Grok 4.6 on AI Gateway, DeepSeek V4 Pro update, and new Replicate agent models
TL;DR
Vendors pushed model updates and agent infrastructure while device makers released new hardware with built-in AI features across the day.
What shipped
On 12 August 2026, AI Gateway added Grok 4.6 and updated DeepSeek weights while Google shipped Pixel 11 devices and several new models appeared on Replicate and Hugging Face. Enterprise and open-weight releases mixed with hardware announcements and cost data from production traffic.
Vendor launches
Google released the Pixel 11 series with new camera hardware, Tensor G6 chip, and Gemini features across multiple posts. Vercel added Grok 4.6 and updated DeepSeek V4 Pro weights to AI Gateway while also publishing internal engineering updates on SDK maintenance and build infrastructure. NVIDIA reported employee rankings and new financing partnerships for AI factories.
- •Exa web search on AI Gateway Exa search became free through August 31 on AI Gateway and now serves as the default for eve agents without needing a separate key.
- •Pixel 11 helpful updates Google added Magic Capture, Rambler, and HiLight tools to Pixel 11 phones for everyday tasks like note taking and quick captures.
- •AI SDK software factory Vercel described its process for handling four moving targets in the AI SDK that sees 20 million weekly downloads.
- •DeepSeek V4 Pro weights Updated DeepSeek V4 Pro weights now run by default on AI Gateway and can be called via deepseek/deepseek-v4-pro-813.
- •Grok 4.6 on AI Gateway Grok 4.6 with 500K context and image input is now available on AI Gateway for coding agents and direct calls.
- •Pixel 11 Pro Fold Google launched the Pixel 11 Pro Fold with a thinner design, new main camera, and enhanced intelligence tools.
- •Pixel Buds updates Google released new colors and upgraded features for Pixel Buds Pro 2 and Pixel Buds 2a.
- •Pixel Watch 5 Google introduced Pixel Watch 5 with proactive Gemini alerts and advanced health tracking features.
- •Pixel 11 series launch The full Pixel 11 lineup includes updated cameras, Tensor G6 chip, and faster Gemini support.
- •Pixel 11 feature roundup Google highlighted biggest camera upgrades in years plus Gemini Intelligence across the new Pixel 11 devices.
- •Gemini connected services Gemini now connects directly to more apps for trip planning, project management, and daily tasks.
- •Sign language AI model DeepMind released sign-language-to-text model SL2T for new features aimed at Deaf and hard of hearing users.
- •DeepSeek volume overtake DeepSeek passed Google in production volume on AI Gateway while average price per token fell 13.6 percent.
Hugging Face trending
Three models trended on Hugging Face Hub with GGUF and transformers builds ready for download and inference. The releases cover image-to-video and large text-generation workloads.
- •MiniMax-H3-GGUF model unsloth released MiniMax-H3-GGUF, an image-text-to-video model built with gguf and available for fine-tuning on the Hub.
- •Qwen3.8-2.4T-A95B model Qwen released Qwen3.8-2.4T-A95B, a text-generation model built with transformers and ready for inference on the Hub.
- •NVIDIA-Nemotron model NVIDIA released NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4, a text-generation model built with transformers and available on the Hub.
Replicate new models
Replicate added five new models including video frame tools, media utilities, and two frontier models for direct API calls. The agent-focused releases target coding and image tasks.
- •video-frame-custom tool littletry launched video-frame-custom on Replicate to extract specific frames from video files via HTTP API.
- •media-metadata tool littletry launched media-metadata on Replicate to read metadata from media files with direct API access.
- •compress-image tool littletry launched compress-image on Replicate to reduce image size while controlling quality and EXIF data.
- •claude-sonnet-5 model Anthropic released claude-sonnet-5 on Replicate for frontier coding and tool use at lower cost than larger models.
- •z-image-turbo model prunaai released z-image-turbo on Replicate, a 6B parameter text-to-image model for fast generation.
Product Hunt picks
Six tools launched on Product Hunt focused on research context, spend tracking, local training, agent control planes, shared folders, and agentic app building.
- •Click tool Click provides live research context inside ChatGPT and Claude sessions for faster information gathering.
- •CodeBurn tool CodeBurn tracks where AI coding spend occurs across projects and teams.
- •Unsloth Desktop Unsloth Desktop lets users run and train AI models locally on desktop hardware.
- •Cohesor tool Cohesor offers a neutral control plane for managing enterprise AI agents across vendors.
- •BearDrive tool BearDrive supplies an open-source shared folder system for team AI agents.
- •LaraCopilot tool LaraCopilot acts as an agentic AI engineer that builds complete applications.
Other
Eleven updates covered real estate video tools, cost attribution on Bedrock, AI contributor guidelines, LangSmith on AWS, new Qwen availability on OpenRouter, and infrastructure financing at Databricks and Lambda.
- •HeyGen for Real Estate HeyGen introduced a new offering tailored for real estate video creation and marketing.
- •Amazon Bedrock cost attribution AWS published part two of its guide on attributing Bedrock costs using Athena and CUDOS.
- •AI-first contributors guide GitHub shared repo instructions for handling AI-generated pull requests while keeping maintainers in control.
- •LangSmith BYOC on AWS LangChain made LangSmith Bring Your Own Cloud generally available on AWS for enterprise observability.
- •Qwen3.8 on OpenRouter Qwen3.8 2.4T A95B launched on OpenRouter with 262k context at $2.00 per million input tokens.
- •Qwen3.8 OpenRouter addition Qwen3.8 2.4T A95B became available on OpenRouter for testing with existing API setups.
Industry news
Industry coverage included open-source debates from Hinton, Li, and Ng, voice-focused AI wearables, data requirements for scaling agents, prompt reverse-engineering risks, breast cancer AI shortfalls, Grok 4.6 benchmarks, and Twitch content training defaults.
- •AI safety open-source debate Hinton, Li, and Ng argued at Ai4 for keeping models open despite rising safety concerns.
- •Sandbar voice ring Sandbar promoted its voice-enabled ring as the next direction for AI wearables beyond note-taking devices.
- •Sandbar hardware strategy Sandbar explained how its ring design aims to avoid the AI hardware failure pattern seen in prior products.
- •Trustworthy data for agents MIT Technology Review examined why clean data infrastructure is required to achieve ROI from agent deployments.
- •LLM prompt reverse-engineering IIT Bombay and Adobe researchers built a method to reconstruct prompts from model output with near-perfect accuracy.
- •Breast cancer AI tools gap A survey of 215 radiologists found AI detection tools missed expected recall rate reductions in real use.
- •Grok 4.6 benchmark results Grok 4.6 tied GPT-5.6 on intelligence index scores while using 53 steps versus 103 on agentic tasks at lower price.
- •Twitch content training default Amazon set Twitch streamer content to train AI models by default with an opt-out option.
What this means for you
For Vibe Builders: You can now route coding agents to Grok 4.6 or updated DeepSeek weights through AI Gateway without extra keys. Replicate added claude-sonnet-5 and z-image-turbo for direct calls while Product Hunt tools like Click and LaraCopilot give ready-made research and build flows. Test one new model or agent tool this week against your current stack.
For Non-techies: Pixel 11 phones and watches bring Gemini features for photos, health tracking, and connected apps without setup. HeyGen launched real estate video tools and Gemini now links to more services for daily tasks. Try the new Pixel hardware or one connected Gemini service to see if it fits your routine.
For Developers: AI Gateway added Grok 4.6 with 500K context and DeepSeek V4 Pro weights while Replicate hosts claude-sonnet-5 for production calls. Hugging Face trending models and OpenRouter Qwen3.8 give open-weight options with known pricing. Benchmark Grok 4.6 against your current agent setup before wider rollout.
What to watch next
Watch for more agent runtime benchmarks on the Artificial Analysis index and any follow-up pricing moves from xAI or Anthropic. Track adoption signals from the new Replicate models and Product Hunt agent tools over the next seven days.
Harsh’s take
The day showed clear separation between infrastructure updates that lower token costs and hardware releases that still require users to buy new devices. Production traffic data revealed DeepSeek gaining share while Google pushed Pixel features that may see limited daily use. The prompt reconstruction research highlights a real security gap that many teams have not addressed.
Vendors continue to ship models and tools faster than most builders can evaluate them. The practical move is to pick one new model or agent tool from the releases and run a short test against existing workflows this week rather than waiting for more announcements.
by Harsh Desai
Sources
Vendor launches
- •Exa web search free through August 31 on AI Gateway and eve
- •Here are 3 ways your Pixel just got even more helpful.
- •Free domain for one year, now for all Pro teams
- •Building a software factory for AI SDK
- •DeepSeek V4 Pro now runs updated weights on AI Gateway
- •NVIDIA CEO Tops Glassdoor’s 2026 List of Best CEOs
- •How we migrated the database behind every Vercel build
- •Grok 4.6 now available on AI Gateway
- •Google’s most sophisticated foldable: Pixel 11 Pro Fold
- •Here’s what’s new with Pixel Buds Pro 2 and Pixel Buds 2a.
- •Pixel Watch 5: Proactive assistance and advanced health tracking on your wrist
- •The Pixel 11 series: Your most personal Pixel yet
- •7 can’t-miss updates from our Pixel 11 launch
- •Now you can connect even more of your favorite apps and services to Gemini.
- •Putting sign language AI into users’ hands
- •NVIDIA AI Factory Compute Is Becoming an Investable Asset Class
- •DeepSeek overtakes Google on volume, cost per token falls 13.6%
Hugging Face trending
- •MiniMax-H3-GGUF by unsloth trends on HuggingFace
- •Qwen3.8-2.4T-A95B by Qwen trends on HuggingFace
- •NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 by nvidia trends on HuggingFace
Replicate new models
- •video-frame-custom by littletry launches on Replicate
- •media-metadata by littletry launches on Replicate
- •compress-image by littletry launches on Replicate
- •claude-sonnet-5 dropped on Replicate today
- •z-image-turbo dropped on Replicate today
Product Hunt picks
Other
- •UpdatesPublishedAugust 11th, 2026Introducing HeyGen for Real Estate
- •Part 2: Amazon Bedrock cost attribution with Amazon Athena and CUDOS
- •Your contributors are AI-first now. Is your project?
- •LangSmith BYOC on AWS is generally available
- •Qwen: Qwen3.8 2.4T A95B now available on OpenRouter (262k context, $2.00/M in, $6.00/M out)
- •Databricks Network Configuration delivery to Tens of Millions of Serverless VMs
- •How Amtrak is building the data backbone for its largest transformation in over 50 years
- •How a major freight railroad scaled pipeline creation with Genie Code
- •Why brand distinctiveness is your strongest moat in the AI era: Colin Kelton’s framework from 36 years at Vanguard
- •Lambda prices $926 million senior secured term loan B facility, the first investment-grade-rated term loan B financing by a private neocloud
Industry news
- •As AI safety concerns mount, three pioneers make the case for staying open
- •Why Stream ring-maker Sandbar says the future of AI wearables is voice
- •Why Sandbar thinks it’s voice-enabled ring can avoid the AI hardware graveyard
- •Scaling AI agents with trustworthy data
- •Researchers can now reverse-engineer LLM prompts from output text with near-perfect accuracy
- •AI tools for breast cancer detection fall short of radiologists' expectations
- •SpaceXAI's Grok 4.6 matches OpenAI's best model and undercuts it on price
- •Amazon will train on Twitch streamers’ content by default, unless they opt out
More AI news
- FeatureLovable generates Trust Center security pages for published apps
Lovable automatically generates Trust Center security pages and machine-readable JSON twins for publicly published apps at /.well-known/trust.html and trust.json.
- FeatureClaude Code adds self-hosted runners, archive plugins, and cross-session messaging
Claude Code introduces self-hosted runners for web, mobile, and desktop. It adds zip archive plugin sources, cross-session SendMessage capabilities, and improved sandbox credential-masking and Remote Control session sync.
- FeatureChatGPT Voice adds file uploads and Projects access in conversations
ChatGPT Voice now supports file uploads during voice conversations and integrates Project context, instructions, and recent chats.