Meituan shipped LongCat-2.5-Preview on September 25, 2026. It is a 1.6-trillion-parameter mixture-of-experts model with a 1-million-token context window, native multimodal input, and dual API compatibility with both the OpenAI and Anthropic formats. For enterprise teams building long-horizon coding agents or document pipelines, it is the most context-dense model available at sub-$1 input pricing today.

What Launched

The LongCat team at Meituan posted the launch on their API platform on September 25. A day later, the coding tool OpenCode listed it as a free model on the Zen and Go tiers, with zero data retention confirmed by the provider. The official X thread from @Meituan_LongCat confirmed 5 million free tokens for every existing platform user.

LongCat-2.5-Preview is the successor to LongCat-2.0, which Meituan open-sourced in June 2026 with MIT-licensed weights. The 2.5 version is a closed API preview for now; no open weights have been published.

Key specifications, per Meituan’s official API platform and confirmed by multiple model aggregators including Models.dev and AnyRouter:

SpecValue
ArchitectureSparse MoE
Total parameters~1.6 trillion
Active parameters per token~48 billion
Context window (in)1,048,576 tokens (1M)
Max output per request131,072 tokens (128K)
ModalitiesText, image, video (URL)
Thinking modeOn by default, switchable
API formatsOpenAI-compatible, Anthropic-compatible
API pricing (direct)$0.30 / $1.20 per million tokens in/out
Release dateSeptember 25, 2026

Why 1M Context Changes the Agentic Equation

Most long-horizon agentic tasks hit a wall not because the model lacks capability, but because the model runs out of context before finishing the job. A typical code migration across a 200-file repository can consume 400,000 to 600,000 tokens of context when you factor in the full diff history, test output, and inline documentation. A model with a 200K window must either chunk the work or forget earlier decisions. A model with 1 million tokens keeps everything in a single pass.

Meituan describes the target use cases as “terminals, browsers, GUIs, spreadsheets, and design tools.” That framing is deliberate: each of those surfaces generates dense, repetitive token streams. A browser automation loop that records every network response, a spreadsheet agent reading 10,000 rows, a GUI controller parsing accessibility trees at each step. LongCat-2.5-Preview’s architecture is sized for exactly that kind of continuous accumulation.

The thinking mode adds another layer. Enabled by default, it lets the model run an internal reasoning chain before emitting a response, which is the same pattern that has lifted Chinese open-weight models like DeepSeek V4.1 Flash onto enterprise evaluation lists over the past year. The difference is that LongCat wraps that reasoning capability inside a context window large enough to track multi-step dependencies without truncation.

Claude Code Compatibility Is a Practical Enterprise Unlock

The Anthropic API endpoint is not a marketing footnote. It means developers who have invested in Claude Code tooling can swap the underlying model in about ten minutes by editing three environment variables:

ANTHROPIC_BASE_URL=https://api.longcat.chat/anthropic
ANTHROPIC_AUTH_TOKEN=<your-longcat-key>
ANTHROPIC_MODEL=LongCat-2.5-Preview

The billing moves to your LongCat balance. The tooling stays exactly as is.

For enterprise teams facing restrictions on routing data through US-based model providers, this is an architecturally significant option. LongCat’s platform is based in China, and Meituan has not published specifics on data residency for non-Chinese customers. Teams with strict data sovereignty requirements should evaluate those terms directly before using the direct API for sensitive workloads. The free OpenCode Zen tier explicitly advertises zero data retention for its limited-time window.

This pattern of building dual OpenAI and Anthropic API compatibility into a Chinese model has become standard practice. Alibaba’s Qwen3.8 Max followed the same playbook earlier this year, and it significantly accelerated enterprise adoption in regions where US providers carry political or data compliance friction.

Pricing: How It Stacks Up

The $0.30 input / $1.20 output per million tokens places LongCat-2.5-Preview in an unusual position: competitive with mid-tier models on price while offering a context window that only frontier models have been delivering.

ModelInput ($/M)Output ($/M)Context
LongCat-2.5-Preview$0.30$1.201M tokens
GPT-6 Luna$0.75$3.00256K tokens
GPT-6 Sol$2.50$10.00256K tokens
Claude Opus 5.5$15.00$75.00200K tokens
DeepSeek V4.1 Flash$0.22$0.8864K tokens
LongCat-2.0 (July)$0.30$1.20256K tokens

Prices as of September 27, 2026. Subject to change. LongCat-2.0 price shown from LongCat platform; OpenRouter quoted higher rates earlier in the year.

Against DeepSeek V4.1 Flash, which has been the dominant cost-efficient option for enterprise agent builds, LongCat-2.5-Preview costs slightly more per token but delivers a context window 16 times larger. For most long-horizon use cases, that tradeoff is decisive.

The Benchmark Gap

Meituan has published no benchmark scores for LongCat-2.5-Preview. That is worth naming plainly: the model you are evaluating for your next coding pipeline has no third-party verified performance numbers. LongCat-2.0 scored 59.5 on SWE-bench Pro when it was open-sourced in July, and that is the last verified data point for the model family.

For enterprise procurement, the right response is to run your own evals on a representative sample of your actual workload before committing infrastructure decisions. The two-week free window on OpenCode Zen makes that test essentially zero-cost. If your pipeline is already instrumented with traces and test cases, point it at LongCat-2.5-Preview for a week, measure what matters to your business, and compare against your current stack.

The absence of benchmarks is not unusual at launch for a preview release. What matters is whether it performs on your workloads. Meituan’s choice to go “free first and figure out pricing later” is a deliberate signal that they are prioritizing adoption over immediate monetization, which means developer feedback in the next two weeks will shape the 2.5 general availability pricing.

What Enterprise Teams Should Do Now

  1. Test it this week. The free OpenCode Zen window runs through approximately October 10. If you run long-context coding agents, this is your cheapest chance to benchmark LongCat-2.5-Preview against your actual workload before the free period ends.

  2. Evaluate the data terms before production use. LongCat’s platform is Meituan-operated. Review the API terms and data residency commitments before routing sensitive enterprise data.

  3. Consider the context window for document intelligence. If your pipeline regularly chunks documents because no model can fit them whole, LongCat-2.5-Preview at 1M tokens is worth a serious evaluation. Legal review, code audit, and financial analysis pipelines in particular can benefit from reducing chunking complexity.

  4. Watch for the open weights. Meituan open-sourced LongCat-2.0 within a few months of its preview. If that pattern repeats, open weights for 2.5-Preview could land before Q1 2027, opening up self-hosted deployment options that eliminate the data residency question entirely.

If you want to think through how LongCat-2.5-Preview fits into your AI agent architecture, the Enera team maps these decisions for enterprise teams every week. We track every significant model launch, run independent evals, and help clients avoid committing infrastructure to models that won’t hold up at scale.

To see how long-horizon models fit into broader enterprise AI transformation, see our AI-native transformation maturity framework.