Groq stopped selling speed and started selling an agent operating system

The changelog tells the story better than any press release. Between September 4 and September 23, 2025, Groq added three major capabilities: a remote Model Context Protocol (MCP) server integration in beta, Moonshot AI's Kimi K2-0905 instruct model, and the Compound agentic systems that bundle web search, code execution, and browser automation into a single API call. By volume alone it looks like a product blitz. Look closer and the pattern is coherent: Groq is no longer selling inference. It is selling a complete agentic operating system for the API era.Parallel agents aren't about speed. They're about…

The MCP launch is the most consequential of the three. Remote MCP, short for Model Context Protocol, Anthropic's open standard for connecting AI models to external tools, lets developers attach any MCP-compliant server to Groq-hosted models. The integration is fully compatible with OpenAI's Responses API and its own remote MCP specification. The practical effect is that any developer currently using OpenAI's tool-calling infrastructure can migrate to Groq without touching a single line of connector code. They change the endpoint, keep the tool definitions, and pay less per token. In announcing the feature, Groq highlighted seven launch partners, BrowserBase, Browser Use, Exa, Firecrawl, HuggingFace, Parallel, Stripe, and Tavily, each of which provides a ready-made MCP server for web automation, search, scraping, or payment workflows.Anthropic Academy: a free platform to learn about Claude

If MCP is the universal adapter, the Kimi K2-0905 release gives developers a model built to exploit it. Moonshot AI's latest variant arrives on GroqCloud with day-zero support, a 256K context window, the largest on the platform to date, and prompt caching that cuts costs by up to 50% on cached prefixes. The model's performance numbers are competitive: 200+ tokens per second at a blended price of $1.50 per million tokens. More importantly, Groq positions it as an agentic coding model, claiming improved reliability on complex multi-turn interactions. That fits the MCP story precisely: a model that can handle multiple tool calls across a long context window is exactly the kind of engine that makes remote MCP useful.Kimi K2.7 Code is faster and cheaper. But open-source…

The Compound and Compound Mini systems, promoted from beta to general availability on September 4, abstract away the tool orchestration entirely. Built on OpenAI's GPT-OSS-120B and Meta's Llama models, they integrate web search, code execution, Wolfram Alpha, and parallel browser automation, up to ten simultaneous browsers, into a single API call. Groq claims a roughly 25% higher accuracy and 50% fewer mistakes compared to OpenAI's Web Search Preview and Perplexity Sonar on internal benchmarks. Compound is not a model; it is a turnkey agent. Developers who do not want to assemble their own toolchain using MCP can simply call groq/compound and get a research assistant that searches, computes, and navigates pages automatically.IBM's new open-source agent framework cuts the…

The model roster continues to deepen. OpenAI's open-weight GPT-OSS-20B and GPT-OSS-120B, released in early August, bring reasoning capabilities, built-in browser search, and structured outputs. Qwen 3 32B, added in June, supports a thinking mode that can be toggled off for low-latency responses. Meta's Llama 4 Scout and Maverick models arrived in April with vision understanding. Groq is deliberately model-agnostic. The play is not to bet on one foundation model but to host all of them, and to provide the glue that makes them useful.

Prompt caching, rolled out in late August, automates cost optimization. The system reuses computation from recent requests that share a common prefix, system prompts, tool definitions, few-shot examples, and applies a 50% discount on cached tokens. Cached data lives in volatile memory and expires within hours, which sidesteps the privacy concerns that dog persistent caches. The feature currently works with Kimi K2, with other models promised soon.

The strategic picture is clearer with each entry in the changelog. The inference-as-a-service market has become a commodity race where providers compete on price-per-token and model roster. Groq is trying to exit that race by becoming the orchestration layer. Remote MCP gives developers a standardized way to attach tools. Compound gives them a pre-assembled agent. Kimi K2 and GPT-OSS give them models with enough context and tooling support to use these capabilities at scale. The pricing remains competitive, $1.50 per million tokens for Kimi K2, $0.29/$0.59 for Qwen 3, but the differentiation is no longer about the unit economics of a single inference call. It is about the total cost of building an agentic application, start to finish.

There are risks. MCP is still a beta standard, and Anthropic's own roadmap for the protocol may shift in ways that complicate Groq's integration. The Compound performance numbers come from Groq's own benchmarks, not independent audits, and real-world agentic reliability tends to degrade under edge cases that controlled evals do not capture. The rapid cadence of model additions also raises the question of maintenance. Each new model requires ongoing compatibility testing across tool integrations, which does not scale linearly.600 files, one command: what moonshot.ai's refactor…

Still, the direction is clear. Groq is building for a world where the application is not a chat interface but a network of tool calls orchestrated by a reasoning model. The changelog of the past five months documents the infrastructure for that world: remote MCP for connecting it, Compound for packaging it, Kimi K2 for running it, prompt caching for paying less for it. The tokens per second still matter, but they are no longer the headline. The headline is that Groq wants to be the platform where agents are assembled, not just where inference happens.Aleph Alpha builds theoretical inference model for…