LongCat-2.5 Is Live: Meituan's 1.6T Agent Coder for $0.30

Meituan released LongCat-2.5 on November 2, 2026, and it is the most aggressive play yet in the agent-coding model race: a natively multimodal model trained to keep working through long task sequences, reporting 88.4 on Terminal Bench v2.1 (up from 70.8 for its predecessor) at an API price of $0.30 per million input tokens. That rate is a limited-time discount, every benchmark is company-reported, and the model weights are still not public. This article covers what LongCat-2.5 actually is, what the numbers mean, what it costs, and the four catches you should know before wiring it into your stack.

What LongCat-2.5 actually is

LongCat-2.5 is Meituan's follow-up to LongCat-2.0, the 1.6-trillion-parameter mixture-of-experts model the company put out this summer. The headline change is multimodal input: LongCat-2.5 processes visual and text inputs inside a single token framework, which Meituan connects to visual coding, creative design and 3D tasks, TechNode reported on launch day.

The training story is what separates this release from a routine version bump. Meituan says it expanded training data drawn from long sequences of agent actions: not single responses, but whole chains of tool calls, file edits and terminal commands, fed into large-scale reinforcement learning so the model learns from the outcome of the full sequence. It also says it converts failures from real agent runs into reproducible training tasks, giving the model examples of work that went wrong and how to recover. For a reader building agents, that is the honest pitch: a model trained on the grind of editing a real codebase rather than on isolated code-completion prompts. An AI Affairs breakdown of the release walks through the same method.

On the engineering side, Meituan says it optimised 4-bit weights and activations (W4A4) for inference on domestic Chinese chips, keeping the numerical behaviour consistent with training. That detail matters mostly inside China. It is the company signalling it can serve a model this size on its own hardware stack, and it does not change the price you pay on the API platform.

The benchmark table, read carefully

Meituan published a full benchmark table with the release. Here it is as the company reported it:

The Terminal Bench jump from 70.8 to 88.4 is the figure that will travel furthest, and the AutomationBench move from 15.8 to 51.8 is the bigger story for anyone running agents: it measures sustained multi-step work, exactly the behaviour Meituan says it trained for. The company also claims this is a real multimodal model doing visual work, with 3DCodeBench and Vision2Web scores to back it.

Now the honest part. Every number in this table is company-reported. Daily AI Brief's launch-day coverage notes plainly that the figures have not been independently verified, and Meituan ran the tests under its own conditions. Also note what the table does not hide: on Program Bench, LongCat-2.5 scores 25.5 against 37 for the top entry in Meituan's own comparison. This is not a model that leads everywhere, and the company published that miss itself. Independent analysts on OpenRouter's model pages have not yet published third-party evaluations of the full release.

The price card, including the fine print

The LongCat API platform lists LongCat-2.5 at the following rates, per the official price table as read this week by multiple independent trackers. One caveat up front: I could not render Meituan's own pricing page in an automated check (it timed out), so the figures below come from the official table as independently read, and all three reads agree on every number.

Conversions use November 2, 2026 market rates, verified at publication (£0.746 and C$1.398 to the dollar). The headline catch is printed on the table itself: "Discounted Price (limited-time)". Meituan has not published a list price on the platform page, and the page warns that prices may change and that the figures on the platform and in your billing records are authoritative.

What would the non-discounted price be? Two independent reads point the same way. ExplainX reports list rates of $0.75 input, $0.015 cached and $2.95 output from its September read of the table, and OpenRouter shows the sibling LongCat-2.0 at 60% off ($0.75 crossed out to $0.30, $3.00 to $1.20). So the honest planning figure is roughly double to triple today's rate: budget at $0.75/$2.95 and treat $0.30/$1.20 as a launch window.

A few more facts from the rate card's fine print. Context is 1 million tokens (1,048,576) with maximum output of 131,072 tokens. At the preview launch on September 25, Meituan granted 5 million free tokens to existing verified users; whether that grant extends to the full release is not confirmed on the pricing page, so do not count on it. There is no published end date for the discount (unlike Gemini 4 Argon's promotional rate, which carries a stated November 30 end date in our Gemini 4 Argon review).

How to use it today

LongCat-2.5 is API-only. There are two doors: the LongCat chat service at longcat.chat for interactive use, and the LongCat API platform, which exposes both OpenAI-compatible and Anthropic-compatible endpoints (api.longcat.chat/openai and api.longcat.chat/anthropic). Meituan's platform documentation includes setup guides for Claude Code, Codex, Hermes, OpenCode, Cline, Kilo Code, CodeBuddy and WorkBuddy, plus OpenClaw support. That is twelve tool integrations by one count, so whatever harness you already use is probably covered. If you need a primer on the tooling side, our Claude Code explainer and OpenCode review cover the two most popular harnesses in the list.

What you cannot do yet: run it yourself. GitHub and Hugging Face releases are planned but not published: there is no repository to download, and third-party catalogues record open weights as false for the full release. LongCat-2.0 was MIT-licensed with weights published, so a 2.5 release may follow the same path, but confirm the model card before relying on it.

How it compares on price

The $0.30 rate looks cheap until you line it up against the rest of the field. All figures are per million tokens, input / output:

The honest read: LongCat-2.5's promotional price undercuts Mistral Large 4 and the Western flagships by a wide margin, but DeepSeek's V4.1 Flash is half its price even today, and our DeepSeek V4.1 Flash 1017 review found the same kind of peak/off-peak discount game underneath. If the promo ends and LongCat reverts to $0.75/$2.95, it lands right beside Mistral Large 4: cheaper than the flagships, no longer the standout. The pricing argument only works if you buy now and watch the billing page.

The catches

Catch 1: the parameter count is disputed. Pandaily and TechNode give LongCat-2.5 1.6 trillion total parameters with about 48 billion active per token. But OrcaRouter reported on October 29 that those figures came from the LongCat-2.0 model card and had been carried across to the preview documentation. The two accounts give different grounds for attaching the same figures to the new model, and Meituan has not settled it with a full-release model card. Treat 1.6T/48B as the working assumption, not a verified spec.

Catch 2: the benchmarks are all Meituan's. No independent lab, leaderboard or arena has published scores for the full release. The Terminal Bench 88.4 and AutomationBench 51.8 are the company's own measurements, and Pandaily explicitly reported them as unverified. Wait for third-party runs before treating the gap over 2.0 as settled fact.

Catch 3: the discount has no end date. A price with no published end date can end any Tuesday. Meituan's own page says the platform and billing records are authoritative, which means the $0.30 rate can move without a press release. Budget any serious workload at the $0.75/$2.95 list rate.

Catch 4: no weights, no local run, no audit. Until the promised GitHub and Hugging Face releases land, LongCat-2.5 is a closed API model with Chinese-platform billing and no way to inspect or self-host. That is a real dependency for production work, and the opposite of what made LongCat-2.0 interesting to the open-source crowd.

Pro tips and common mistakes

Do: route cached prompts aggressively: $0.006 per million cached tokens is a 50x discount on input, and agent workflows resend the same system context every turn. Do: set a spend alert tied to the list rate, not the promo rate, so a discount expiry shows up as a line item instead of a surprise. Do: use the Anthropic-compatible endpoint if your harness already speaks Claude's API shape; the OpenAI-compatible endpoint works with everything else.

Don't: quote the 88.4 Terminal Bench score as if it were independently verified. Every current citation traces back to Meituan. Don't: assume the September preview's 5-million-token grant applies to the full release. Don't: plan a self-hosted deployment around a weights release that has no date; the "planned" GitHub and Hugging Face releases are a roadmap line, not a commitment with a calendar.

Frequently asked questions

Is LongCat-2.5 open source?

Not yet. LongCat-2.0 was MIT-licensed with published weights, but LongCat-2.5 is currently API-only. Meituan says GitHub and Hugging Face releases are planned, with no date given.

How much does LongCat-2.5 cost?

The platform lists $0.30 per million uncached input tokens, $0.006 cached, and $1.20 output as a limited-time discount. Reported list rates are $0.75, $0.015 and $2.95. Budget at the list rate.

What is the context window?

1 million tokens (1,048,576), with maximum output of 131,072 tokens per request, in line with the current 1M-context generation.

Does it work with Claude Code and OpenCode?

Yes. Meituan's documentation includes setup guides for Claude Code, Codex, Hermes, OpenCode, Cline, Kilo Code, CodeBuddy, WorkBuddy and OpenClaw, through both OpenAI-compatible and Anthropic-compatible endpoints.

Is the 1.6 trillion parameter figure confirmed?

No. Pandaily attributes 1.6T total / ~48B active to the new model, but OrcaRouter traced the same figures to the LongCat-2.0 model card and says they were carried into the preview documentation. A full-release model card would settle it.

Who should try it first?

Developers already running agents in Claude Code, OpenCode or similar harnesses who want a cheap experiment: at $0.30 per million input tokens, a weekend of test runs costs pocket money. Our roundup of the 7 best AI coding agents in 2026 covers the harness landscape if you are choosing one.

The takeaway

LongCat-2.5 is a solid release with a real hook: a model trained specifically for long-horizon agent work, with a company-reported Terminal Bench jump to match, priced to win developers at a promotional $0.30 per million tokens. The discount is temporary, the benchmarks are unverified, the parameter count is contested, and the weights are not out. Try it cheap while the rate lasts; build on it only once the list price and a third-party score are public.

This is general information, not procurement or legal advice.

Sources

• TechNode — "Meituan launches LongCat-2.5 multimodal AI model for extended coding tasks" (November 2, 2026) — technode.com
• AI Affairs — "Meituan releases LongCat-2.5 for long-running coding tasks" (November 2, 2026) — aiaffairs.com
• Daily AI Brief — "Meituan releases LongCat-2.5, a 1.6 trillion parameter model for agentic coding" (November 2, 2026) — dailyaibrief.com
• LongCat official site — longcat.chat (pricing page 200-verified; render timed out in automated check, rates cross-read from independent trackers)
• ExplainX — "LongCat 2.5: Meituan 1.6T Model Targets AI Agents" (September 2026, updated) — explainx.ai
• agmazon — "LongCat-2.5-Preview Explained: 1M Context at $0.30" (October 2026) — agmazon.com
• OrcaRouter — "LongCat-2.5-Preview vs Grok 4.6" (October 2026) — orcarouter.ai
• OpenRouter — LongCat 2.0 pricing and benchmarks — openrouter.ai