Summary:
Meituan lists LongCat-2.5 as available through its API. Selected developer-run benchmarks show mixed results, while GitHub and Hugging Face distribution remain marked coming soon.Meituan’s LongCat-2.5 is now listed as available in the official API changelog dated November 2. The release follows the September 25 LongCat-2.5-Preview. Its technical announcement, dated November 1, focuses on visual understanding and extended coding tasks.
Availability needs a qualification: the announcement still labels GitHub and Hugging Face distribution “Coming soon.” API access is confirmed by the documentation; downloadable weights were not established in this check.
What developers can access
The API documentation lists a 1,048,576-token context window and a maximum output setting of 262,144 tokens. Context includes input and output together, so those limits should not be added. Thinking mode has a fixed temperature of 1.0; disabling thinking sets it to 0.7, and other temperature values are overridden.
The release notes describe compatibility with development harnesses including Claude Code, Codex and Hermes. That is an integration claim from Meituan, not evidence that every existing workflow will behave identically. Teams should test tool calls, long sessions and recovery from failed steps before moving production work.
A small, sourced comparison instead of an overall winner
The scorecard uses two developer-run evaluations from the November 1 announcement. Every selected cell is identified there as a self-run result rather than a score imported from another report. These are indicative company results, not independent testing. Model labels are the versions in that source, not a claim to compare today’s newest frontier models.
| Benchmark, higher is better | LongCat-2.5 | Claude Opus 5 | GLM-5.3 | DeepSeek V4.1 Flash |
|---|---|---|---|---|
| SWE Atlas — Test Writing | 47.8 | 62.2 | 42.2 | 48.9 |
| SpreadsheetBench 2 | 34.9 | 33.0 | 35.8 | 33.0 |
Bold marks the best numeric cell in each row among these four models. Source and methods: LongCat-2.5 evaluations. Both selected tests use Claude Code v2.1.116. Test Writing uses an Opus 4.5 judge and three-run averaging; SpreadsheetBench 2 follows its official protocol with a Gemini 3.6 Flash judge for visualization. Exact evaluation dates and full per-peer generation settings are not disclosed, limiting reproducibility. No overall ranking is calculated.
Pricing is promotional, not a permanent promise
The current pricing page lists limited-time rates per million tokens: $0.30 for uncached input, $0.006 for cached input and $1.20 for output. It provides no promotion end date and says platform prices and billing records are authoritative. Cached-input pricing only applies where caching actually qualifies; it should not be used to estimate all incoming tokens.
For a practical evaluation, run representative tasks against the same harness and budget, then compare correctness, completion time, tool failures and total spend. Long context and low listed token prices are useful inputs to that decision, but neither establishes reliability on a particular codebase.
Image: AI-assisted editorial scorecard, checked against the cited source. The authentic Meituan provider logo is used as a labeled fallback, not a distinct LongCat-2.5 logo, from Meituan’s media library, which permits editorial use. This is not an official Meituan graphic or endorsement.
