Long-context coding hits the million-token era
Long-context window coding models are large language models that can process up to a million tokens of code, documentation, and tool output in one continuous context, enabling coding agents to reason over entire repositories, long logs, and multi-step workflows without constant truncation or manual prompt curation.
The headline shift is that the million token LLM is no longer a research toy; it is becoming a practical tool for developer workflow AI. GLM-5.2 arrives as an open-source coding agent model with a one million-token context window licensed under MIT, targeting work on large codebases, multi-step engineering tasks, and extended reasoning sessions. LongCat-2.0 responds from the commercial side as a 1.6 trillion-parameter Mixture-of-Experts system with about 48 billion parameters active per token and the same 1 million-token context window for agentic coding and long-context work. The real story is not who pushed the bigger number, but how these designs change what coding agents can realistically do in a single run.
Specs: two million-token LLMs with very different guts
On paper, GLM-5.2 and LongCat-2.0 look similar: both are coding agent models tuned for long context window coding with a one million-token context limit. In practice, their architectures and access modes diverge sharply. GLM-5.2 is released as an open-source model under an MIT license, built for coding tasks that run across longer workflows and aimed at project-scale engineering work where the model can process larger codebases and retain more task history in one workflow. It also adds multiple thinking-effort levels, including High and Max effort modes, so users can trade latency for deeper reasoning on harder tasks.
LongCat-2.0, by contrast, is “a 1.6 trillion-parameter Mixture-of-Experts system, with approximately 48 billion parameters active per token” and a maximum output length of 128K tokens through its API. It introduces LongCat Sparse Attention for long-horizon tasks and was trained on more than 35 trillion tokens, including hundreds of billions at 1M context, explicitly positioning it for large repositories, long documents, and multi-step agent tasks. This is a horsepower-first design: you are not meant to self-host it today; you are meant to plug into its managed API.
| Spec | GLM-5.2 | LongCat-2.0 |
|---|---|---|
| Context window | 1M tokens (open-source model) | 1M tokens via API; 128K max output |
| Architecture | Long-context LLM with IndexShare and multi-effort modes | 1.6T-parameter Mixture-of-Experts, ~48B active per token |
| Primary focus | Long-horizon coding-agent workflows, large codebases, multi-step engineering | Agentic coding, tool use, long-context work, automated workflows |
| Access model | Full open-source under MIT; self-hosting supported | Hosted API with OpenAI- and Anthropic-compatible formats; not a full weights drop yet |
Openness vs managed APIs: where each model wins
GLM-5.2’s biggest win is philosophical and practical: it is an MIT-licensed million token LLM you can run on your own infrastructure. That matters if you want strict data control, custom fine-tuning, or deep integration with existing tooling. Documentation lists support for Transformers, vLLM, SGLang, Docker Model Runner, KTransformers, and deployment on NPU platforms. In other words, GLM-5.2 is the long-context coding model for teams willing to own their stack. It also brings concrete performance gains over GLM-5.1 on coding benchmarks like SWE-bench Pro and Terminal-Bench 2.1, with scores Z.ai reports as 62.1 and 81.0 respectively, closing in on some closed-model competitors.
LongCat-2.0’s advantages live in the cloud. The model is exposed through an API that supports OpenAI-compatible and Anthropic-compatible formats and is already integrated with workflows such as Claude Code, OpenClaw, and related coding tools. The pay-as-you-go pricing is transparent: USD 0.75 (approx. RM3.50) per 1M uncached input tokens, USD 0.015 (approx. RM0.07) per 1M cached input tokens, and USD 2.95 (approx. RM13.60) per 1M output tokens. For teams who care more about getting a reliable developer workflow AI today than about owning weights tomorrow, LongCat-2.0’s managed infrastructure and flexible billing are hard to ignore.
Real developer workflows: whole codebases in a single shot
The real promise of these coding agent models is not synthetic benchmarks; it is changing how developers organize work. GLM-5.2 is aimed at work involving large codebases, multi-step engineering tasks, and extended reasoning sessions, with use cases spanning large-scale implementation, automated research, performance optimization, and complex debugging. Because its one million-token context window can include related documentation and tool outputs, coding agents can move between code, commands, and test results in a single long-horizon session instead of juggling brittle prompt windows.
LongCat-2.0 is built with the same goal of long-context window coding but pushes scale via its MoE design and LongCat Sparse Attention. With training on hundreds of billions of tokens of 1M-context data and a maximum output length of 128K tokens, it is clearly tuned for agentic coding and automated workflows over large repositories. This is the kind of model you point at an entire monorepo, plus architecture docs and logs, then let run multi-step plans involving tool use and complex instruction following. Both models answer the pent-up demand for developer workflow AI that can keep the whole project in mind at once, rather than thinking in 8K-token fragments.
Shared caveats and an early verdict on million-token coding
For all the hype, both models share the same uncomfortable truth: developers still need to test them in anger. GLM-5.2 has earned praise from early adopters, including technology executives who describe it as meeting their bar for daily use, but its broader performance will depend on independent testing and real-world deployment. LongCat-2.0, meanwhile, ships as a hybrid: there is a public GitHub repository under an MIT license, but the model weights themselves are not yet released, leaving it usable through the API while self-hosting remains a promise for later.
The takeaway: million token LLMs for coding agents are no longer science fiction, but they are not plug-and-play replacements for your IDE either. GLM-5.2 represents the open, self-hosted path where you accept infrastructure complexity in exchange for control and long-context customization. LongCat-2.0 represents the managed path where you rent a powerful, 1.6T-parameter Mixture-of-Experts service with flexible billing and focus on building agent workflows today. The smart move is not to pick a side on ideology, but to pilot both, measure how they handle your real repositories and tooling, and let results—not marketing—shape your next-generation developer workflow AI stack.
