Cursor (Anysphere) AI code editor and agentic coding platform (Agent, Tab, Composer, Bugbot, Cloud Agents).
Cursor, built by Anysphere, is the most prominent AI coding company and one of the fastest-growing software businesses ever. Its stack is a deliberate hybrid: it routes to every frontier lab (Anthropic, OpenAI, Google, xAI) while training its own frontier models (Composer, Tab) with large-scale RL on thousands of NVIDIA Blackwell GPUs, on in-house infrastructure (Anyrun, custom PyTorch+Ray pipelines). Build-vs-buy philosophy: build the model/RL/eval layer, buy best-of-breed infra (Turbopuffer, Fireworks, AWS, Datadog).
cursor.comSources last verified Jul 27, 2026
Flagship in-house model family specialized for agentic coding.
Receipts · 2
“It is specialized for software engineering through reinforcement learning (RL) in a diverse range of development environments.” — cursor.com, Oct 29, 2025
“Composer 2.5 is Cursor's own model, trained to be highly capable for agentic coding.” — cursor.com, Jul 2026
Source last verified Jul 26, 2026
Receipts · 2
“Claude Sonnet 5” — cursor.com, Jul 2026
“make API calls out to OpenAI, Anthropic, Gemini and xAI” — simonwillison.net, May 11, 2025
Source last verified Jul 26, 2026
GPT-5 series
by OpenAI OpenAI frontier and Codex models offered in the editor's model picker.
Receipts · 3
“Cursor supports frontier models from OpenAI, Anthropic, Google, SpaceXAI, and more” — cursor.com, Jul 2026
“How Cursor uses GPT-5” — openai.com, Aug 7, 2025
“GPT-3.5 and GPT-4 to ask questions and receive answers based on their code” — latent.space, Aug 22, 2023
Source last verified Jul 27, 2026
Grok 4.5
by SpaceXAI (xAI) Jointly trained frontier model offered as a first-party option across plans.
Receipts · 1
“Grok 4.5 is jointly trained by Cursor and SpaceXAI for long-running coding and knowledge work.” — cursor.com, Jul 2026
Source last verified Jul 26, 2026
Receipts · 1
“Continued pretraining on an open base model, Kimi K2.5” — cursor.com, Mar 27, 2026
Source last verified Jul 26, 2026
Fireworks AI Inference host serving Cursor's custom fine-tuned models.
Receipts · 4
“Fireworks deployed Cursor's special fine-tune of Llama-3-70b for the coding task 'Fast Apply' using the speculative API flag.” — fireworks.ai, Jun 23, 2024
“Fireworks provides the inference layer that makes these RL loops practical.” — fireworks.ai, Jun 26, 2026
“host their own custom models on Fireworks” — simonwillison.net, May 11, 2025
“We'd also like to thank Fireworks and Colfax for their collaboration and partnership.” — cursor.com, Mar 27, 2026
Source last verified Jul 27, 2026
Receipts · 2
“We are very much a 'cloud shop.' We mostly rely on AWS and then Azure for inference.” — newsletter.pragmaticengineer.com, Jun 10, 2025
“AWS for primary infrastructure, Azure and GCP for "some secondary infrastructure"” — simonwillison.net, May 11, 2025
Source last verified Jul 26, 2026
Together GPU Clusters
by Together AI Runs production inference on dedicated NVIDIA Blackwell GPU clusters.
Receipts · 1
“Cursor partnered with Together AI to deploy production inference on NVIDIA Blackwell” — together.ai, Jul 2026
Source last verified Jul 27, 2026
Model Context Protocol
by Anthropic (open standard) Extensibility layer connecting the agent to external tools and data.
Receipts · 1
“Model Context Protocol (MCP) enables Cursor to connect to external tools and data sources.” — cursor.com, Jul 2026
Source last verified Jul 26, 2026
turbopuffer Hosted vector store for codebase-index embeddings powering semantic code retrieval.
Receipts · 2
“turbopuffer is one of the few pieces of infrastructure we haven't had to worry about as we've scaled.” — turbopuffer.com, Jan 2026
“At our server, we chunk and embed the files, and store the embeddings in Turbopuffer” — simonwillison.net, May 11, 2025
Source last verified Jul 26, 2026
Indexes codebases with client-side change detection and server-side embeddings for search.
Receipts · 2
“features a cryptographic hash of every file, along with hashes of each folder” — cursor.com, Jan 27, 2026
“it improved response accuracy by 12.5% on average” — cursor.com, Jan 27, 2026
Source last verified Jul 26, 2026
Internal benchmark suites built from real usage gate model releases.
Receipts · 2
“Our benchmark, Cursor Bench, consists of real agent requests from engineers and researchers at Cursor” — cursor.com, Oct 29, 2025
“offline using BugBench, a curated benchmark of real code diffs” — cursor.com, Jan 15, 2026
Source last verified Jul 26, 2026
Receipts · 1
“heavy users and find the developer experience of Datadog vastly superior to the alternatives” — newsletter.pragmaticengineer.com, Jun 10, 2025
Source last verified Jul 26, 2026
Ray
by Anyscale (open source) Underpins the custom asynchronous reinforcement-learning training infrastructure with PyTorch.
Receipts · 2
“We built custom training infrastructure leveraging PyTorch and Ray to power asynchronous reinforcement learning at scale.” — cursor.com, Oct 29, 2025
“Build our distributed training, inference, and RL infrastructure” — cursor.com, Jul 2026
Source last verified Jul 26, 2026
NVIDIA GPUs (Blackwell)
by NVIDIA Trains mixture-of-experts models at low precision on large GPU clusters.
Receipts · 2
“allowing us to scale training to thousands of NVIDIA GPUs with minimal communication cost” — cursor.com, Oct 29, 2025
“Custom low-precision kernels for efficient MoE training on Blackwell GPUs” — cursor.com, Mar 27, 2026
Source last verified Jul 26, 2026
Internal compute platform running sandboxed cloud environments for RL rollouts.
Receipts · 2
“Anyrun, our internal compute platform for running hundreds of thousands of sandboxed coding environments” — cursor.com, Mar 27, 2026
“running hundreds of thousands of concurrent sandboxed coding environments in the cloud” — cursor.com, Oct 29, 2025
Source last verified Jul 26, 2026
Trains the Tab model continuously on live accept and reject feedback.
Receipts · 2
“rolling out new models to users frequently throughout the day and using that data for training” — cursor.com, Sep 12, 2025
“train frontier coding agents and scale RL on real user data” — cursor.com, Jul 2026
Source last verified Jul 26, 2026
Engineers hand-curate gold-standard solutions for evaluation and benchmark data.
Receipts · 1
“along with hand-curated optimal solutions to these requests” — cursor.com, Oct 29, 2025
Source last verified Jul 26, 2026
Runs agent shell commands in OS-level sandboxes with restricted filesystem access.
Receipts · 2
“Seccomp blocks unsafe syscalls, while Landlock enforces filesystem restrictions.” — cursor.com, Feb 18, 2026
“A mistaken agent can delete databases, ship broken code, or leak secrets.” — cursor.com, Feb 18, 2026
Source last verified Jul 26, 2026
Receipts · 2
“real agent requests from engineers and researchers at Cursor” — cursor.com, Oct 29, 2025
“whether they wrote it by hand, or had Cursor generate it” — newsletter.pragmaticengineer.com, Jun 10, 2025
Source last verified Jul 26, 2026
Reviews every internal pull request with Cursor's own AI reviewer.
Receipts · 1
“We also run Bugbot on all internal code at Cursor.” — cursor.com, Jan 15, 2026
Source last verified Jul 26, 2026
