Small models now carry the agent workload, and the tools are following
GPT-5.4 mini and nano and Forge, 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-5.4 mini and nano
DEPLOYif you run coding assistants or subagents
WHAT SHIPPEDGPT-5.4 mini and nano released. Mini hits SWE-Bench Pro 54.4% and OSWorld-Verified 72.1%, running more than 2x faster than GPT-5 mini. Nano costs $0.20 per 1M input tokens and $1.25 per 1M output tokens.
THE CATCHMini approaches GPT-5.4 on OSWorld-Verified at 72.1% versus 75.0%, but long-context retrieval drops sharply: OpenAI MRCR v2 8-needle 64K-128K is 47.7% against GPT-5.4's 86.0%.
IF YOU RUNhigh-volume coding or computer-use workloads where latency shapes the product experience
SOURCEOpenAI, Introducing GPT-5.4 mini and nano
Mistral
Forge
WATCHif you need models trained on internal knowledge
WHAT SHIPPEDForge introduced as a system for enterprises to build models grounded in proprietary knowledge. It supports pre-training, post-training, and reinforcement learning, plus dense and MoE architectures.
THE CATCHMistral names ASML, DSO National Laboratories Singapore, Ericsson, European Space Agency, Home Team Science and Technology Agency Singapore, and Reply as partners already training on proprietary data.
IF YOU RUNregulated environments where models must reflect compliance requirements and internal governance
SOURCEMistral, Introducing Forge
The Deploy Log
Models and APIs
| WATCHif you run throughput-bound computer-use agents Holotron-12B. Holotron-12B, post-trained from NVIDIA Nemotron-Nano-2 VL, raises WebVoyager from 35.1% to 80.5% and achieves over 2x higher throughput than Holo2-8B on a single H100. Hugging Face |
Tools and products
| DEPLOYif you create executable scripts in sandboxes Sandbox SDK 1.9.0. Vercel Sandbox SDK 1.9.0 lets you set file permissions directly in writeFiles via a mode property, removing the extra chmod round-trip. Vercel |
| DEPLOYif you need fast agentic workflows MiniMax M2.7 on AI Gateway. MiniMax M2.7 is available on Vercel AI Gateway in standard and high-speed variants, with the high-speed variant delivering the same performance for 2x the cost at about 100 tokens per second. Vercel |
| WATCHif you render streaming markdown in React Streamdown 2.5. Streamdown 2.5 adds inline KaTeX support, staggered streaming animations with a default 40ms stagger, and fixes for code blocks, CSV exports, and Tailwind v3 compatibility. Vercel |
Left out
- Mistral Small 4: a figure in the draft could not be found on the publisher's own page
- Custom instructions for AI Meeting Notes: the publisher's page does not show it available yet
- Vercel plugin for coding agents: a figure in the draft could not be found on the publisher's own page
- GPT-5.4 Mini and Nano on AI Gateway: the publisher's page does not show it available yet
- v0 diff view: the publisher's page does not show it available yet
- Notion Academy in 6 languages: the publisher's page does not show it available yet
- Vercel GitHub App permissions: the publisher's page does not show it available yet
- LiteLLM server on Vercel: it did not carry a number that moved
- Workflow encryption: the condition it attaches to the call did not hold up
- Vercel in Stripe Projects: it did not carry a number that moved
- Chat SDK: the condition it attaches to the call did not hold up
- Custom skills for Notion AI: it did not carry a number that moved
The pattern
Small models now carry the agent workload, and the tools are following. GPT-5.4 mini and nano hit OSWorld-Verified 72.1% and SWE-Bench Pro 54.4% at $0.20 per 1M input tokens, while Holotron-12B lifts WebVoyager from 35.1% to 80.5% with over 2x higher throughput than Holo2-8B on a single H100. MiniMax M2.7 on Vercel AI Gateway serves the same performance at about 100 tokens per second for 2x the cost in its high-speed variant.
The infrastructure around these models is tightening the loop: Vercel Sandbox SDK 1.9.0 removes the extra chmod round-trip in writeFiles, Streamdown 2.5 adds inline KaTeX and staggered streaming animations with a default 40ms stagger, and Mistral Forge lets enterprises like ASML, Ericsson, and the European Space Agency pre-train, post-train, and reinforcement learn on proprietary knowledge. The catch is that long-context retrieval still drops sharply, with OpenAI MRCR v2 8-needle 64K-128K at 47.7% for mini against GPT-5.4's 86.0%.