Safety classifiers and agent packaging both moved from custom to shared
GPT-5.6 Sol, Shieldstral and Grok Imagine Image 2.0 Preview, and The Deploy Log for this week. Every price and claim read off the publisher's own page, with a call on each.
OpenAI
GPT-5.6 Sol
DEPLOYif you pay for Plus or Pro
WHAT SHIPPEDGPT-5.6 Sol in ChatGPT now gives more focused answers and makes factual errors 68% less common than GPT-5.5 Instant on financial, medical, and legal prompts. A new slider controls how much thought each response gets.
THE CATCHThe updated Sol only runs in the Chat experience. The version powering Work and Codex is unchanged, so coding and work products do not get these gains.
IF YOU RUNa support or research workflow where factual errors on dates, numbers, or rules cost you trust
SOURCEOpenAI, Improving GPT-5.6 Sol in ChatGPT, and expanding access to GPT-5.6 Luna for free users
Mistral
Shieldstral
DEPLOYif you moderate text and images in production
WHAT SHIPPEDShieldstral is a 3B open-weights multimodal safety classifier under Apache 2.0 that matches or outperforms open guard models up to 7x its size on text safety, refusal detection, policy adaptability, and multimodal benchmarks. It runs on a single 16GB NVIDIA GPU.
THE CATCHYou write the policy as a plain-language question at inference time and get a calibrated yes/no probability from a single forward pass. No retraining when your safety definitions change.
IF YOU RUNa product where the same content is fine for one audience and harmful for another
SOURCEMistral, Introducing Shieldstral
Vercel
Grok Imagine Image 2.0 Preview
WATCHif you generate infographics or posters programmatically
WHAT SHIPPEDGrok Imagine Image 2.0 Preview from xAI is available on AI Gateway. It follows detailed instructions closely and plans typography and layout together, so dense visuals like infographics, posters, and title screens hold their structure and small text stays legible.
THE CATCHIt also supports image editing, keeping subjects and details consistent across repeated generations. Set model to xai/grok-imagine-image-2.0-preview and call generateImage from the AI SDK.
IF YOU RUNa pipeline that needs multi-part visuals with legible small text at 1k or 2k resolution
SOURCEVercel, Grok Imagine Image 2.0 Preview from xAI is now available on AI Gateway
The Deploy Log
Models and APIs
| DEPLOYif you call open-weight LLMs through the Hub Baseten Inference Provider. Baseten is now a supported Inference Provider on the Hugging Face Hub, launching support for conversational and text-generation tasks with open-weight LLMs such as Kimi K3, DeepSeek V4 Flash, and GLM-5.2. Hugging Face |
| DEPLOYif your team asks health or lab questions Fable 5 biology safeguards. Anthropic reduced biology-related fallbacks on Claude Fable 5 by about 85% across product surfaces, so users see far fewer switches to Opus 5 on everyday health and educational questions. Anthropic |
Tools and products
| DEPLOYif you author agent extensions for multiple clients Agent Plugins 1.0.0. Agent Plugins 1.0.0 is publicly available as an open, vendor-neutral standard for packaging Agent Skills and MCP servers into portable plugins, supported at launch across ChatGPT and Codex, Cursor, GitHub Copilot, Kiro, and VS Code. Vercel |
| DEPLOYif you embed app generation in your product v0 API. The new v0 API is generally available, giving programmatic headless access to v0's app-building agent that generates an app, starts a dev server in a Vercel Sandbox, and returns a preview URL you can embed. Vercel |
| WATCHif you run large sandbox fleets Vercel Sandbox quotas. Vercel Sandbox default quotas on Pro and Enterprise plans rose from 2,000 to 10,000 concurrent sandboxes and from 200 or 400 vCPUs per minute to a dynamic rate up to 5,000 vCPUs per minute. Vercel |
Left out
- Agent Plugins 1.0.0 changelog: Duplicate of the Agent Plugins announcement lead with no new facts.
- AI Gateway on AWS Marketplace: the publisher's page does not show it available yet
- Chat SDK durable approvals: a figure in the draft could not be found on the publisher's own page
- Skill packs: the publisher's page does not show it available yet
- Free domain with Pro: a figure in the draft could not be found on the publisher's own page
- agent-browser for eve: it did not carry a number that moved
- eve iMessage channel: it did not carry a number that moved
- Sandbox egress firewall on Hobby: it did not carry a number that moved
- Vercel WAF for Blob: it did not carry a number that moved
- Seedance 2.5: the condition it attaches to the call did not hold up
The pattern
This week the same move showed up in two very different places: safety and packaging both stopped being custom work. Anthropic cut biology fallbacks by 85% by retraining a classifier, and Mistral shipped a 3B classifier that takes your policy as a plain-language question at inference time instead of baking a taxonomy into the weights.
On the packaging side, Agent Plugins 1.0.0 gives Skills and MCP servers one portable manifest across ChatGPT, Codex, Cursor, GitHub Copilot, Kiro, and VS Code, while Vercel pushed sandbox quotas to 10,000 concurrent instances and 5,000 vCPUs per minute. The deployment lens says the same thing in both lanes: the parts you used to rebuild for every client or every policy are now shared, and the number that moved is how much less you rebuild.