Confirmed in a primary source 25 Sept 2026 Models

A Chinese delivery company released a new model built to run inside American developer tools.

On September 25 Meituan, the Chinese food delivery and local services company, released LongCat-2.5-Preview on its own platform, adding image understanding to a model line whose previous version was a 1.6 trillion parameter system trained entirely on Chinese chips. The change it advertises first is not capability but fit: the changelog headline is deep compatibility with Claude Code and other mainstream development environments, naming five of them.

The caveat that matters is that this is a preview on a vendor's own API, the parameter and pricing figures circulating come from third parties rather than the changelog, and no independent evaluation of the new version exists yet.

Until now

Chinese labs competed on benchmark scores and on price, and the agent tooling developers actually use was an American layer above them.

Now

The tooling is the distribution channel, and a Chinese model now ships advertising that it slots into it.

On its own, this is a routine release. A preview model on a vendor API, an added image capability, and a set of coding claims nobody outside the company has tested. The numbers being repeated around it, the trillion and a half parameters and the sub-dollar pricing, are not in the changelog and belong to the previous version or to third party trackers. A card that leads with those numbers would be carding a press cycle.

What makes it worth the space is where Meituan put the emphasis. The changelog gives its own heading to being deeply compatible with Claude Code and four other harnesses, and lists them by name. For two years the assumption has been that American labs hold the top of the stack because developers work inside their tools, and that a Chinese model competing on price is still competing one layer down. Meituan is treating that layer as a socket rather than a moat: keep the workflow, change the engine, pay a fraction. That it comes from a food delivery company, using accelerators the export controls were written to deny it, is the part worth remembering when the next chip rule is drafted. The tooling is the distribution channel, and a Chinese model now ships advertising that it slots into it.

How it got here

Jul 5
Meituan publishes LongCat-2.0, 1.6 trillion parameters, open weights, one million token context, trained entirely on Chinese chips, priced below GPT-5.5 and Claude Sonnet 5.
Sep 22
Alibaba says its Zhenwu model has been improving itself for a month and publishes the score; Xiaomi's open weights model tops the global rankings the next day.
Sep 23
Alibaba ships the Qwen-Audio 3.1 stack and cuts voice API prices by up to 95 percent.
Sep 25
Meituan's changelog announces LongCat-2.5-Preview with image understanding and, as its own headline item, deep compatibility with Claude Code, Hermes, OpenClaw, OpenCode and Kilo Code.
Who has to react
Anthropic, whose coding agent is now a named compatibility target for a Chinese model; OpenAI and Google, who sell into the same harnesses; the harness maintainers themselves, Hermes, OpenClaw, OpenCode and Kilo Code, who become the switching layer whether they want to be or not; OpenRouter and the other routers, where price substitution actually happens; American enterprise buyers whose procurement rules say nothing about which model sits under an approved tool; the Bureau of Industry and Security, given that the predecessor was trained entirely on domestic accelerators; Alibaba and Xiaomi, who released in the same week; the Chinese labs that have taken commercial rights away from their open releases this month, since this one is a hosted preview with no weights promised; and every developer choosing an engine on price this quarter.

The evidence

ObservedFriday September 25, 2026. The LongCat API platform changelog carries an entry dated 2026-09-25 announcing that LongCat-2.5-Preview is available.

The whole observation, as recorded

The changelog gives three capability claims in the company's own words: a "new image understanding capability, able to parse image content, supporting cross-modal Q&A, content summarization, and complex visual reasoning"; that it "excels in code generation, code understanding, and automated programming tasks"; and, as its own headed item, "Deeply Compatible with Claude Code and Other Mainstream Dev Environments," naming Claude Code, Hermes, OpenClaw, OpenCode and Kilo Code. Figures reported elsewhere and not stated in the changelog: 1.6 trillion total parameters with about 48 billion active per request, a one million token context window, list pricing of 0.75 dollars per million input tokens and 2.95 dollars per million output tokens with promotional rates below that, and cache reads at 0.015 and 0.006 dollars. The predecessor sets the context. LongCat-2.0 was published on July 5, 2026 as an open weight mixture-of-experts model of 1.6 trillion parameters with a native one million token context and Meituan's own sparse attention design, trained entirely on Chinese accelerators, priced below GPT-5.5 and Claude Sonnet 5, and it led usage on the OpenRouter routing service. Meituan's main business is food delivery and local services in China. The changelog does not say whether the preview weights will be published, and the entry appears twice with identical text. Context inside the same week: Alibaba released the Qwen-Audio 3.1 stack on September 23 and cut voice API prices by up to 95 percent, and Alibaba's Qwen3.8-Omni-Flash appeared on September 25.

Established

Meituan's LongCat API platform changelog carries an entry dated 2026-09-25 announcing LongCat-2.5-Preview as available.
The changelog states that the model adds image understanding with cross-modal question answering, content summarisation and complex visual reasoning, and that it excels at code generation, code understanding and automated programming.
The changelog gives its own heading to deep compatibility with Claude Code and other mainstream development environments, naming Claude Code, Hermes, OpenClaw, OpenCode and Kilo Code.
The predecessor LongCat-2.0, published July 5, 2026, was a 1.6 trillion parameter open weight mixture-of-experts model with a one million token context, trained entirely on Chinese accelerators and priced below GPT-5.5 and Claude Sonnet 5.

Not established

That LongCat-2.5-Preview matches frontier models on anything. No independent evaluation of this version has been published.
That its parameter count, context length or pricing are as reported. Those figures are not in the changelog and some of them belong to the previous version.
That it was trained on Chinese chips. That is established for LongCat-2.0 and has not been stated for this release.
That the weights will be published. The changelog calls it a preview and promises no open release.

Why it weighs this much

Signal strength

CostMedium

CommitmentMedium

ReversibilityEasy

PersistenceMonths

We are scoring this modestly and deliberately, because it is a preview release on a slow news day and we would rather publish an honest three than dress it up. Cost is medium: training and serving a model at this scale is expensive, and the predecessor's training run on domestic accelerators was a genuine capital commitment, but nothing new was disclosed about this version's cost. Commitment is medium: a changelog entry naming five rival harnesses is a public product direction that the company will be held to by the developers who take it up, but it is not a contract, a price floor or a weights release. Reversibility is easy, which is the honest reading of the word preview: pricing can change, the endpoint can close, and no open weights have been promised. Persistence is months rather than years for the same reason. The strength here is not in the act but in the direction, and the bars should say so.

What to watch

strengthens A Chinese model appears in the published usage or routing rankings of a major agent harness or router as one of the top five engines by volume, or a Western vendor changes its licence or tooling terms to restrict third-party model substitution, by March 31, 2027.
weakens LongCat-2.5 does not reach general availability, or its stated harness compatibility is withdrawn, or the endpoint closes, by March 31, 2027.

The words

Mixture of experts: A design where only a fraction of a model's parameters run on any one request. It is how a 1.6 trillion parameter model can be served at the cost of a much smaller one.
Claude Code: Anthropic's command line coding agent, one of the main harnesses developers run models inside. Compatibility with it means a rival model can be substituted underneath the same workflow.
Context window: How much text a model can hold at once. A million tokens is roughly a large codebase, which is what agentic coding needs.
Preview: A release the vendor has not committed to as final. Pricing, availability and behaviour can all change, and open weights are not promised.

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.