Every time your coding agent forgets what it read three files ago, you lose minutes resetting context that should never have dropped in the first place.
Constant context resets are the silent tax on every coding session
Coding agents running on smaller context windows spend a surprising amount of inference budget summarizing what they already read, instead of actually writing code. The result is agents that drift, forget earlier decisions, and force you to re-explain the same architectural choices repeatedly.
One million tokens means the whole repo stays in the room
Z.ai Launches GLM accepts up to 1,000,000 input tokens through its glm-5.2[1m] variant, letting a coding agent hold source files, tests, configuration, and conversation history simultaneously. You point your agent at the repo, set your effort level to High or Max, and the model works without the summarization interruptions that smaller windows force. Max effort is the setting Z.ai recommends for complex, multi-step coding tasks, and in Claude Code it maps directly to the xhigh, max, and ultracode options.
Senior engineers running large codebases feel this first
- Senior engineers refactoring legacy monoliths who need the full call graph visible before touching a single function
- DevOps engineers reviewing infrastructure-as-code repos where a missed dependency in a distant config file breaks the whole pipeline
- Staff engineers writing multi-file features who need test files, type definitions, and implementation in the same working memory at once
The productivity gap between engineers using 200K-token agents and those running 1M-token agents is about to become visible in sprint velocity.
The context window arms race just jumped five times in one release
GLM-5.2 arrives roughly four months after the GLM-5 line launched, and its 1M-token window is a five-fold increase over GLM-5.1’s 200,000-token ceiling, a jump that puts direct pressure on Claude’s context positioning in agentic coding workflows. If Z.ai ships a usable retrieval layer on top of this window before competitors close the gap, the default agent stack for large-repo work could shift by end of year.
What you can actually do with this today
- Load an entire mid-sized repository and ask for a cross-file refactor without context loss
- Run Max-effort reasoning on multi-step debugging across hundreds of interdependent files
- Keep conversation history, specs, and implementation in a single session without resetting
- Configure effort level per task type directly inside Claude Code using existing slash commands
No benchmark numbers were published at launch, which is either disciplined or telling depending on where GLM-5.2 lands on independent evals.
Pricing not listed — check our directory.
GLM-5.2 does not publish architecture details, and the 744B-parameter MoE base is community-inferred rather than officially confirmed, so production teams should verify behavior before committing to it in critical pipelines.
Gemini 1.5 Pro holds the same 1M-token ceiling but is a general-purpose model rather than a coding-first release. If raw context depth without coding-specific tuning is acceptable, it remains a direct alternative.
The 1M-token era for coding agents is starting without a benchmark safety net
The shift from retrieval-augmented agents to true long-context agents is happening faster than evaluation frameworks can track it. We cover tools like this every Friday — subscribe here and we’ll send the best ones straight to you.