Anysphere

Anysphere, Inc. is an applied research company working on automating coding and the maker of Cursor. Its work combines a programming environment with custom models for predicting edits, applying code changes and carrying out software-engineering tasks. Cursor is the product; Anysphere is its legal maker, although company announcements also use the Cursor name.[1][3][4]

The organization expanded from AI assistance inside an editor to reinforcement-learning agents and code-review workflows. On August 14, 2026, it announced that SpaceX had completed its acquisition of Cursor, following an April model-training partnership with SpaceXAI. The earlier experiments that preceded the coding editor explain the team's experience with training data, model capabilities and integration into existing software.[11][14][5][2]

In his August 2023 Latent Space interview, Aman Sanger recalled an earlier project with Michael Truell, also called Cursor, for computer-aided design, or CAD. They trained models from scratch to predict CAD operations. Sanger said limited training data caused larger models to overfit—learning examples without generalizing well—and that spatial reasoning posed another obstacle: a model needed to track how drawing and extrusion operations changed a three-dimensional structure, then refer to the resulting geometry.[2][3]

Sanger also recalled difficulties building plugins for SolidWorks, Onshape and other CAD tools, followed by a brief text-to-image experiment. Around December 2022, he and Truell obtained early GPT-4 access through OpenAI's Converge program. Sanger distinguished that access from investment: they had participated in the program before building Cursor but had not yet taken money from OpenAI. Testing the model on generation, refactoring, editing and debugging convinced him that programming required a broader interface than autocomplete.[2]

According to Sanger, Arvid Lunnemark and Sualeh Asif, whom he and Truell knew from MIT, were independently building a company called Anysphere and had also begun pursuing an AI programming environment. After several months, the two pairs joined forces. The 2024 founding-team interview introduced Truell, Asif, Lunnemark and Sanger as Cursor's founding members.[2][3]

The team built Cursor as a fork of VS Code. Truell recalled that late-2022 GPT-4 access widened their ambitions and explained why they chose an editor rather than a plugin: owning the environment gave them control over how programmers interacted with increasingly capable models. Lunnemark described developing the interface, prompt construction, context retrieval and model behavior together, often with the same people working across those layers.[3]

The team's early internal prompting tool, Priompt, managed variable amounts of code, documentation and conversation history within a fixed token budget. Sanger described rules that could first remove older chat messages, then truncate file contents as space became scarce. Its JSX approach expressed priorities declaratively. By 2024, Lunnemark said separating input data from prompt rendering also made it easier to replay old inputs and evaluate changes to prompt construction.[2][3]

By 2024, the team combined frontier models with task-specific models. Cursor Tab predicted edits and navigation rather than only the next characters. Sanger described sparse mixture-of-experts models suited to long inputs and short outputs, with caching to avoid reprocessing the same code on every keystroke. Apply used a trained model to turn a frontier model's rough sketch into a file change, addressing difficulties such as incorrect line numbers while reducing the tokens the more expensive model needed to generate.[3]

Their speculative edits technique used existing code as a prediction of the rewritten file. The model could verify unchanged chunks in parallel, generate new text where it disagreed, then resume processing chunks once the output matched the original again. This accelerated rewriting and let programmers begin reviewing streamed changes before completion. Lunnemark's design priority was succinct: “Fast is fun.” He said the team discarded experiments that were not enjoyable to use.[3]

Codebase retrieval brought a separate systems problem: keeping local files and the server's semantic index synchronized. Asif described a Merkle tree of file and folder hashes that let the system compare the project root, then descend only into mismatching branches. Sanger described caching embedding vectors by code-chunk hash so colleagues using largely identical repositories did not repeatedly embed the same code. In that account, the indexing system stored vectors rather than source code on its servers.[3]

On October 29, 2025, the company introduced Composer, a mixture-of-experts agent model developed after Tab and a faster agent prototype called Cheetah. Reinforcement learning specialized Composer for software engineering with file editing, terminal commands and codebase-wide semantic search. Training rewarded efficient tool use and parallelism. Its internal Cursor Bench evaluation used employee requests and curated solutions to assess correctness and adherence to a repository's abstractions and engineering practices.[11]

Composer's training infrastructure used PyTorch and Ray for asynchronous reinforcement learning, native MXFP8 low-precision kernels and distributed GPU parallelism. The team adapted its Background Agents infrastructure by rewriting the virtual-machine scheduler for bursty training workloads. It reported training across thousands of NVIDIA GPUs and running hundreds of thousands of concurrent sandboxed coding environments. Its claimed fourfold generation-speed advantage came from internal comparisons; the accompanying qualification said GPT-5 and Sonnet 4.5 outperformed Composer on quality.[11]

In the March 27, 2026 Composer 2 account, Sasha Rush identified Kimi K2.5 as the open base model. Cursor's subsequent work combined code-focused continued pretraining with reinforcement learning in realistic Cursor sessions using the deployed tools and harness. Infrastructure included low-precision kernels for Blackwell GPUs, asynchronous training across regions and Anyrun, its internal sandbox compute platform. The account credited the Kimi K2.5, Ray, ThunderKittens and PyTorch teams and acknowledged Fireworks and Colfax as collaborators and partners.[12]

• August 22, 2024: The company announced $60 million raised in Series A funding from participants including Andreessen Horowitz, Thrive Capital and OpenAI. It reported more than 40,000 Cursor customers and work on next-edit prediction, retrieval across billions of files and speculative-inference code rewrites.[6]
• January 16, 2025: It announced $105 million raised in Series B funding from Thrive Capital, Andreessen Horowitz, Benchmark and existing investors. It reported millions of programmers using Cursor, proprietary models editing more than one billion characters daily and more than $100 million in recurring revenue.[7]
• June 6, 2025: It announced $900 million in funding at a $9.9 billion valuation from Thrive, Accel, Andreessen Horowitz and DST. It reported more than $500 million in annual recurring revenue (ARR) and use within more than half of Fortune 500 companies, including NVIDIA, Uber and Adobe; that adoption figure did not specify company-wide deployments.[8]
• November 13, 2025: It announced a $2.3 billion Series D at a $29.3 billion post-money valuation, with Accel separately confirming its co-lead role. New partners included Coatue, NVIDIA and Google. The company reported more than 300 team members and more than $1 billion in annualized revenue, a run-rate figure rather than revenue earned over a completed year.[9][10]

On June 16, 2025, the company changed Pro from request-based pricing to a $20 monthly pool of model usage at API prices. In his July 4 explanation and apology, Truell said longer agent tasks could consume substantially more tokens, making request counts a poor reflection of cost. He acknowledged that the company had not clearly explained that unlimited model usage applied to Auto, rather than every model, and offered refunds for unexpected usage charges between June 16 and July 4. These were historical plan terms.[13]

On October 17, 2025, Lunnemark announced his decision to leave Cursor, saying he had informed the team that day. His note explicitly connected Cursor with Anysphere, the name used for most of his time at the organization.[4]

The organization also expanded toward reviewing generated code. Participating investor Accel described Bugbot as Cursor's first add-on product and reported more than 3,000 customers within weeks of launch. On December 19, 2025, Cursor announced a definitive agreement to acquire Graphite. The company argued that review, safe merging and collaboration were becoming bottlenecks as code writing accelerated. It planned to retain Graphite's team and product, operate it independently and explore connections between local development, pull requests and code review; the announcement established an agreement, not a completed acquisition.[10][14]

The April 2026 SpaceXAI training partnership preceded the joint release of Grok 4.6 on August 12. The collaborative release targeted long-running agents, coding, knowledge work and visual projects, with supplemental training followed by supervised fine-tuning and agentic reinforcement learning. Two days later, Cursor announced the completed SpaceX acquisition. It presented greater compute access as a way to build stronger models more economically and lower customer costs; those benefits were expectations, and the announcement did not specify a purchase price or Anysphere's precise resulting legal ownership structure.[15][5]