Hey everyone!
It’s been a busy month at Groq- new models, better pricing, SDK upgrades, and community updates galore. Here’s a roundup of everything that dropped in October 
New Model Drop: GPT-OSS-Safeguard 20B
OpenAI’s first safety-focused open-weights model landed on Groq with day-zero support!
Safeguard 20B lets you build customizable Trust & Safety AI: bring your own taxonomy and policies to classify content however you need.
Why it’s awesome:
- 131K token context window and 65K max output tokens
- Prompt caching enabled → 50 % cost savings on cached input tokens
- Full reasoning traces for audits & debugging
- Supports tool use, browser search, code execution, and JSON modes
- Runs at 1000 + t/s on Groq, same price as the base GPT-OSS-20B model
Get started with GPT-OSS-Safeguard 20B
Price Drops + Prompt Caching
Good news for your wallet - GPT-OSS models are now even cheaper and faster on Groq:
Plus, automatic prompt caching is now live for GPT-OSS 120b. Cached hits automatically provide:
- 50 % cost savings on cached input tokens
- Lower latency
- Higher effective rate limits (cached tokens don’t count toward rate limits)
- Zero setup, caching activates automatically when requests share common prefixes
Learn more about prompt caching here
SDKs, Tools, & Ecosystem Updates
-
Python SDK v0.33.1 and TypeScript SDK v0.34.0 now include improved prompt caching support
-
Openbench v0.5.0 added 350+ new evals, new dev experience upgrades, and a plugin system for external benchmarks
-
The Groq API Cookbook now has two new tutorials:
- Using Groq with E2B MCP Gateway
- Using Groq with Box’s MCP
-
Models on Groq are now part of OpenRouter’s exacto tier: a curated set of endpoints that only route to the highest-quality models, selected through rigorous benchmarking and testing on billions of real-world API calls
-
New video on how to add web search to Kimi K2 on Groq, great if you’re building agentic tools
That’s it for last month’s updates! Keep building FAST!