LongCat-2.0: Configuring Meituan’s 1.6T Open-Source Agentic Coding Model as a Codex CLI Provider

Meituan’s LongCat-2.0 arrived on 30 June 2026 under an MIT licence, carrying 1.6 trillion parameters in a Mixture-of-Experts architecture that activates roughly 48 billion per token ¹. It scored 59.5 on SWE-bench Pro — edging past GPT-5.5’s 58.6 ² — and ships with a native one-million-token context window built on LongCat Sparse Attention, which reduces the usual quadratic attention cost to linear complexity ¹. The model had been running anonymously on OpenRouter as “Owl Alpha” for two months before the reveal, reaching first place on the Hermes Agent workspace leaderboard ³.

For Codex CLI users, the question is practical: can you wire LongCat-2.0 into your existing workflow, what does it cost, and where does it actually outperform the models you already use?

Architecture in Brief

LongCat-2.0 is not a dense transformer. Its ScMoE (Scalable Mixture-of-Experts) routing dynamically activates between 33 and 56 billion parameters per token, averaging 48 billion ¹. Post-training splits experts into three specialised categories:

Agent Experts — tool use, function calling, structured output
Reasoning Experts — multi-hop logic, chain-of-thought
Interaction Experts — instruction following, conversational coherence

An additional 135 billion parameters are allocated to N-gram embeddings covering 5-gram combinations, which improve token prediction without increasing inference-time compute ³.

The native million-token context comes from LongCat Sparse Attention (LSA), a mechanism that avoids the windowed-approximation approach used by many long-context models. Training ran on 50,000+ domestic Chinese ASIC accelerators over 30 trillion tokens, making LongCat-2.0 the first trillion-parameter model trained entirely on non-NVIDIA hardware ¹.

Benchmark Position

The headline numbers tell a nuanced story. LongCat-2.0 leads GPT-5.5 on the coding-specific benchmarks but trails on general agent work:

Benchmark	LongCat-2.0	GPT-5.5	Claude Sonnet 5	Gemini 3.1 Pro
SWE-bench Pro	59.5	58.6	63.2	54.2
SWE-bench Multilingual	77.3	—	—	—
Terminal-Bench 2.1	70.8	—	80.4	—
FORTE	73.2	77.8	—	—
BrowseComp	79.9	—	—	—

The 0.9-point SWE-bench Pro lead over GPT-5.5 falls within evaluation margins; all scores are vendor-reported and await independent verification ². Claude Sonnet 5 still leads on both SWE-bench Pro (63.2) and Terminal-Bench 2.1 (80.4) ². LongCat-2.0’s strength is the combination of near-frontier coding performance at substantially lower cost.

⚠️ All benchmark scores are self-reported by Meituan and have not yet been independently reproduced.

Pricing and Cost Position

This is where LongCat-2.0 changes the calculus:

Model	Input (per MTok)	Output (per MTok)	Approx. Cost per Coding Task
LongCat-2.0 (standard)	$0.75	$2.95	~$1.10
LongCat-2.0 (promo)	$0.30	$1.20	~$0.45
GPT-5.6 Luna	$0.20	$1.20	~$0.61
GPT-5.6 Terra	$2.00	$12.00	~$3.96
Claude Sonnet 4.6	$3.00	$15.00	~$5.20

At promotional pricing, LongCat-2.0 undercuts everything except GPT-5.6 Luna on raw token cost. At standard pricing, it sits between Luna and Terra — but with SWE-bench Pro performance closer to Terra’s tier ⁴.

Cache reads are free on the LongCat API, which matters when you are re-running agents against the same repository context ⁴.

Configuring Codex CLI with LongCat-2.0 via OpenRouter

LongCat-2.0 is available through OpenRouter, the LongCat native API, and self-hosted inference. The OpenRouter path is the simplest for most Codex CLI users ⁵.

Step 1: Set the API key

Set your OpenRouter key in the environment using the standard OPENROUTER variable that Codex CLI reads.

Step 2: Add the provider to `~/.codex/config.toml`

[model_providers.openrouter]
base_url = "https://openrouter.ai/api/v1"
wire_api = "responses"
env_key = "OPENROUTER"

Three provider names are reserved (openai, ollama, lmstudio) — openrouter is safe to use ⁵.

Step 3: Set LongCat-2.0 as the model

model_provider = "openrouter"
model = "meituan/longcat-2.0"

⚠️ The exact OpenRouter slug for LongCat-2.0 should be verified against the current OpenRouter model catalogue; slugs occasionally change.

Step 4: Verify the connection

codex doctor

The doctor output should confirm the provider is reachable and the model responds to a health-check prompt.

Direct LongCat API Configuration

For teams wanting to avoid OpenRouter’s margin, the LongCat API is accessible directly:

[model_providers.longcat]
base_url = "https://api.longcat.chat/v1"
wire_api = "responses"
env_key = "LONGCAT_KEY"

model_provider = "longcat"
model = "longcat-2.0"

⚠️ The LongCat API runs under Chinese regulatory oversight. Enterprise teams should evaluate data-residency and jurisdictional requirements before routing production traffic through it.

Named Profiles for Model Routing

LongCat-2.0 slots naturally into a tiered model-routing strategy. Use named profiles in config.toml to switch models by task complexity:

[profile.quick]
model_provider = "openrouter"
model = "meituan/longcat-2.0"

[profile.heavy]
model_provider = "openai"
model = "gpt-5.6-sol"

[profile.review]
model_provider = "openai"
model = "gpt-5.6-terra"

Then invoke by profile:

codex --profile quick "refactor the auth module to use dependency injection"
codex --profile heavy "redesign the event-sourcing pipeline"

This pattern lets you route straightforward refactoring and code-generation tasks to LongCat-2.0’s cost-effective tier, reserving Sol or Sonnet for complex architectural work where the benchmark gap matters ⁵.

Sub-Agent Model Overrides

With Multi-Agent V2, you can assign LongCat-2.0 to specific sub-agents whilst keeping a different model for the orchestrator:

model_provider = "openai"
model = "gpt-5.6-terra"

[subagent_overrides.codegen]
model_provider = "openrouter"
model = "meituan/longcat-2.0"

[subagent_overrides.review]
model_provider = "openai"
model = "gpt-5.6-sol"

This is particularly effective for code-generation sub-agents, where LongCat-2.0’s SWE-bench scores suggest it can handle the task at a fraction of the cost.

The Million-Token Context Advantage

LongCat-2.0’s native million-token context is genuinely useful for repository-scale operations. Where GPT-5.6 models cap at 272,000 tokens ⁶, LongCat-2.0 can ingest roughly 3.7 times more context before hitting its limit. This matters for:

Whole-repo analysis — scanning an entire monorepo without chunking
Cross-file refactoring — understanding dependency chains across hundreds of files
Migration tasks — framework upgrades that touch every layer of the stack

graph LR
    A[Repository Context<br/>~800K tokens] --> B{Context Window Check}
    B -->|GPT-5.6 272K| C[Compaction Required<br/>Information Loss]
    B -->|LongCat-2.0 1M| D[Full Context Retained<br/>No Truncation]
    C --> E[Partial Understanding]
    D --> F[Complete Understanding]

However, Codex CLI’s model_auto_compact_token_limit still applies. If your config.toml sets a compaction threshold below the model’s native window, compaction will trigger early regardless:

# Let LongCat-2.0 use its full context before compacting
model_auto_compact_token_limit = 900000

Risks and Considerations

Jurisdictional Exposure

LongCat-2.0’s API endpoints operate under Chinese data-protection law. For enterprises with GDPR, SOC 2, or FedRAMP obligations, routing code through these endpoints may create compliance friction. Self-hosting the MIT-licensed weights on your own infrastructure sidesteps this, but requires significant GPU capacity (or compatible ASIC hardware) ¹.

Benchmark Verification

All published benchmarks are vendor-reported. The SWE-bench Pro margin over GPT-5.5 is 0.9 points — well within the noise of evaluation variance. Until independent reproductions confirm the scores, treat them as indicative rather than definitive ².

Wire API Compatibility

Codex CLI requires wire_api = "responses" for custom providers since the older chat protocol was removed in February 2026. Ensure your provider configuration includes this setting — omitting it causes startup failures ⁵.

Token-Pack Distribution

The LongCat direct API distributes promotional token packs via flash sales four times daily at Beijing time (10:00, 16:00, 21:00, 23:00). This model is unusual and may not suit teams that need predictable billing ⁴.

When to Use LongCat-2.0

The model fits a specific niche in your Codex CLI configuration:

flowchart TD
    A[Incoming Task] --> B{Task Complexity}
    B -->|Simple codegen,<br/>refactoring| C[LongCat-2.0<br/>$0.75/$2.95 per MTok]
    B -->|Complex architecture,<br/>multi-step reasoning| D{Budget Constraint?}
    D -->|Cost-sensitive| E[GPT-5.6 Terra<br/>$2/$12 per MTok]
    D -->|Quality-first| F[Claude Sonnet 5<br/>or GPT-5.6 Sol]
    A --> G{Context Size}
    G -->|>272K tokens| H[LongCat-2.0<br/>1M native context]
    G -->|<272K tokens| B

Use LongCat-2.0 when you need near-frontier coding capability at open-source pricing, or when your context requirements exceed GPT-5.6’s 272K window. Route elsewhere for tasks where the 3.7-point gap to Claude Sonnet 5 on SWE-bench Pro matters, or where jurisdictional constraints rule out Chinese-hosted inference.

Citations

Meituan, “LongCat-2.0: 1.6T Agentic Coding LLM — 1M Context, Open Source,” longcatai.org, June 30, 2026. https://www.longcatai.org/models/longcat-2 ↩ ↩² ↩³ ↩⁴ ↩⁵
“LongCat-2.0 vs Claude Code & Claude Sonnet 5 — Price & Coding Benchmarks,” longcatai.org, July 2026. https://www.longcatai.org/benchmarks/longcat-vs-claude-code ↩ ↩² ↩³ ↩⁴
“LongCat-2.0,” Awesome Agents model card, July 2026. https://awesomeagents.ai/models/longcat-2-0/ ↩ ↩²
“Meituan reveals LongCat-2.0, undercuts GPT-5.5 and Claude Sonnet 5 on pricing,” Crypto Briefing, July 2026. https://cryptobriefing.com/meituan-longcat-2-undercuts-gpt-claude-pricing/ ↩ ↩² ↩³
“Codex CLI with OpenRouter: config.toml Setup and Models,” OpenRouter Blog, 2026. https://openrouter.ai/blog/tutorials/codex-cli-openrouter/ ↩ ↩² ↩³ ↩⁴
“Codex CLI v0.144.6 — Refreshed bundled instructions for GPT-5.6 Sol, Terra, and Luna, corrected context windows to 272,000 tokens,” OpenAI Codex changelog, July 29, 2026. https://developers.openai.com/codex/changelog ↩

LongCat-2.0: Configuring Meituan’s 1.6T Open-Source Agentic Coding Model as a Codex CLI Provider

Architecture in Brief

Benchmark Position

Pricing and Cost Position

Configuring Codex CLI with LongCat-2.0 via OpenRouter

Step 1: Set the API key

Step 2: Add the provider to ~/.codex/config.toml

Step 3: Set LongCat-2.0 as the model

Step 4: Verify the connection

Direct LongCat API Configuration

Named Profiles for Model Routing

Sub-Agent Model Overrides

The Million-Token Context Advantage

Risks and Considerations

Jurisdictional Exposure

Benchmark Verification

Wire API Compatibility

Token-Pack Distribution

When to Use LongCat-2.0

Citations

Step 2: Add the provider to `~/.codex/config.toml`