LongCat-2.5 at $0.30/$1.20: Thinking Locks Temperature, 1M Context, and Vendor Benches Between Opus 4.8 and 5

Attributed compile + one small offline run on my box

Primary source: Introducing LongCat-2.5 — Native Multimodality for Agentic Tasks by Meituan / LongCat team, published 2026-11-1.

Pricing / billing: LongCat-2.5 pricing and API Pay-As-You-Go (limited-time discounted rates; platform billing authoritative).

API: Chat Completions + Platform Quick Start — OpenAI-compat https://api.longcat.chat/openai, Anthropic-compat https://api.longcat.chat/anthropic, model id LongCat-2.5.

Secondary CN press: 美团发布LongCat-2.5大模型，“国模国芯”实现新突破 (Meituan / 经济观察报 framing, 2026-11-02) for the “scores between Claude-Opus-4.8 and Claude-Opus-5” claim and product/API availability.

Benchmarks and prices below are vendor-reported unless I say otherwise. This box did not call chat completions. Offline preflight = public HTTP scrapes + list-price arithmetic only.

Meituan shipped LongCat-2.5 on 2026-11-1 (blog) with wide coverage on 2026-11-2: a 1.6T-parameter MoE (~** 48B** activated per token) pitched as a native multimodal model for long-horizon agentic work. For builders, the useful story is not “another 万亿 MoE.” It is four concrete checks: (1) OpenAI- and Anthropic-compatible endpoints with a single model string, (2) a thinking object that forces temperature (1.0 on / 0.7 off), (3) limited-time list price $0.30 / $0.006 cached / $1.20 per 1M tokens, and (4) vendor benches that Meituan’s own CN framing places between Claude Opus 4.8 and Opus 5 — which is a reason to run your golden eval, not a reason to merge their table into your prod decision.

Try-it surface is longcat.ai; developers get a prepaid Pay-As-You-Go API the same week. That is enough to A/B against a current default without waiting for open weights — and also enough to over-trust a vendor dashboard if you skip your own tickets.

LongCat-2.5 builds on LongCat-2.0 with stronger native visual understanding and agentic coding (vendor). Visual signals are represented as multi-level discrete tokens (RVQ over a Semantic-and-Aligned Encoder / SAE-style path), so vision and text share a unified token stream. Training emphasis is long-horizon RL: more long trajectories in long-sequence training, failure trajectories turned into verifiable tasks, mixed training across harnesses. Infra notes (for context only): long-context training and W4A4 inference on domestic AI accelerators; vendor online stress vs W8A8 INT8 cites prefill +43.3%, decode +42.77%, TTFT −15.91%, TPOT −19.72%, GMM kernel ~** 1.73×**. That is serving math, not permission to skip quality gates.

Site / docs language also says the model works across Claude Code, Codex, and Hermes harnesses. Treat that as compatibility marketing until you wire one harness and measure your own tasks. Blog eval footnotes themselves show different harnesses per bench (Claude Code on several coding rows, Codex on SkillsBench, custom browse harness on BrowseComp) — another reminder that “works across” is not “one config wins every leaderboard.”

Skip the deep train–inference / supernode digression for day-one integration. Remember only: long context (1M) is a first-class claim, multimodal is native (not a bolted vision sidecar in the marketing story), and agentic RL is the capability bet.

Platform docs advertise two bases:

Model id: LongCat-2.5. Context window: 1M tokens; max output: 128K. Default top_p on chat docs: 0.95.

Thinking is an object, not a free-form string:

"thinking": { "type": "enabled" }

or { "type": "disabled" }.

Trap: chat docs say temperature is fixed at 1.0 with thinking enabled and 0.7 with thinking disabled; other values are overridden. If your client still sends temperature: 0 from an older Anthropic/OpenAI template, you are not getting determinism — you are getting the forced default for the thinking mode you picked. Eval note on the blog: LongCat-2.5 rows use enable_thinking=on, temperature=1.0, top_p=0.95 unless specified; other models marked * are public reports.

OpenAI SDK shape is the usual base_url="https://api.longcat.chat/openai" + Bearer key. Anthropic SDK path uses https://api.longcat.chat/anthropic with the docs’ header pattern (Authorization: Bearer … plus Anthropic-style body). Pick one client stack for the A/B; do not flip formats mid-eval or you will confuse tool-schema bugs with model bugs.

Builder rule: set thinking explicitly per route, assume temperature will be clamped, and pin the base URL + model string in config — do not guess. If you need lower temperature for classification, you likely want thinking disabled (and you still get 0.7, not 0.0). Document that in the route table so the next engineer does not “fix” sampling with a dead knob.

From the LongCat-2.5 blog metrics table (selected high-signal rows). LongCat-2.5 only below; peer columns exist on the source page — do not invent third-party re-scores here.

Meituan’s 2026-11-02 CN press frames results as sitting between Claude-Opus-4.8 and Claude-Opus-5 on several mainstream benches (and exceeding Opus-5 on some). That is a vendor / press claim next to their own charts — useful as a map of where they want you to look, useless as a substitute for your golden set. Same caution as this series always uses: do not paste “between Opus 4.8 and 5” into a slide that implies an independent Artificial Analysis or third-party harness unless that harness is cited.

Jump vs 2.0 on Terminal Bench v4.0 (3.0 → 38.5) and DeepSWE (** 9.7 → 64.6**) is the internal “agentic coding moved” signal. Still: those are their runs with their thinking/temp settings. Replicate on your repos before you retire the incumbent model string.

Pay-As-You-Go currently lists LongCat-2.5 at discounted (limited-time) rates:

Prices are subject to change; platform UI + billing records win. Failed requests (401 / 403 / 429 / 500) are not billed; only HTTP 200 is.

Offline preflight used those three USD rates only. Assumptions: one trajectory = 80k input + 20k output; “80% cache hit” means 80% of input tokens bill at the cached rate; no tax, retries, or tool-overhead tokens.

Cache is the swing factor on long agent loops. Thinking-on also tends to inflate reasoning / output tokens relative to thinking-off — price the knob and the rate card together, and remember temperature will not save you from that inflation. For a rough sanity check: at these list rates, 1M in + 1M out is $1.50 before cache; an 80k/20k turn with fat prefix reuse lands closer to ~$0.03 if the 80% cache assumption holds. Measure real cached vs uncached fields on usage before you put the middle rows in a forecast.

CNY list (¥2 / ¥0.04 / ¥8) is on the same docs pages — use it if you bill in RMB; this post’s arithmetic sticks to USD for series consistency.

LongCat-2.5 on either api.longcat.chat/openai or …/anthropic. Confirm which SDK headers your client needs (Anthropic path wants the usual Anthropic-shaped request). thinking explicitly.temperature: 0 from old templates. /workspace/reputation/runs/31-longcat-2.5/ — 2026-11-02 America/Toronto:

LONGCAT_API_KEY: NOT SET
Chat completions: NOT ATTEMPTED

HTTP 200: longcat.ai/blog/longcat-2.5/
HTTP 200: platform pricing + pay-as-you-go + chat + docs + homepage
HTTP 200: meituan.com CN press (Opus 4.8–5 framing)
Needle hits: $0.30 / $0.006 / $1.20, LongCat-2.5, openai/anthropic bases,
             thinking + temperature lock language, 1M/128K, Claude Code/Codex/Hermes

List-price math (assumptions above):
  1 traj no cache $0.0480 | 80% cache $0.0292
  40-turn no cache $1.9200 | 80% cache $1.1674
  1M+1M $1.5000

NOT tested: live chat; thinking/temp override live; harness wiring;
Terminal Bench / DeepSWE / BrowseComp re-score; real cache hits; W4A4 latency

Costs are arithmetic on LongCat’s stated limited-time list rates under assumptions I state. They are not a bill, not latency, and not quality.

Only after a callable path and a golden eval win on your tickets:

If you are already deep in Anthropic- or OpenAI-shaped agent code, LongCat-2.5 is a same-week candidate for a side-by-side — not because the Opus framing is gospel, but because the integration surface is familiar and the list price is easy to model. If you need open weights, local GPUs, or a third-party index row before you trial anything, this drop is the wrong trigger; wait for a surface that matches that constraint.

LongCat-2.5’s builder-relevant release is the dual-compat API, the thinking/temperature lock, the $0.30/$1.20 limited-time card, and vendor benches framed near Opus 4.8–5. Treat those as labeled inputs, run your own eval, and only then change the production model string.

YongBo Yu · Toronto AI engineer · https://yongbo-yu.vercel.app · GitHub YongBoYu1