Frontier reasoning drops to mid tier prices while security shifts to defenders
Claude Sonnet 4.6, Gemini 3.1 Pro and Claude Code Security, and The Deploy Log for this week. Every price and claim read off the publisher's own page, with a call on each.
Anthropic
Claude Sonnet 4.6
DEPLOYif you run agentic coding or computer use
WHAT SHIPPEDClaude Sonnet 4.6 is a full upgrade across coding, computer use, long-context reasoning, agent planning, knowledge work, and design. Pricing remains the same as Sonnet 4.5, starting at $3/$15 per million tokens, with a 1M token context window in beta.
THE CATCHIn Claude Code, users preferred Sonnet 4.6 over Sonnet 4.5 roughly 70% of the time and even preferred it to Opus 4.5, the frontier model from November, 59% of the time. It hit 94% on an insurance benchmark for computer use.
IF YOU RUNa team doing agentic coding at scale or automating legacy software without APIs
SOURCEAnthropic, Introducing Claude Sonnet 4.6
Google DeepMind
Gemini 3.1 Pro
DEPLOYif you pay for Google AI Pro
WHAT SHIPPEDGemini 3.1 Pro is rolling out in preview via the Gemini API in Google AI Studio, Gemini CLI, Google Antigravity, Android Studio, Vertex AI, Gemini Enterprise, the Gemini app, and NotebookLM. On ARC-AGI-2 it achieved a verified score of 77.1%, more than double the reasoning performance of 3 Pro.
THE CATCHThe model generates website-ready animated SVGs directly from text prompts, built in pure code rather than pixels, so they remain crisp at any scale and maintain incredibly small file sizes compared to traditional video.
IF YOU RUNcomplex problem solving where a simple answer is not enough
SOURCEGoogle DeepMind, Gemini 3.1 Pro: A smarter model for your most complex tasks
Anthropic
Claude Code Security
WATCHif you maintain an open source repository
WHAT SHIPPEDClaude Code Security, a new capability built into Claude Code on the web, is now available in a limited research preview. It scans codebases for security vulnerabilities and suggests targeted software patches for human review.
THE CATCHUsing Claude Opus 4.6, Anthropic's team found over 500 vulnerabilities in production open-source codebases, bugs that had gone undetected for decades despite years of expert review. Nothing is applied without human approval.
IF YOU RUNa security team with more vulnerabilities than people to address them
SOURCEAnthropic, Making frontier cybersecurity capabilities available to defenders
The Deploy Log
Models and APIs
| DEPLOYif you make social content or Shorts Lyria 3 music generation. Lyria 3, Google DeepMind's latest generative music model, is rolling out in beta in the Gemini app, letting anyone make 30-second tracks from text or images with custom cover art generated by Nano Banana. Google DeepMind |
| WATCHif you run local inference GGML and llama.cpp join HF. GGML, creators of llama.cpp, are joining Hugging Face, with Georgi Gerganov and team dedicating 100% of their time maintaining llama.cpp with full autonomy and leadership on technical directions and the community. Hugging Face |
| DEPLOYif you fine-tune small models cheaply Unsloth and HF Jobs training. Unsloth and Hugging Face Jobs enable fast LLM fine-tuning of LiquidAI/LFM2.5-1.2B-Instruct through coding agents like Claude Code and Codex, with Unsloth providing approximately 2x faster training and about 60% less VRAM usage compared to standard methods. Hugging Face |
Tools and products
| WATCHif you generate video programmatically AI Gateway video generation. AI Gateway now supports video generation through AI SDK 6, with four initial video models and 17 variations covering text-to-video, image-to-video, first and last frame, and reference-to-video generation types. Vercel |
| WATCHif you produce print-ready marketing assets Recraft V4 on AI Gateway. Recraft V4 is now available on AI Gateway, a text-to-image model built for professional design and marketing use cases with improvements in photorealism, realistic skin, natural textures, and fewer synthetic artifacts. Vercel |
| DEPLOYif you install skills from skills.sh Skills.sh security audits. Skills on skills.sh now have automated security audits from partners Gen, Socket, and Snyk, allowing rapid scaling to audit over 60,000 skills and counting. Vercel |
| DEPLOYif you ingest cost data into FinOps tools Billing usage cost data API. Vercel now supports programmatic access to billing usage and cost data through the API and CLI, with the new /billing/charges endpoint returning data in the FOCUS v1.3 open-standard format. Vercel |
| WATCHif you render markdown in streaming Streamdown 2.3. Streamdown 2.3 ships with a fully cleared bug backlog, applying a unified nested-card design to tables, code blocks, and Mermaid diagrams, with sticky action buttons and plain text code blocks that render immediately. Vercel |
| DEPLOYif you debug with runtime logs Runtime Logs search redesign. The Runtime Logs search bar in the project dashboard has been redesigned with structured filters that parse level:error or status:500 into visual pills, smarter suggestions based on actual log data, and automatic conversion of pasted Vercel Request IDs into filters. Vercel |
| DEPLOYif you export large log datasets Streaming Runtime Logs exports. Runtime Logs exports now stream directly to the browser, with all plans able to export up to 10,000 requests per export and Observability Plus subscribers able to export up to 100,000 requests. Vercel |
| DEPLOYif you vibe code single file apps Gradio gr.HTML one-shot apps. Gradio 6 shipped gr.HTML with custom templates, scoped CSS, and JavaScript interactivity, letting Claude or any other frontier LLM generate frontend, backend, and state management in a single Python file with no build step. Hugging Face |
Left out
- Vercel AI Gateway video generation blog: The page text is a duplicate of the individual video model changelogs and adds no new deployment facts beyond what the Grok Imagine, Wan, Kl
- Vercel Blob private storage: the publisher's page does not show it available yet
- Vercel Sandbox snapshot retention: a figure in the draft could not be found on the publisher's own page
- Notion Library: a figure in the draft could not be found on the publisher's own page
- Vercel Agent build fix suggestions: it did not carry a number that moved
- Veo video models on AI Gateway: its own name for the thing was too long to print
- Kling video models on AI Gateway: its own name for the thing was too long to print
- Wan models on AI Gateway: it did not carry a number that moved
- Grok Imagine Video on AI Gateway: its own name for the thing was too long to print
- Qwen 3.5 Plus on AI Gateway: its own name for the thing was too long to print
- Claude Sonnet 4.6 on AI Gateway: its own name for the thing was too long to print
- Gemini 3.1 Pro on AI Gateway: its own name for the thing was too long to print
The pattern
Claude Sonnet 4.6 at $3/$15 per million tokens beats Opus 4.5 from November 59% of the time, and Gemini 3.1 Pro doubles the reasoning performance of 3 Pro on ARC-AGI-2 with a verified score of 77.1%. Frontier capability is now a mid tier price point, and the same models that produce website-ready animated SVGs and 30-second Lyria 3 tracks are also the ones finding over 500 vulnerabilities in production open-source codebases. The security shift is not a separate product line: Claude Code Security uses Claude Opus 4.6, and skills.sh audits from Gen, Socket, and Snyk scale to over 60,000 skills because the model output is the attack surface.
Vercel, Hugging Face, and Google DeepMind all push the same direction: AI Gateway supports 17 video generation variations, Gradio 6 lets Claude generate frontend, backend, and state management in a single Python file, and Gemini 3.1 Pro ships across Gemini API, Vertex AI, Android Studio, and NotebookLM. The infrastructure now treats model output as code, logs, billing data in FOCUS v1.3, and security patches, all streamed, exported, and audited. GGML joining Hugging Face with 100% of time on llama.cpp and Unsloth providing approximately 2x faster training with about 60% less VRAM usage on LiquidAI/LFM2.5-1.2B-Instruct show the same pattern at the local layer: reasoning, generation, and defense are converging into one deployable stack.