LongCat-2.5-Preview Is Free on OpenCode: 1M Context, Image Input, and a Deadline Nobody Published

Meituan's LongCat-2.5-Preview costs nothing to call through OpenCode right now, with no request cap — and the interesting fact is not the price. It is that the end of the free window has not been written down anywhere a reader can check. OpenCode's own documentation says "free for a limited time" for the LongCat 2.5 Preview Free entry, in both the Go subscription docs and the Zen pricing table. A two-week trial period is circulating in aggregator write-ups. Neither OpenCode's Go page, its Go docs, its Zen docs, nor the models.dev metadata that powers its model picker carries a date, an expiry marker, or a trial length. So the honest version of this story is: unlimited free access, effective 25 September 2026, with an unannounced cliff edge.

That is worth stating plainly before anyone points a production agent at it.

What is actually documented, and by whom

Three separate sources have to be kept apart here, because they carry different authority.

OpenCode's own pages are the authority on the offer. The Go page lists "LongCat 2.5 Preview Free" with a "New" tag and "limited time" beside it, and an unlimited marker in both the estimated-requests-per-5-hours column and the monthly-usage column — the only other model with that treatment is Space Bunny Free. The Go documentation's pricing table repeats it: LongCat 2.5 Preview Free is Free on input, output and cached read, with a monthly limit of "Unlimited". The Zen pricing table lists the same model at Free across all three columns. The Zen free-model notes say: "LongCat 2.5 Preview Free is free on OpenCode for a limited time. Its provider follows a zero-retention policy and does not use your data for model training." The Zen privacy table puts a number on that: model training "Not used", data retention "0 days". For comparison, the same table gives Grok 4.6 and Grok 4.7 thirty days of retention, and GPT 6 Luna the same.

Zero retention is the part worth reading twice, because it is doing real work. A free coding model that keeps your prompts for training is a different proposition from one that does not, and OpenCode is explicit that this provider does neither.

The model underneath

The LongCat 2.5 Preview Free entry in OpenCode's model catalogue — served from models.dev, the metadata registry OpenCode maintains — gives the load-bearing specs. Context window 1,000,000 tokens. Maximum output 131,072. Attachment support true, which is the catalogue's flag for image input. Reasoning true, exposed as a toggle, with the reasoning trace arriving in an interleaved reasoning_content/ field. Tool calling true. Open weights false. Release date 25 September 2026.

Meituan's own API documentation agrees on the frame and adds its own detail. The vendor's changelog carries a dated entry, "Version: 2026-09-25 — LongCat-2.5-Preview Now Available", listing three claimed capabilities: image understanding that can parse image content for cross-modal question answering, content summarisation and complex visual reasoning; coding performance in generation, comprehension and automated programming; and compatibility with "Claude Code and other mainstream dev environments", naming Hermes, OpenClaw, OpenCode and Kilo Code. The vendor's pricing page states the pay-as-you-go rate for anyone calling the model directly rather than through OpenCode: $0.30 per million uncached input tokens, $0.006 per million cached input tokens, $1.20 per million output tokens — flagged on the page itself as a limited-time discounted rate, with the note that "the prices shown on the LongCat API Platform and your billing records are authoritative."

The scale figures — roughly 1.6 trillion total parameters with about 48 billion active per token on a mixture-of-experts backbone — come from Meituan's own site metadata and from Chinese trade coverage of the 25 September listing, not from the API documentation. Chinese outlets reported the launch on 26 September with the same numbers. No technical report, no model card and no benchmark table accompanied them.

One inconsistency is worth flagging because it affects whether you can rely on the feature the launch leans on hardest. The changelog presents image understanding as the headline addition. But the example response in Meituan's own "Retrieve Model" documentation still shows input_modalities/ as ["text"]/ with a text->text/ modality string. That sample may simply be stale. It is nonetheless the vendor's published contract, so treat image input as vendor-claimed until you have tested it yourself.

What "two weeks" refers to, and why you should not plan around it

The trial length is the least verifiable number attached to this launch. OpenCode says "limited time" and stops there. The models.dev entry carries no expiry field. The Go usage table's footnote reads, in full: "LongCat 2.5 Preview Free: Free for a limited time."

Aggregator pages have described the promotion as a two-week trial. That claim is plausible — OpenCode has run similarly shaped windows before, and the DeepSeek V4.1 Flash promotion in Go carried an explicit "Ends Sep 20" marker that this one lacks — but a plausible number is not a published one, and the two should never be printed as though they were the same thing. The practical read: budget for the free window to end without warning, and do not put anything on it that cannot survive a price appearing overnight.

How to actually use it

• Model id — longcat-2.5-preview-free/, exposed through OpenCode's provider config under both opencode/ (Zen) and opencode-go/ (Go).

• Endpoint for direct API use — https://opencode.ai/zen/v1/chat/completions/, described as OpenAI-compatible.

• Authentication — the same OpenCode API key that covers the rest of Zen or Go. There is no separate signup for the free model.

• Subscription requirement — none for the Zen path. The model appears in the Zen free list, so a Go subscription is an alternative route to it, not a prerequisite.

• Quota behaviour — under Go it does not consume the monthly dollar allowance, because its listed monthly limit is unlimited rather than a dollar figure. If you hit your paid limits elsewhere, OpenCode's own documentation notes you can keep using the free models.

• Reasoning — a toggle, not a sampled parameter, with the trace returned inline. Anything that parses chat completions without expecting an interleaved reasoning field should be checked first; harnesses that silently drop it will lose the model's thinking and, in practice, much of its usefulness.

Where this sits next to paying per token

A free window is a way to evaluate, not a way to run. The moment it closes, the same model costs $0.30 and $1.20 per million tokens on the vendor's platform — a rate the vendor itself labels as promotional, which means the number that survives the promotion is not this one.

That is the argument for keeping the fallback trivial. On OrcaRouter, the models we serve sit behind one key at provider list price with zero markup, so a vendor's price change reaches your bill the same day rather than at the next contract renewal, and automatic failover moves a degraded request to a healthy provider without your code knowing. LongCat-2.5-Preview is not one of our routes — we do not serve it, and you should not read any part of this piece as claiming otherwise — but the same discipline applies to whatever you point the agent at next: the model you evaluate for free is rarely the model you end up paying for, and the switch should cost you a config line.

Who should act on this, and who should not

Act now if you have an evaluation you have been deferring because per-token costs made it annoying — a long-context refactor, an image-heavy review pass, a harness compatibility check against a model with an interleaved reasoning field. Unlimited free access with zero retention is close to the ideal conditions for exactly that, and the retention policy means you can point it at a private repository without a data-processing conversation.

Wait if the thing you need is a benchmark number. Meituan has published none for this model: no vendor chart, no independent score, no leaderboard entry. Anyone quoting a LongCat-2.5-Preview benchmark this week is either quoting a different LongCat model or making it up. The free window is a chance to generate your own numbers on your own tasks, which is the only comparison that was ever going to transfer anyway.