LongCat-2.5 Leads the Most Crowded AI Launch Week of 2026
LongCat-2.5 is the name on every AI engineer's lips this week, and with good reason: Meituan's new 1.6-trillion-parameter mixture-of-experts model, which went from a quiet preview in September to general availability on the company's developer platform on November 2, 2026, is being pitched as a purpose-built engine for long-horizon agentic work. LongCat-2.5 is not just another big model release; it is the headline act of the most crowded AI launch week of 2026, arriving within days of AGIBOT's GO-3 Preview embodied model, Deep Principle's open-weight Hydrogen-1.0, Nace AI's Drex Max v1, a whole wave of new benchmarks, a landmark New York Times lawsuit against Brave, and a round of usage-limit cuts at OpenAI and xAI. This guide explains what LongCat-2.5 actually is, what it costs, what its benchmarks really prove, how to try it, and everything else that launched alongside it, all in one place.
If you have been following the AI news cycle closely, you already know this week has been relentless. AGIBOT unveiled its GO-3 Preview embodied foundation model on November 3. SemiAnalysis dropped GDPx, an agentic benchmark spanning eleven economically valuable domains, on November 2. The New York Times and several other major newspapers filed copyright lawsuits against Brave on November 2. OpenAI told users on Monday, November 2 that ChatGPT and Codex will share one usage limit from December, and Elon Musk spent the Halloween weekend cutting Grok Bot limits because, in his words, serving costs were "unsustainably high relative to revenue." For context on last week's security turbulence in the industry, see the story about how OpenAI's AI agents broke out of their confines before reading on.
Why lead with LongCat-2.5, though, when there is so much competition for attention? Three reasons. First, the scale is genuinely unusual: 1.6 trillion total parameters with roughly 48 billion activated per token is one of the largest MoE configurations ever put behind a public API. Second, the positioning is sharp: Meituan is not selling a general chatbot, it is selling an agent that lives inside terminals, browsers, GUIs, spreadsheets and design tools. Third, the API speaks both OpenAI and Anthropic request formats, which means it drops into existing tooling with a one-line change. That combination of scale, positioning and compatibility is why this article leads with LongCat-2.5 and then rounds up the rest of the week's launches. Every benchmark number in this article that comes from a vendor is fenced as a vendor claim, and every price is shown exactly as published, because an honest guide is worth more than a hyped one.
What LongCat-2.5 Actually Is
LongCat-2.5 is Meituan's flagship large language model, released by the Chinese technology giant best known for its food-delivery and local-services super-app, which has spent the last two years quietly building one of the most ambitious AI labs in China. The model was first made available as a preview on September 25, with some documentation dating the preview to September 27, then officially announced on Meituan's blog on November 1, 2026, and moved to general availability on the company's developer platform on November 2. The core idea behind LongCat-2.5 is simple to state and hard to execute: a natively multimodal model, built on a unified token framework that treats text and images as the same kind of sequence, designed from the ground up for long-horizon agentic work.
"Long-horizon agentic work" is the phrase Meituan keeps returning to, and it is worth unpacking because it is the model's entire reason for existing. Most frontier models are optimized for producing excellent single responses to single prompts. LongCat-2.5 is optimized for the opposite pattern: tasks that unfold over dozens or hundreds of steps, where the model has to open a terminal, run commands, read the output, browse the web, manipulate a spreadsheet, click through a GUI, or iterate inside a design tool. Meituan explicitly lists terminals, browsers, GUIs, spreadsheets and design tools as the model's target environments. That is a bet that the next phase of AI value is not better answers but better workers: agents that stay on task for an hour instead of a minute.
The native multimodality matters for the same reason. A unified token framework across text and images means the model does not bolt vision onto a text model as an afterthought; it processes a screenshot of a GUI, a chart inside a spreadsheet, or a mockup in a design tool as natively as it processes a paragraph. For agentic coding and computer-use tasks, where half the information is visual, that is a meaningful architectural choice rather than marketing language. One more detail that caught the industry's attention: LongCat-2.5 was reportedly trained on domestic Chinese chips, according to reporting by TechNode. That is reported, not confirmed by Meituan in the materials summarized here, but if accurate it would make LongCat-2.5 one of the highest-profile demonstrations that frontier-scale training can be done without the usual hardware supply chain.
LongCat-2.5 Key Specifications at a Glance
Here are the headline specifications, each verified from Meituan's published materials and third-party summaries of the release. The single most important row in this table is the parameter line: 1.6 trillion total parameters with only about 48 billion activated per token, which is what makes a model of this scale economically feasible to serve.
| Specification | LongCat-2.5 |
|---|
| Developer | Meituan |
| Architecture | Mixture-of-experts (MoE) |
| Total parameters | 1.6 trillion |
| Parameters activated per token | ~48 billion |
| Modality | Natively multimodal, unified token framework across text and images |
| Context window | 1,048,576 tokens (1 million) |
| Maximum output | 128K tokens (some documentation lists up to 256K) |
| Preview release | September 25, 2026 (some docs say September 27) |
| Official announcement | November 1, 2026 (Meituan blog) |
| General availability | November 2, 2026 (Meituan developer platform) |
| API compatibility | Both OpenAI and Anthropic request formats |
| Training hardware | Domestic Chinese chips (reported by TechNode, not officially confirmed) |
| Open weights | Not yet released, marked "coming soon" on GitHub and Hugging Face |
| Pricing (limited-time discount) | $0.30 per million input tokens (cache miss), $0.006 per million cached input tokens, $1.20 per million output tokens |
A 1-million-token context window puts LongCat-2.5 in the top tier of long-context models, and the 128K maximum output (with some documents listing up to 256K) is unusually generous: most frontier models cap output far lower, which is a real constraint for agents that need to write long documents, large code diffs, or extended tool-call transcripts. The dual API compatibility is the other standout line. Supporting both the OpenAI request format and the Anthropic request format means the model drops into a Claude Code setup or any OpenAI-compatible harness by changing a base URL and an API key, and Meituan ships an official Claude Code setup guide to make that migration as frictionless as possible. That is a deliberate land-grab for the developer-agent market, and it is working: the model is already listed on the Vercel AI Gateway at the same rates, which gives teams outside China a familiar way to try it without touching Meituan's own platform.
What "48 Billion Activated per Token" Actually Means
The headline number, 1.6 trillion parameters, is impressive, but the number that determines what LongCat-2.5 costs to run is the second one: roughly 48 billion parameters activated per token. To understand why, you need to know how a mixture-of-experts model works, and it is worth two minutes of your time because MoE is now the dominant architecture for frontier-scale models. In a mixture-of-experts model, the 1.6 trillion parameters are divided into many specialist sub-networks called experts, and for each token the model routes the computation to only a small subset of them: in LongCat-2.5's case, about 48 billion parameters' worth, or roughly 3 percent of the total.
Think of it like a hospital. A hospital might employ a thousand specialists, but any single patient sees only a handful: the cardiologist for a heart issue, the radiologist for a scan. The patient benefits from the combined knowledge of the whole institution while consuming the time of only a few doctors. LongCat-2.5 works the same way: the knowledge of the full 1.6 trillion parameters is available across a task, but each individual token costs roughly as much compute as a 48-billion-parameter dense model. That is why a model with more total parameters than anything else on a public API can still be priced at $0.30 per million input tokens: the per-token compute bill looks like a mid-size model, while the capability ceiling reflects the total parameter count.
This has three practical consequences. First, inference cost scales with activated parameters, not total parameters, so MoE models punch far above their price class. Second, the routing itself is a learned skill: the model has to learn which experts to consult for which tokens, and that routing quality is part of what Meituan's training had to get right at unprecedented scale. Third, serving a 1.6-trillion-parameter model still requires enormous memory to hold all the experts, even if only a few fire at a time, which is why the model is currently API-only with weights "coming soon": running this locally is not a realistic option for almost anyone until the weights actually appear and the community figures out efficient serving.
Pricing: What LongCat-2.5 Actually Costs
Meituan's published pricing for LongCat-2.5 is a limited-time discount, and the company is explicit that it has no announced end date and that prices may change. The headline figures are $0.30 per million input tokens on a cache miss, $0.006 per million cached input tokens, and $1.20 per million output tokens, also listed as ¥2, ¥0.04 and ¥8 respectively. The same rates apply on the Vercel AI Gateway listing, which is convenient for international teams that prefer a familiar billing relationship.
| Token type | Price per million tokens (USD) | Price per million tokens (CNY) |
|---|
| Input, cache miss | $0.30 | ¥2 |
| Input, cached | $0.006 | ¥0.04 |
| Output | $1.20 | ¥8 |
Let us do the real cost math, because API pricing only becomes meaningful when you translate it into work. One million input tokens at $0.30 means a million tokens of fresh context costs thirty cents. A cached million tokens costs six-tenths of a cent: fifty times cheaper. That 50x gap between cached and uncached input is the single most important number on the price sheet for agentic workloads, because agents are the heaviest repeat-context users in existence. An agent that keeps a 200,000-token conversation history across a long session pays $0.06 for that history the first time, and $0.0012 every subsequent time it is re-sent with a cache hit. Over a hundred-turn agent run, caching is the difference between a session that costs a few dollars and one that costs a few cents.
Now price a realistic coding-agent session. Suppose an agent reads 500,000 tokens of repository context (cached after the first read), generates 50,000 tokens of code, reasoning and tool calls, and runs for fifty turns. The first input pass costs $0.15; the remaining cached passes cost roughly a tenth of a cent each; the output costs $0.06. The whole session lands well under a dollar at list prices. That is the economic argument Meituan is making: agentic workloads burn enormous token volumes, so the model built for agents has to be priced for enormous token volumes. Two cautions, both worth stating plainly. First, these are promotional prices with no end date and an explicit warning that they may change, so do not build a business whose margins require $0.30 input pricing to hold forever. Second, I am not inventing or quoting other vendors' prices in this article for comparison; the honest statement is that LongCat-2.5 is priced aggressively for an agent-oriented frontier model, and you should check current pricing on the platform you actually use before committing to a migration.
Benchmarks: Read These as Meituan's Own Figures, Not Independent Verdicts
Meituan published an extensive evaluation package for LongCat-2.5: eight official demos and twenty-five official benchmarks, with the headline comparisons benchmarked against GPT-5.5, GPT-5.6 Sol, Claude Opus 4.8 and Claude Opus 5. Everything in this section is Meituan's own figures, not independently verified: treat them as the vendor's best case for its own model until independent evaluations replicate them.
The strongest reported numbers are on BrowseComp, where Meituan reports 91.9, and 3DCodeBench, where it reports 55.4; according to third-party summaries of the release, those are the highest scores among the five models in Meituan's own comparison set. The generational leaps over the predecessor are striking: on Terminal Bench v4.0 the score rose from 3.0 on LongCat-2.0 to 38.5 on LongCat-2.5, and on DeepSWE v1.1 it rose from 9.7 to 64.6. Those are exactly the benchmarks you would expect an agent-oriented model to target: terminal work and software engineering are the core of the long-horizon pitch. Meituan's overall claim is that LongCat-2.5 scores between Claude Opus 4.8 and Claude Opus 5, beating Opus 5 on some tasks. One important gap to note explicitly: the current Claude Opus 5.5 is not on Meituan's comparison chart at all, so the chart is comparing against the previous generation of Anthropic's flagship, not the current one.
How should you read all of this? With interest and with skepticism in equal measure. The Terminal Bench and DeepSWE jumps are genuinely large, and if even half of that improvement holds up in independent testing, LongCat-2.5 is a serious agentic coding model. But one Hacker News commenter, after a few prompts with the model, already called it "benchmaxxed," which is worth mentioning as one person's reaction rather than a fact: it captures a real and healthy suspicion that models optimized for benchmark suites can look better on paper than in the terminal. The honest position is that Meituan has published an unusually thorough first-party evaluation, the agentic benchmarks are the right ones to care about, and the verdict that matters, independent replication, has not arrived yet. For the release details and API guide, Meituan's published documentation and the third-party guides linked in the Sources section are the places to check.
How to Try LongCat-2.5 Today
You do not need to wait for anything to start experimenting with LongCat-2.5: the model is live behind an API right now, and there are three practical routes in. The fastest route for most developers is the Vercel AI Gateway, where LongCat-2.5 is listed at the same rates as Meituan's own platform, which gives teams outside China a familiar billing and integration path with no new vendor relationship to set up.
The second route is Meituan's own developer API platform, where the model reached general availability on November 2. Because the API speaks both the OpenAI request format and the Anthropic request format, integration is a one-line change for most existing setups: point your base URL at the new endpoint, swap the API key, and your existing prompts, tools and harnesses keep working. That dual compatibility is doing real work here: teams running Claude Code can drop LongCat-2.5 in as the backend, and Meituan publishes an official Claude Code setup guide that walks through the configuration step by step. The third route is Meituan's own chat service, which is the simplest way to get a feel for the model's conversational and multimodal behavior before wiring it into an agent harness.
A note on open weights, since the community is asking: the weights are not released yet. Both the GitHub repository and the Hugging Face page mark them as "coming soon," which means exactly what it says and nothing more. The predecessor, LongCat-2.0, is available on Hugging Face under the MIT license, which is a genuinely permissive license and a good sign for the eventual release, but it is not a promise about 2.5. If your plans depend on self-hosting or fine-tuning, wait for an actual release announcement rather than building on a "coming soon" badge.
Honest Caveats Before You Build on LongCat-2.5
Every new model launch deserves a clear-eyed reading, and LongCat-2.5 is no exception. The single biggest caveat is the one already stated above: the impressive benchmark numbers are Meituan's own figures, and the only verdict that counts, independent replication on real workloads, has not happened yet. The "benchmaxxed" reaction on Hacker News may be one person's hot take, but it rhymes with a real pattern in this industry: models tuned against public benchmark suites sometimes underperform the moment they leave the benchmark's distribution. Run your own evals on your own tasks before migrating anything important.
The second caveat is the pricing. The current rates are explicitly a limited-time discount with no end date and an explicit warning that prices may change. Promotional API pricing has a long history of reverting upward once adoption is established, and a 50x cached-input discount is exactly the kind of lever a vendor adjusts when the economics shift. Treat today's prices as a great reason to experiment, not as a permanent cost structure to bet a company on.
The third caveat is the weights. "Coming soon" on GitHub and Hugging Face is not a release date, and open-weight releases from major labs have slipped before. The MIT-licensed LongCat-2.0 is encouraging precedent, but precedent is not a commitment. Finally, one planning risk worth flagging in fenced form: earlier reporting this week has mentioned a draft Chinese export-catalogue measure that could, in some readings, affect AI models and related technology leaving China. I cannot confirm its final scope or status from the sources available for this article, so treat it as a risk to monitor rather than a fact to act on. If you are building a product with a multi-year horizon on any single vendor's model, that is true regardless of which vendor you pick.
GO-3 Preview: AGIBOT's Bid to Give Robots a Brain
While Meituan was making noise in the agent world, AGIBOT, the Chinese robotics company, made its own on November 3 with GO-3 Preview, an embodied foundation model for robots announced via PRNewswire. GO-3 Preview is built on what AGIBOT calls the WRAM architecture, short for World Reasoning Action Model, an embodied-native design meant to unify how a robot perceives the world, reasons about it, and acts in it, and the company claims it ranks number one on three benchmarks: RoboDojo, RoboTwin and RoboColiseum.
The release is more than a single model checkpoint. GO-3 Preview arrives paired with Genie Evolver, a hybrid post-training method for refining robot behavior, and Genie Master, an embodied agent framework for building applications on top of the model. That three-part package, model plus post-training plus framework, mirrors the way the LLM world matured: first the base model, then the tooling that makes it usable. The detail that matters most for the open-source community is AGIBOT's commitment to fully open-source GO-3 by the end of 2026. If that happens on schedule, the robotics world gets its first truly open foundation-model stack spanning perception, reasoning and action, which would be a genuine inflection point for smaller labs and startups that cannot train embodied models from scratch.
Why does an embodied model matter in a week dominated by language-model news? Because the industry's center of gravity is shifting from models that talk about the physical world to models that operate in it. LongCat-2.5 wants to drive your terminal and browser; GO-3 Preview wants to drive a robot body. Both are bets on the same thesis: the next phase of AI is agency, not answers. The difference is only in which environment the agent inhabits, and a serious AI watcher in November 2026 needs to track both.
Hydrogen-1.0 and Drex Max v1: The Open-Weight Counterpunch
Not every launch this week was a closed API. Two open-weight releases gave the open-source community plenty to chew on, and both target specialized territory rather than trying to be everything to everyone. Deep Principle's Hydrogen-1.0 is an open-weight, 753-billion-parameter mixture-of-experts model optimized for scientific computation and research, while Nace AI's Drex Max v1, released November 1, is a 27-billion-parameter multimodal decision model with open weights under the Open RAIL-M license.
Hydrogen-1.0's positioning is the more unusual of the two. At 753 billion parameters, it is enormous for an open-weight release, and the explicit focus on scientific computation and research sets it apart from the general-purpose chat models that dominate open releases. Scientific workloads, simulations, data analysis, literature-scale reasoning, are exactly the tasks where long context and specialized training pay off, and an open-weight model in that niche lets research labs fine-tune on their own data without sending sensitive work through a third-party API. The MoE architecture keeps the per-token cost manageable despite the headline parameter count, the same economic logic that makes LongCat-2.5 viable.
Drex Max v1 is the more technically detailed release. Nace AI describes a hybrid architecture: 48 Gated DeltaNet layers combined with 16 non-causal full-attention layers, a 262,144-token context window, and multimodal inputs aimed at decision-making tasks rather than open-ended chat. The reported numbers are specific: 67.52 on Decision Index 0.4, which Nace AI says is 6.27 points above Drex v2.0, and 97.3 on InfographicVQA. Those are vendor-reported figures, so apply the usual fencing, but the direction is clear: smaller, cheaper, specialized models are closing in on tasks that used to require frontier-scale systems. The Open RAIL-M license matters too: it is an open license with responsible-use restrictions, which has become the standard compromise for labs that want wide adoption without unrestricted commercial or harmful use.
For context on the rest of the late-October landscape these releases are landing in: Google's Gemini 4 Carbon arrived October 28 as a cyber-focused model that Google describes as its strongest frontier model, with API introductory pricing reported at $2 per million input and $10 per million output tokens, reportedly rising later to $4 and $20; and Mistral's Large 4 arrived October 29 at around a trillion parameters with open weights, which Mistral claims tops open models on the Artificial Analysis Coding Agent Index while using 47 percent fewer tokens per task than GLM-5.3. Both of those characterizations are the vendors' own claims, reported here as claims. For the full breakdown of Mistral's release, see the Mistral Large 4 launch analysis. The pattern across all of these releases is unmistakable: the industry is segmenting, with different models optimized for agents, science, decisions, robotics and security rather than one model to rule them all.
The Benchmark Wave: GDPx, LifeVal and the Momentic QA Bench
New models need new tests, and this week delivered three of them, each trying to measure something the old benchmarks miss. Taken together, GDPx, LifeVal and the Momentic QA Bench signal a shift in how the industry evaluates AI: away from exam-style questions and toward economically valuable work, real-world chores, and dollars-per-task efficiency.
GDPx, released by SemiAnalysis on November 2, is the most ambitious of the three. It is an agentic benchmark spanning eleven economically valuable domains, including chip design, cybersecurity, biology, legal work, quantitative trading, materials science and software engineering. The results are sobering: Claude Opus 5.5 leads at 12.0 percent, followed by Gemini 4 Carbon at 10.9 percent and GPT-6 Astra at 9.6 percent. And the detail that should stop every reader cold: on finance and legal tasks, every model tested scored zero percent. Not low. Zero. GDPx is essentially a map of where frontier AI can do real economic work today, and the map says the territory is small but the frontier is moving.
OpenAI's LifeVal attacks the problem from the opposite direction: instead of economically valuable professional work, it evaluates assistants on the unglamorous personal admin that eats real human hours, childcare coordination, food assistance paperwork, healthcare navigation, personal finance. The benchmark is built from time-use surveys plus a 3,995-person poll, which grounds it in what people actually spend time on rather than what researchers find interesting. The headline result is a genuine leap: GPT-6 Astra scores 80.6 percent overall, against 49.3 percent for GPT-5.2 just nine months earlier, and hits 100 percent on ten of the sixteen primary categories. If GDPx measures whether AI can do your job, LifeVal measures whether it can run your life, and the answer is increasingly yes.
The Momentic QA Bench, published around November 2 by the Y Combinator Winter 2024 startup Momentic, measures something refreshingly concrete: how many real bugs an AI agent can find in a running web application. The dataset contains 1,396 human-labelled bugs across four open-source projects, Gitea, Metabase, Twenty and Saleor, spanning twenty product areas. The vendor's own exported figures, dated October 31, 2026, put Momentic's Mo agent at 250.3 bugs matched in sixty minutes at $2.88 per bug, against 94.3 for Codex at $0.56 per bug. Those are the vendor's own numbers, so fence them accordingly, but the metric itself is the innovation: cost per bug found is a business number, not a leaderboard number.
So what does 12 percent on GDPx actually tell you? Two things. First, that even the best model in the world fails roughly seven out of eight hard agentic tasks in economically valuable domains, which means anyone selling you "AI employees" today is selling a future, not a present. Second, that the benchmark conversation has finally grown up: GDPx prices capability in economic terms, LifeVal grounds it in human time-use, and the QA Bench prices it in dollars per bug. The next time a vendor quotes you a benchmark score, ask what it costs per unit of work, not just what percentage it scored. That is the question this week's benchmark wave is really asking.
NYT vs Brave: The Lawsuit That Could Reshape the AI Data Economy
On November 2, the New York Times filed a federal lawsuit in Manhattan against Brave Software, and it was not alone: the same day, the Daily News, the Chicago Tribune and the Orlando Sentinel filed a similar suit. The core allegation is that Brave scraped publishers' copyrighted content and sold it through the Brave Search API to AI companies including Anthropic, Perplexity, Mistral, xAI and Cohere, building a business on content it did not pay for. Everything in this section describes allegations, not findings: Brave has not answered in court and there has been no ruling.
The complaints are detailed and worth understanding precisely. The Times filed a 72-page complaint; the other outlets filed a 69-page complaint. They allege that Brave built its Brave Index, a corpus of more than 30 billion web pages, using an unidentified crawler that publishers cannot block, then sold access to that index through the Brave Search API to the named AI companies. The suits go further than scraping: they also target a Brave browser feature that allegedly bypasses publishers' paywalls, and they accuse Brave of blocking publishers' own advertisements while selling its own ads against their content. The legal claims number five: direct, contributory and vicarious copyright infringement, violation of the DMCA's anti-circumvention provisions, and unjust enrichment. The Times is seeking an injunction and damages. Notably, the AI companies named as Brave's customers are not defendants in these suits.
The backstory adds important context. The Times says it sent cease-and-desist letters to Brave CEO Brendan Eich in August 2025 and again in April 2026, and alleges both went unanswered. Brave, founded in 2015 and reported to have raised around $250 million, is no stranger to this fight from the other side: in March 2025 it sued News Corp seeking a judicial declaration that its use of publisher content was fair use, and News Corp countersued in July 2026. The Times is meanwhile already suing OpenAI and Microsoft, and has a separate suit against Perplexity. And the complaint cites a study finding that AI crawlers send 95.7 percent less referral traffic than traditional Google search, which is the economic wound underlying the legal theory: publishers allege they are losing both the traffic and the compensation while AI systems monetize their work.
Why does this matter for anyone who is not a media lawyer? Because every model discussed in this article, LongCat-2.5 included, was trained on web-scale data, and the legal rules for that data are being written right now, case by case. If publishers win the right to block unidentified crawlers or to demand payment for training and retrieval use, the cost structure of building frontier models changes, the composition of training data changes, and the moat around companies with licensed data grows deeper. The Brave suits are also strategically different from the OpenAI and Microsoft cases: they target the data pipeline, the index and the API layer, rather than the model itself. Watch this space, because the outcome will shape what every future model is allowed to learn from.
Usage Limits Are Tightening Across the Industry
Two announcements this week confirmed what heavy AI users have been sensing for months: the era of effectively unlimited inference is ending. OpenAI told users on Monday, November 2 that ChatGPT and Codex will share a single usage limit starting in December for Plus, Pro and Business subscribers, and over the Halloween weekend Elon Musk cut the limits on xAI's Grok Bot, saying on Sunday that "the cost of serving Grok Bot was unsustainably high relative to revenue."
The OpenAI change matters most for the exact audience this article serves: developers and agent builders. Today, many subscribers treat ChatGPT and Codex as separate pools of capacity; from December, heavy Codex sessions will eat into the same allowance as ChatGPT conversations. For teams running long agentic coding sessions, which is precisely the workload LongCat-2.5 is built for, a shared cap forces real budgeting decisions: which tasks go to the premium model, which go to a cheaper alternative, and how much caching and prompt discipline can stretch the allowance. The new limits for Grok Bot were not disclosed, but Musk's stated reason, serving costs outrunning revenue, is the same arithmetic every AI provider is doing right now.
There is a straight line connecting this section to the pricing section above. Inference at frontier scale is expensive, promotional pricing is temporary, and usage caps are how providers keep the economics working. The rational response is not outrage but architecture: cache aggressively, route easy tasks to cheap models, reserve the frontier models for the steps that actually need them, and keep an eye on aggressively priced alternatives like LongCat-2.5's $0.30-per-million input tier. The providers are telling you, in plain language, that compute is the scarce resource. Believe them and plan accordingly.
Should You Use LongCat-2.5? A Decision Guide
With the facts on the table, the practical question is who should actually reach for LongCat-2.5 and who should wait. LongCat-2.5 is worth trying now if you run agentic coding or computer-use workloads, care about token costs at scale, and can tolerate a new vendor with first-party-only benchmarks; it is worth waiting on if you need open weights, audited compliance posture, or a proven track record on your specific tasks.
You should try it this week if you are a developer running long agent sessions in tools like Claude Code, because the Anthropic-compatible API makes the experiment nearly free in engineering time. You should try it if your token bills are large enough that a $0.30-per-million input tier with 50x cheaper cached reads changes your unit economics; run one representative workload through both your current model and LongCat-2.5 and compare cost per completed task, not cost per token. You should try it if you work with Chinese-language or China-market tooling, where a model from a major Chinese lab may have better local context. And you should try it if you are simply a model evaluator by trade: the community needs independent numbers, and early independent evals are a genuine public service.
You should wait if your organization requires data-residency guarantees, security audits, or contractual SLAs that a brand-new API platform has not yet demonstrated. You should wait if your roadmap depends on fine-tuning or self-hosting, because the weights are "coming soon" and that is not a date. You should wait if your use case is casual chat rather than agentic work; LongCat-2.5 is optimized for hundred-step tasks, and a general chatbot may serve you better for single-turn questions. And everyone, without exception, should run their own evaluations before migrating production workloads: vendor benchmarks are a starting point for your testing, not a substitute for it.
Five Mistakes to Avoid With LongCat-2.5
Mistake one: treating Meituan's benchmark figures as independent verification. BrowseComp 91.9 and the Terminal Bench jump from 3.0 to 38.5 are Meituan's own numbers on Meituan's chosen comparisons, and the current Claude Opus 5.5 is not even on the chart. Use these figures as a reason to test the model, not as proof of where it ranks.
Mistake two: hard-coding the promotional price into your budget. The $0.30 input and $1.20 output rates are a limited-time discount with no end date and an explicit warning that prices may change. Model your costs at two or three times the promo rate; if the economics still work, you have a robust plan, and if the promo holds, you have a pleasant surprise.
Mistake three: building your roadmap on "coming soon" weights. The GitHub and Hugging Face pages say the weights are coming, and LongCat-2.0's MIT license is encouraging, but until files are actually downloadable, you have an API dependency, not an open model. Plan for the API you can use today.
Mistake four: deploying an agent on high-stakes tasks without human review. Long-horizon agents are exactly powerful enough to make expensive mistakes at machine speed: a terminal agent can delete the wrong directory, a browser agent can submit the wrong form. LongCat-2.5's agentic strengths amplify both capability and risk, so keep humans in the loop for anything irreversible until you have measured the failure rate on your own tasks.
Mistake five: ignoring the vendor and compliance questions. Meituan is a Chinese company, the model is served from its platform, and some readers will have regulatory, data-residency or procurement constraints that make that a non-starter regardless of the technology. That is a legitimate business consideration, not a technical judgment: check your organization's policies before you route sensitive work through any new provider.
Frequently Asked Questions
What is LongCat-2.5?
LongCat-2.5 is Meituan's flagship AI model, released to general availability on November 2, 2026. It is a natively multimodal mixture-of-experts model with 1.6 trillion total parameters and about 48 billion activated per token, designed specifically for long-horizon agentic work across terminals, browsers, GUIs, spreadsheets and design tools. Its API speaks both OpenAI and Anthropic request formats, and it offers a 1-million-token context window.
Who makes LongCat-2.5?
LongCat-2.5 is developed by Meituan, the Chinese technology company best known for its food-delivery and local-services super-app. Meituan has been building its AI lab quietly for roughly two years, and LongCat-2.5 is its most ambitious public release to date. The model was previewed in late September 2026, officially announced November 1, and made generally available on Meituan's developer platform November 2.
How big is LongCat-2.5 really?
The honest answer has two numbers. LongCat-2.5 has 1.6 trillion total parameters, but only about 48 billion are activated for each token, so its per-token compute cost resembles a 48-billion-parameter dense model while its capability ceiling reflects the full 1.6 trillion. That is the mixture-of-experts trade-off: enormous total knowledge, economical per-token inference, and very large memory requirements to hold all the experts during serving.
How much does LongCat-2.5 cost?
Meituan's published rates are a limited-time discount: $0.30 per million input tokens on a cache miss, $0.006 per million cached input tokens, and $1.20 per million output tokens, with no announced end date and an explicit warning that prices may change. The 50x discount on cached input is the most important figure for agentic workloads, which re-send large contexts repeatedly. The same rates are listed on the Vercel AI Gateway.
Is LongCat-2.5 open source? Are the weights available?
Not yet. The weights are marked "coming soon" on GitHub and Hugging Face, which is a status, not a release date. The predecessor, LongCat-2.0, is available on Hugging Face under the permissive MIT license, which is encouraging precedent but not a guarantee about 2.5. Until weights actually appear, LongCat-2.5 is API-only.
Can I use LongCat-2.5 with Claude Code?
Yes, that is one of the model's explicit selling points. LongCat-2.5's API supports the Anthropic request format, so it drops into Claude Code by changing the base URL and API key, and Meituan publishes an official Claude Code setup guide. The OpenAI request format is supported too, so OpenAI-compatible harnesses work the same way. This dual compatibility is why developers can evaluate the model with almost no integration effort.
How does LongCat-2.5 compare to Claude Opus 5?
According to Meituan's own figures, LongCat-2.5 scores between Claude Opus 4.8 and Claude Opus 5, beating Opus 5 on some tasks. Two important qualifications: these are vendor-reported numbers, not independent results, and Meituan's comparison chart does not include the current Claude Opus 5.5 at all. Independent benchmarks, including SemiAnalysis's new GDPx where Claude Opus 5.5 leads at 12.0 percent, will be the real test once LongCat-2.5 is evaluated there.
What is the context window of LongCat-2.5?
LongCat-2.5 supports a 1-million-token context window, precisely 1,048,576 tokens. The maximum output is 128K tokens, with some documentation listing up to 256K, which is unusually generous and directly serves the model's agentic positioning: long tasks need long outputs. Combined with the 50x discount on cached input tokens, the long context is priced to be used, not just advertised.
What is GO-3 Preview?
GO-3 Preview is AGIBOT's embodied foundation model for robots, announced November 3, 2026. It uses what AGIBOT calls the WRAM (World Reasoning Action Model) architecture, claims the top rank on the RoboDojo, RoboTwin and RoboColiseum benchmarks, and is paired with Genie Evolver post-training and the Genie Master agent framework. AGIBOT says GO-3 will be fully open-sourced by the end of 2026, which would make it a landmark release for open robotics.
Why did the New York Times sue Brave?
On November 2, the Times sued Brave Software in federal court in Manhattan, joined the same day by similar suits from the Daily News, Chicago Tribune and Orlando Sentinel. The suits allege Brave scraped copyrighted content into a 30-billion-page index with a crawler publishers cannot block, sold access through the Brave Search API to AI companies including Anthropic, Perplexity, Mistral, xAI and Cohere, bypassed publishers' paywalls in its browser, and blocked publishers' ads while selling its own. These are allegations only: Brave has not responded in court and there is no ruling. The named AI companies are not defendants.
The Bottom Line on the Biggest AI Week of November 2026
Step back from the individual launches and the pattern of this week is clear. The AI industry has stopped asking "how smart can one model get?" and started asking "how much real work can models do, at what cost, trained on whose data, and under what legal rules?" LongCat-2.5 answers the cost-and-agency question with 1.6 trillion parameters served at 48 billion per token and priced for hundred-step tasks. GO-3 Preview, Hydrogen-1.0 and Drex Max v1 answer the specialization question: robots, science and decisions each get their own model. GDPx, LifeVal and the Momentic QA Bench answer the measurement question by pricing capability in economic terms. The NYT-vs-Brave lawsuits answer the data question, or at least start the argument. And the OpenAI and xAI usage-limit cuts answer the economics question bluntly: inference is the scarce resource now.
For readers, the actionable summary is short. Try LongCat-2.5 this week through the Vercel AI Gateway or Meituan's API if you run agentic workloads, and measure cost per completed task rather than benchmark scores. Watch for independent evaluations before believing any vendor's chart, this vendor's included. Keep an eye on AGIBOT's open-source commitment and the Brave lawsuits, because both could reshape what you are allowed to build and train on next year. And budget your inference like the scarce resource it now officially is. November 2026 will be remembered as the week AI grew up from impressive demos to real economics, and this guide will be updated as the independent numbers come in.
Sources
Every key fact in this article is traceable to the sources below; vendor-reported figures are identified as such in the text.
LongCat-2.5 native multimodal agent model guide: https://help.apiyi.com/en/longcat-2-5-native-multimodal-agent-model-guide-en.html
LongCat-2.5 release guide and specifications: https://www.tokrace.com/en/guides/longcat-2-5-release
Meituan LongCat-2.5 multimodal coding agent model and API overview: https://www.thedevdrift.com/blog/meituan-longcat-2-5-multimodal-coding-agent-model-api
LongCat-2.5 model page: https://artificialwatch.com/model/longcat-2-5
AI news roundup, November 3, 2026: https://aitoolsrecap.com/Blog/ai-news-november-03-2026
AI model releases tracker, November 2026: https://af.net/realtime/ai-model-releases-november-2026/
AI news for the week of November 1, 2026: https://www.toolglade.com/blog/ai-news-week-of-2026-11-01
AGIBOT GO-3 Preview announcement (PRNewswire via AAP): https://aapnews.aap.com.au/aapreleases/cision20261102AE87486
TechCrunch on the New York Times and other outlets suing Brave: https://techcrunch.com/2026/11/02/new-york-times-and-other-outlets-sue-web-browser-brave-over-alleged-content-scraping/
Brave Search API, official page: https://brave.com/search/api/
AI daily newsstand, November 3, 2026: https://byobot.ai/ai-news/ai-daily-newsstand-november-3-2026
Daily AI rankings, November 3, 2026: https://theairankings.com/daily/2026-11-03/
Everything that happened in AI today, Tuesday November 3, 2026: https://www.theneuron.ai/digest/everything-that-happened-in-ai-today-tuesday-november-3-2026/
AI daily briefing, November 3, 2026: https://www.communeify.com/en/blog/ai-daily-2026-11-03/
Comments
Post a Comment