eesel on LongCat 2.5: DeepSeek Wins Most Coding Rows

On November 3, eesel published the most detailed third-party reading of LongCat 2.5: after going through the official documentation page by page, its verdict was a cheap visual coding agent worth trying — but DeepSeek V4.1 Flash, at the same price, leads on most coding benchmark rows. One timing caveat: eesel wrote on November 3, before the weights shipped, so its "API-only" conclusion became outdated on November 4, when the weights went up on Hugging Face under the MIT license. Apart from that, its documentation audit remains the most thorough third-party record available.

Vision input: the real value and the limits

Vision is the biggest addition over 2.0. Images support PNG, JPG, JPEG, WebP and GIF, up to 10MB each and 50 per request, submitted as Base64 or public URLs; video joined at GA, up to 100MB and under 3,600 seconds per clip, up to 10 per request, public URLs only. SVG and HEIC are not supported, and images with an aspect ratio above 200:1 error out. Two points matter: output is still text-only — vision goes in but does not come out; and images and video carry no separate charge, billed only as the tokens they convert into. eesel argues the real value is native-resolution processing: the model tokenizes images at their original resolution instead of compressing first, which matters more for reading charts and UI screenshots.

What changed from 2.0, and the 128K output dispute

2.5 shares 2.0's 1.6T-parameter base (about 48B active), adding vision capability and large-scale long-horizon agent training: 2.0 went GA on June 30, the 2.5 preview arrived September 25 (image input only), the technical blog announced it on November 1, and video input and Codex access joined at GA on November 2. On tooling, Codex joined the official harness list, and a new OpenAI Responses endpoint arrived (effort adjustable from none to xhigh). But the docs disagree on one number: Quick Start, the FAQ and Meituan's Claude Code configuration all say maximum output is 128K (131,072 tokens), while the Model Parameters page allows max_tokens up to 262,144. eesel's advice is to plan on 128K — treat 262K as a ceiling figure, not a promise you can rely on.

Getting started, and what the community says

For access, the official docs provide 13 configuration guides including Claude Code and Codex; temperature is fixed — 1.0 with thinking on, 0.7 with it off — and other values are silently overridden rather than rejected. Rate limits (RPM/TPM) are unpublished; the docs only note that a 429 returns retry_after 60 and recommend exponential backoff. One more detail worth noting: 2.0 and the 2.5 preview have disappeared from the official Supported Models table, with no dated retirement notice, and whether they can still be called is unverified. Community feedback is mixed: Reddit users report long tasks falling into loops, with the model "losing state" after roughly 200K context; narrow roles work well — a cheap visual agent, small tasks and committing code are fine, but it is not suited to planning or creative writing. One developer self-ran Terminal-Bench 2.1 on the free preview and got 48.7%, but that was a single-person test on the preview with a different harness — evidence that independent reproduction is missing, not a score for 2.5. On Hacker News, the discussion centred on pricing and the choice of comparison set — the official charts skip Opus 5.5 and other current frontier models.

The missing data policy, and who should pick what

eesel's audit found that neither the FAQ nor the billing page states how long prompts are retained, or whether API traffic is used for training — LongCat 2.5 currently has no public data policy. For an individual user just writing code, that may not matter; for anyone routing customer conversations or support content through the API, it is a red light: there is nothing to show compliance. eesel's overall recommendation splits four ways: try it now on the free or cheap tier if you want a taste; wait for independent evaluations if long-task stability is a hard requirement; skip it for now if you need a clear data policy or verified long-context performance; and if the top priority is the strongest coding at this price, DeepSeek V4.1 Flash still leads on most coding rows.