Inter-office memo

To
Everyone who uses Cursor
From
Mo, Chief Executive
Date
November 20, 2026, 9:13 AM CST
Re
Composer and Chinese models

Cursor told Congress it may phase out Chinese models. What that means for Composer

On November 20, 2026, a House report said Cursor’s maker, Anysphere, “plans to explore phasing out PRC-developed open-weight models because of geopolitical risk.” Cursor’s own model, Composer 2.5, is built on Moonshot AI’s Kimi K2.5, by Cursor’s account. Nothing has changed yet. Composer 2.5 is still in Cursor, and Cursor hasn’t given a date or said what would replace it.

What Cursor told Congress

On November 20, 2026, the House Homeland Security Committee and the House Select Committee on China published a joint report on American companies’ use of AI models from Chinese labs. It closes an investigation the committees opened on April 29 with letters to Airbnb and to Anysphere, the company behind Cursor, about “Cursor’s Composer 2 model, which was reportedly built on an open-weight model developed by Moonshot AI.” DoorDash got a letter in July.

The report sets the three companies side by side. DoorDash told the committees it hadn’t changed how it runs its Moonshot model after the White House accused Moonshot of illicit distillation, and Airbnb kept Alibaba’s models approved after the Department of War listed Alibaba as a Chinese military company. Then: “By contrast, Anysphere informed the Committees that it plans to explore phasing out PRC-developed open-weight models because of geopolitical risk.” Its footnote cites a briefing Anysphere gave the committees’ majority staff.

That one sentence is all anyone has. CNBC, which saw the report before it came out, says Cursor and SpaceX, which bought Cursor in August, didn’t respond to its questions, and we found no statement from Cursor on its blog or changelog. “Explore” isn’t a decision. There’s no date, and no list of which models it means.

Which Cursor models it touches

By Cursor’s own account, it touches Cursor’s own model. Composer 2.5, Cursor’s agentic coding model since May, “is built on the same open-source checkpoint as Composer 2, Moonshot’s Kimi K2.5,” its Composer page says, and “85% of its compute goes into additional training and RL.” Cursor’s technical report on Composer 2 describes the same start: “continued pretraining on an open base model, Kimi K2.5.” Moonshot AI, which makes Kimi, is a Chinese lab.

Cursor didn’t name the base when Composer 2 came out in March. The House report says it “became public only after an outside developer identified the underlying model,” and eWeek reported then that Cursor’s Lee Robinson put the base model’s share of the compute at about a quarter.

Here is what Cursor’s models page lists, checked November 20. Composer is the only model on it that Cursor describes as built on a Chinese model.

Model Made by How you pay for it Per million tokens, in and out
Composer 2.5 Cursor, on Moonshot’s Kimi K2.5 Weekly pool, every paid plan $0.50 and $2.50
Composer 2.5 (Fast) Cursor, on Moonshot’s Kimi K2.5 Weekly pool, every paid plan $3 and $15
Grok 4.7 SpaceXAI Weekly pool, every paid plan $2 and $6
Grok 4.6 SpaceXAI Weekly pool, every paid plan $2 and $6
Claude Opus 5.5 Anthropic Prepaid credit, Premium and up $4 and $20
Claude Sonnet 5.5 Anthropic Prepaid credit, Premium and up $2 and $10
Gemini 4 Neon Google Prepaid credit, Premium and up $3 and $15
GPT-5.6 Sol OpenAI Prepaid credit, Premium and up $4 and $20

The prices for the weekly pool’s models apply when you buy usage beyond it.

What changes for you, and when

Nothing yet. Composer 2.5 is on Cursor’s models page with no notice beside it, and Cursor hasn’t said it will stop offering it. What a phase-out would mean depends on your plan:

  • Plus, at $8 a month, has only Cursor’s own models: Grok 4.7, 4.6 and 4.5, and Composer 2.5, with no Auto and no third-party models. Composer is the cheapest of them. Cursor hasn’t said what Plus would include without it.
  • Premium, Super and Ultra have the same pool, plus Claude, Gemini and GPT models at API prices from prepaid credit. Picking a model by name is the surest way to know which one answers; Cursor’s docs don’t list the models Auto’s Cost mode uses.
  • Teams and Enterprise can use Cursor Router for Auto’s Balance and Intelligence modes. Cursor’s help page lists its pool as GPT-5.5, Claude Opus 5, Grok 4.6 and Claude Fable 5, without Composer, though it adds that “The pool changes over time.” On Enterprise, admins can turn Composer off under Team Settings → Models, under Cursor Models, per Cursor’s docs. The catch: team and group settings combine “most-permissive (union),” so a group that allows Composer overrides a team that blocks it.

Where Composer runs is a separate question from where it came from. Cursor’s sub-processor list says “Our custom models are hosted with Together,” in the US, Canada and the UK, under a zero data retention agreement, and none of the inference providers it lists is a Chinese company or gives China as a location. The committees don’t think that settles it: “Running a model on American servers keeps data out of the PRC, but it does not change how the model was built, what it was trained to do, or how it behaves under conditions no one thought to test.”

What the report asks for, and what’s law

The report makes 12 recommendations. The ones that would bind anyone would bar “PRC-developed models in federal systems and federal contract work,” remove them “from high-risk functions of critical infrastructure,” and stop federal cybersecurity grants from paying for them. Another asks for a standard for documenting a model’s lineage, base model included, and an “AI bill of materials,” required for federal contractors and high-risk critical infrastructure and “voluntary elsewhere.”

None of that is law. Congress would have to pass it. What’s in force is narrower: a Homeland Security order, issued in July and made public in October, that removes DeepSeek, and third-party products built on it, from federal civilian agencies. Separately, Reuters reported on November 18 that the administration is weighing a rule to make companies publicly disclose their use of Chinese AI models; it wasn’t clear whether the idea was gaining traction.

So for most people, this is a question of policy, not law: your employer’s, your customers’, or your own.

What to do

  1. If Composer works for you, keep using it. Nothing has been withdrawn, and Cursor hasn’t set a date.
  2. If your company restricts models from Chinese labs, count Composer 2.5 as one. Cursor says so itself. Choose Grok or another lab’s model by name, and on Enterprise, turn Composer off for the team and check every Organization Group’s model settings too.
  3. If you do federal contract work, watch Congress. The report’s contractor rule would start with the largest contractors and reach the rest after a transition period. For DeepSeek, Congress has already barred Department of War contractors from using it on their contracts.
  4. If you’d rather use an agent built by a lab for its own models, Claude Code, Codex, Antigravity and Grok Build each are. Enjoy runs them side by side, each on the plan you already pay for, and lets you switch agents mid-conversation. It doesn’t run Cursor’s agent or Composer. Our Cursor alternatives guide compares the editors too.

What Cursor hasn’t said

  • Whether Composer 2.5 will be withdrawn, rebuilt on another base, or kept, and when.
  • What replaces it. Its Composer page says that with SpaceXAI it’s “training a significantly larger model from scratch, using 10x more total compute,” on SpaceXAI’s Colossus 2, but it doesn’t say when, or whether that model is the answer to the report.
  • What “PRC-developed open-weight models” covers beyond Composer’s base, and whether anything else in Cursor uses one.
  • What happens to Plus, whose only other models are Grok’s.

We’ll update this memo when Cursor says more.

7.Questions.

Is Cursor getting rid of Composer?

Not yet. Cursor told two House committees it “plans to explore phasing out PRC-developed open-weight models because of geopolitical risk,” and Composer 2.5 is built on Moonshot AI’s Kimi K2.5. But it hasn’t announced a change, a date or a replacement, and on November 20, 2026, Composer 2.5 was still on Cursor’s models page with no notice beside it.

Is Cursor’s Composer a Chinese model?

Its base is. Cursor says Composer 2.5 “is built on the same open-source checkpoint as Composer 2, Moonshot’s Kimi K2.5,” and that 85% of its compute went into Cursor’s own training on top. Cursor trains it and hosts it with Together; Moonshot AI, the Chinese lab that made Kimi, isn’t on Cursor’s list of sub-processors.

Does Composer send my code to China?

Not by Cursor’s own account. Its sub-processor list says “Our custom models are hosted with Together,” in the US, Canada and the UK, and none of the inference providers it lists is in China. The House committees’ worry is different: running a model on American servers, they write, “does not change how the model was built, what it was trained to do, or how it behaves.”

Can my company block Composer in Cursor?

Yes, on Enterprise. Admins turn models off under Team Settings → Models, in the Cursor Models list. Check Organization Groups too: a model is allowed if the team or any of a user’s groups allows it, so a group can let Composer back in.

Will the US ban Chinese AI models?

Not for most people, so far. The report asks Congress to bar PRC-developed models from federal systems, federal contract work and high-risk parts of critical infrastructure, but those are recommendations, not law. The only ban in force is DHS’s order removing DeepSeek from federal civilian agencies.

What will replace Composer?

Cursor hasn’t said. Its Composer page says that with SpaceXAI it’s “training a significantly larger model from scratch, using 10x more total compute,” but it gives no date and doesn’t tie that model to the phase-out.

Enjoy Agentic Industries · Checked against its sources on