From short snippets to project-scale thinking
Long-context coding models are large language models for software development that can hold and reason over up to a million tokens of code, documentation, and interaction history in one context, so AI coding agents can work on entire projects instead of isolated snippets at a time.
The important shift is that context size is no longer a side note; it is the feature that defines which workflows are possible. GLM-5.2 and LongCat-2.0 both support a one million-token context window, meaning an AI agent can read huge repositories, related documentation, and long tool traces in a single request instead of juggling dozens of small prompts. For developers, that changes the design question from “How do I compress this task into 8K?” to “What do I dare hand the model in one shot?”
Chinese AI company Z.ai has released GLM-5.2 as an open-source model for long coding workflows, while Meituan has unveiled LongCat-2.0 as its new flagship long-context system. The race is now about who can turn that million-token context window into the most useful AI coding agents.
GLM-5.2: open-source muscle for long-horizon agents
GLM-5.2 is opinionated software: it is built specifically for long-horizon coding-agent scenarios, not generic chat. It supports a one million-token context window and is tuned for project-scale engineering work, where an agent may need to inspect large codebases, run tools, and keep long task histories in mind. That makes it a natural fit for AI coding agents that act more like junior engineers than autocomplete widgets.
The benchmarks show why developers should care. On SWE-bench Pro, GLM-5.2 scores 62.1 versus 58.4 for GLM-5.1, and on Terminal-Bench 2.1 it jumps from 62.0 to 81.0. The company also reports a best harness result of 82.7. In other words, the model is not just longer; it is better at the tedious command-line and code-inspection work that real agents rely on.
Equally important, GLM-5.2 is open-source under an MIT license and can be self-hosted using tools like Transformers, vLLM, SGLang, and Docker-based runners. The open-source release gives developers the option to run the model on infrastructure they control, which improves control over deployment and data handling but also shifts infrastructure and tuning responsibilities onto the user. For organizations that care about privacy or custom tooling, that trade-off is often worth it.
LongCat-2.0: a production-first MoE for agentic coding
LongCat-2.0 takes a different bet: instead of self-hosting, it aims to be the production workhorse for long-context AI coding agents via API. The model is a 1.6 trillion-parameter Mixture-of-Experts system, with about 48 billion parameters active per token. It is aimed at agentic coding, tool use, long-context work, automated workflows, and complex instructions.
Like GLM-5.2, LongCat-2.0 offers a 1 million-token context window, but it also supports a maximum output length of 128K tokens through the LongCat API platform. That combination invites new workflows where agents can not only absorb giant repositories and long documents but also emit long-form plans, refactors, or test logs in one run. LongCat Sparse Attention and training on hundreds of billions of 1M-context tokens show that the architecture is built for these long-horizon tasks.
The model is exposed through OpenAI-compatible and Anthropic-compatible APIs and supports workflows similar to Claude Code and other code-focused tools. Pricing is set per million tokens, with pay-as-you-go rates of 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. RM13.70) per 1M output tokens. That structure is clearly targeted at production teams who want predictable billing while making heavy reuse of cached context.
What million-token context really changes for developers
The real impact of a million-token context window is that AI coding agents can now treat entire codebases as first-class inputs instead of piecemeal chunks. GLM-5.2 and LongCat-2.0 can ingest large repositories, related documentation, and long chains of tool output in a single pass. That does not just save time; it changes how developers think about agent architecture.
Because GLM-5.2 can process more task history, it fits long workflows like large-scale implementation, automated research, performance optimization, and complex debugging. LongCat-2.0, trained on more than 35 trillion tokens and hundreds of billions of 1M-context examples, is positioned for similar long multi-step agent tasks across large repositories and documents. In practice, this means an agent can follow a bug from issue report to log file to source file to test suite without dropping context every few steps.
The trade-off is that long-context tokens cost money and compute. Z.ai’s IndexShare tries to reduce per-token FLOPs by 2.9 times at a one million-token context length, while LongCat-2.0’s pricing rewards heavy cache reuse. Developers building AI coding agents now have a new design constraint: use the million-token window where it matters most, and avoid turning every task into an unbounded transcript.
A lower barrier—and higher expectations—for specialized coding agents
The most encouraging change is how low the barrier has become to building specialized AI coding agents on top of long-context models. GLM-5.2 is open-source under an MIT license, with self-hosting guides for multiple runtimes, and LongCat-2.0 ships with a public GitHub repository under an MIT license even as its weights remain pending. Together, they mean you no longer need access to closed models to experiment with million-token workflows.
Early developer reactions suggest GLM-5.2 is good enough for daily use, and LongCat-2.0 already underpins widely-used agentic models via platforms like OpenRouter. That combination—serious capability plus accessible tooling—raises expectations. Teams will judge agents not by whether they can autocomplete a function, but by whether they can understand a whole system, reason over long histories, and respect the constraints of real deployments.
The takeaway is simple: million-token context windows are not a curiosity. They are a new baseline for long-context coding models and AI coding agents, and developers who design with that baseline in mind will build tools that feel less like chatbots and more like reliable collaborators.
