free, no signup
53 rows are listed for in The Deploy Log, from 24 companies. Amazon Web Services shipped the most, 8 rows, then Anthropic with 7. Listed this week: Bedrock prompt caching for open-weight models, Bedrock tabular search, Nova-3 36 new languages, Qwen4 Max on OpenRouter, Tencent Hy4 flash and 48 more. Every item sourced to the publisher's own page.
| Company | Rows |
|---|---|
| Amazon Web Services | 8 |
| Anthropic | 7 |
| GitHub | 4 |
| OpenAI | 4 |
| 3 | |
| Google Workspace | 3 |
| NVIDIA | 3 |
| OpenRouter | 3 |
| Vercel | 3 |
| Block | 1 |
| Browser Use | 1 |
| Databricks | 1 |
| Deepgram | 1 |
| ElevenLabs | 1 |
| Langfuse | 1 |
| LiteLLM | 1 |
| Microsoft | 1 |
| MLX-VLM | 1 |
| OpenHands | 1 |
| Pydantic | 1 |
| Red Hat | 1 |
| Snowflake | 1 |
| Ultralytics | 1 |
| vLLM | 1 |
| Figure | Value |
|---|---|
| Rows | 53 |
| Rows the week before | 357 |
| Change against the week before | -304 |
| Companies | 24 |
| Companies seen for the first time | 0 |
| Models seen for the first time | 0 |
| Rows with no shipped date, placed by their edition or the day they were added | 0 |
| Undated until its day | 2 |
| Lane | Rows | Rows the week before |
|---|---|---|
| Models and APIs | 9 | 43 |
| Tools and products | 41 | 301 |
| Robotics, hardware and chips | 3 | 6 |
| Breakthroughs | 0 | 0 |
| Funding and acquisitions | 0 | 5 |
Open-weight models on Amazon Bedrock now support explicit prompt caching, with cache reads billed at a lower rate than standard input.
Amazon Bedrock Managed Knowledge Base now supports tabular search for CSV, XLSX, and XLS files.
Nova-3 adds 36 monolingual languages for batch and streaming speech-to-text.
Qwen4 Max is available on OpenRouter at $1.50 input and $4.50 output per 1M tokens with a 1M-token context window.
Tencent Hy4 flash is available on OpenRouter at $0.04375 input and $0.1756 output per 1M tokens with 1M context.
Agent tools as code (tool builder) and unhandled client tool call reporting.
Adds tool search types, static file processing options, and batch inputs for Gemini embedding models.
js-genai v3.4.0 adds tool search types, static file processing options, and batch inputs with usage metadata for Gemini embedding models.
python-genai v3.4.0 released with tool search types, batch inputs for Gemini embeddings, and bug fixes.
Pydantic AI v2.69.0 adds Gemini 4 Hydrogen support on GoogleModel and GroqModel service_tier support.
AWS Professional Services released a deployable solution running Z-Image-Turbo on AWS Inferentia2 via Amazon ECS, with 512px images at $0.003 and 1024px at $0.011.
Private Inference connects your own model endpoints to AI Gateway with the same authentication, SDKs, APIs, and observability as catalog models.
AWS DevOps Agent is now available in AWS GovCloud (US-West).
AgentCore runtime instances are now available in six additional AWS Regions: US West (N. California), Canada (Central), Europe (London), Europe (Paris), Europe (Stockholm), and Asia Pacific (Seoul).
Hybrid search combining vector similarity and full-text search in a single QueryVectors call, available today in all commercial AWS Regions.
New capabilities let users set up, connect, and govern AI agents faster.
Adds source and error codes to RBAC Role read endpoints, Update Bundle API, and owner-only agent toolset files.
Adds source and error codes to RBAC Role read endpoints, Update Bundle API, and Vertex default credential scope.
Adds source and error codes to RBAC Role read endpoints, Update Bundle API, and streaming tool input validation.
You can now start AI Scans for up to 20 repositories at once.
claude-agent-sdk 0.2.183 is available on PyPI.
Claude Code v2.1.326 adds Code Mode with agent teams, startup warnings, MCP loading messages, and many fixes.
Claude Fable 5.2 and Gemini 4 Carbon are available to Copilot Pro+, Max, Business, and Enterprise users.
Gemini app skills can now be shared with View or Edit permissions and include files from Google Drive, Gemini Notebook, and Google Photos.
Gemini side panel in Apps Script is generally available, letting users create, customize, and troubleshoot scripts with plain language.
Enterprise admins can now require selected GitHub Copilot customizations to come from plugins or managed settings.
GitHub MCP Server 2.2.0 adds filtered MCP skills, optional secret hiding, and public agent automation tools.
Databricks provider drops thinking blocks bound to a different prefix.
Tencent's Hy4 Flash model is available on AI Gateway with 50% off tokens through November 29, 2026.
Langfuse v4.81.0 adds Google Vertex AI as a gateway provider and Gemini decision models in evaluators.
LiteLLM v1.108.10 released with backported fixes for OpenAI, Vertex AI, and cost tracking.
MLX-VLM v0.8.0 adds QAT training, distributed streaming and cluster management, and support for MiniMax-H3-Image, Ling 3.1 Flash, Ripple-65B-Preview, DFlash2, DeepSeek V4.1 DSpark, ERNIE 5.5 Flash Base, and Inkling 2.
Google Pics upgraded to Nano Banana 2.1 with higher visual quality, stronger edit fidelity, and sharper text rendering, rolling out from .
ONNX Runtime WebGPU Plugin EP v0.6.0 adds new kernels, FP16 GRU/LSTM, MatMulBlockQuantizedFp8Weight, and shader-cache controls.
OpenAI Java SDK v4.106.0 adds schedule pause reasons and delivers beta Agents session errors as stream events.
OpenHands v1.41.0 adds agent reasoning effort selection, named hooks files, automation setup analytics, and grouped automation dashboard.
OpenRouter routing now powers Hugging Face Inference Providers, with 160 models from 38 labs available today.
OpenShift Service Mesh 3.5 is generally available with Istio 1.31, Kiali 2.33, and the Kubernetes MCP server.
KV cache-aware routing for real-time inference endpoints is available today in all AWS Regions where SageMaker AI is available.
ByteDance's Seedance 2.5 Flash is available on AI Gateway, generating videos with synchronized audio from text or references.
Snowflake AI & Data Sharing now lets providers share Apache Iceberg tables with any consumers using Iceberg REST Catalog-compatible clients.
Segmentation masks can now be plotted, converted, and prompted without building a dense image-sized mask stack.
Providers can add Unity Catalog agent skills to OpenSharing shares and Databricks Marketplace listings.
v2.9.0 adds source and error codes to RBAC Role read endpoints and Update Bundle.
browser-use 1.0.6 makes BU 4.0 the default model, with bu-4.0-mini and bu-4.0-pro options, and tokens cost $1.80/$9.00 per 1M input/output.
Claude Agent SDK TypeScript v0.3.326 adds mcpServersDelivery initialize option, getSessionInfo(), and message_uuid in cancel responses.
OpenAI Go SDK v3.100.0 adds MCP approvals, Live transport events, schedule pause reasons, and beta Agents session errors as stream events.
Adds schedule pause reasons and delivers beta Agents session errors as stream events.
Adds schedule pause reasons and delivers beta Agents session errors as stream events.
vLLM v0.34.0 ships 10 new models, extensible KV cache on by default, and Model Runner V2 as the only GPU runner.
New CUDA-X libraries cuPhoton, cuPotential and cuSpa are available, with cuPhoton.xBeam cutting black hole correlation from over a month to eight minutes.
The Vera CPU rack-scale platform combines 192 Vera CPUs with Quantum-X800 InfiniBand, delivering over 200TB/s aggregate peak memory bandwidth.
Nemotron for Science brings Nemotron 3.5 Super VL, LitmusBench v0.2 with nearly 50,000 chemistry questions, and post-training recipes for science agents.
One line every time AI actually ships, around the clock, free on Threads and Bluesky. Every one is in The Deploy Log with its source.