GLM-4.6 from Z.ai expanded the window from GLM-4.5's 128K to 200K tokens (204,800) and lifted the output cap to 131,072, all at the unchanged $0.60 per Mtok input and $2.20 per Mtok output. GLM-4.6 became the default open-weights coding model for a lot of teams for exactly that reason: a larger working set for no additional cost. Inside Atlas, GLM-4.6 is a 357B total parameter MoE with 32B active per forward pass, and the weights are published at zai-org/GLM-4.6 on HuggingFace, so the model you benchmark is the model you can self-host.
What did GLM-4.6 change versus GLM-4.5?
GLM-4.6 expanded the context window from GLM-4.5's 128K to 204,800 tokens and lifted the output cap to 131,072, a 33 percent lift over GLM-4.5's 98,304, while Z.ai held the price at $0.60 per Mtok input and $2.20 per Mtok output.
A free window upgrade is rare enough to be worth stating plainly. GLM-4.6 gives Atlas roughly 56 percent more input room than GLM-4.5 at identical cost, which changes what fits in a single turn. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, and the practical effect of 204,800 tokens is that more of the fused result set survives into the prompt: more call sites, more tests, more of the surrounding module. The 131,072 output cap then matters when an agent emits a large multi-file diff in one turn rather than dribbling it across several.
How many parameters does GLM-4.6 have?
GLM-4.6 is a Mixture of Experts model with 357B total parameters and 32B active per forward pass. Z.ai publishes the weights at zai-org/GLM-4.6 on HuggingFace, so the model you benchmark is the model you can self-host.
The 32B active count is what determines the compute cost of a forward pass, and 357B total is what determines the memory footprint. That combination is why GLM-4.6 serves cheaply at $0.60 per Mtok input while still holding the reasoning depth to be useful in Atlas's plan-then-build loop. Open weights close the loop on trust: a benchmark you run against the hosted zai/glm-4.6 endpoint reflects the same artifact you would deploy internally, which is not true of a hosted-only model whose serving stack can change under you.
How does GLM-4.6 pricing compare in 2026?
GLM-4.6 costs $0.60 per Mtok input and $2.20 per Mtok output, which is exactly what GLM-4.5 charged in July 2025. Z.ai raised the context from 131,072 to 204,800 tokens and left the price alone, which is why GLM-4.6 spread so quickly.
The pricing story of GLM-4.6 is that there is no pricing story. Nothing moved. What moved is capability, and that asymmetry made GLM-4.6 the default open-weights coding model for a lot of teams during its run. For Atlas specifically, where subagents fan out and multiply whatever the per-token rate happens to be, holding output at $2.20 per Mtok while doubling the practical working set is the kind of change that shows up directly in what workloads are economically viable to automate.
Is GLM-4.6 superseded by GLM-4.7?
Yes, within its own family. GLM-4.7 ships the same 204,800 context and the same $0.60 / $2.20 price with better agentic scores, so a new Atlas deployment starting fresh in 2026 has no pricing reason to choose GLM-4.6 over GLM-4.7.
This is the same pattern Z.ai ran with GLM-4.5, and it is good for buyers even when it is awkward for any single checkpoint. GLM-4.6 remains a completely reasonable model: 204,800 tokens, 131,072 output, open weights, $0.60 in and $2.20 out. What it no longer has is a differentiated position, because GLM-4.7 matches it on every one of those numbers and scores better on agentic tasks. The case for GLM-4.6 in 2026 is an existing validated deployment, a fine-tune on the published zai-org/GLM-4.6 weights, or a pin you do not want to move.
How do you run GLM-4.6 without China-hosted inference?
The Z.ai API is China-hosted, so route GLM-4.6 through OpenRouter or another gateway if data residency is a hard requirement. Atlas loads OpenRouter via `@openrouter/ai-sdk-provider`, and the GLM-4.6 model id on that gateway is `z-ai/glm-4.6`.
Residency is a filter, not a preference, for a lot of regulated codebases, and no price makes a China-hosted endpoint acceptable when policy forbids it. GLM-4.6 has an unusually clean answer because the weights are open: route through a US-hosted gateway with the `z-ai/glm-4.6` id, or self-host zai-org/GLM-4.6 outright. Atlas supports both paths, and Atlas can also build its code index with local Ollama embeddings, keeping code off third-party servers, so the retrieval half of the pipeline can stay in-house regardless of where inference runs.
Setup
- 01Export ZHIPU_API_KEY, or authenticate with `atlas login` and select Z.ai.
- 02Run `atlas models zai` and confirm that `glm-4.6` resolves.
- 03Pin "model": "zai/glm-4.6" in atlas.json.
- 04If you need US hosting, add OpenRouter instead: Atlas loads it via `@openrouter/ai-sdk-provider`.
- 05On OpenRouter the GLM-4.6 model id is `z-ai/glm-4.6`, not `zai/glm-4.6`.
Frequently asked questions
- what is the glm-4.6 context window
- GLM-4.6 has a 200K tokens (204,800) context window with a 131,072 max output, up from GLM-4.5's 128K context and 98,304 output.
- how much does glm-4.6 cost per million tokens
- GLM-4.6 costs $0.60 per Mtok input and $2.20 per Mtok output, the same price GLM-4.5 charged on a smaller window.
- how many parameters does glm-4.6 have
- GLM-4.6 is a Mixture of Experts model with 357B total parameters and 32B active per forward pass.
- how do I use glm-4.6 with atlas
- Export ZHIPU_API_KEY or run `atlas login`, confirm the model with `atlas models zai`, then pin "model": "zai/glm-4.6" in atlas.json.
- glm-4.6 vs glm-4.7 which should I pick
- GLM-4.7 ships the same 204,800 context and the same $0.60 / $2.20 price with better agentic scores, so a fresh deployment should pick GLM-4.7.
- can I run glm-4.6 outside china
- Yes. Route GLM-4.6 through OpenRouter, which Atlas loads via `@openrouter/ai-sdk-provider`, using the model id `z-ai/glm-4.6`.
- are glm-4.6 weights open
- Yes. GLM-4.6 weights are published at zai-org/GLM-4.6 on HuggingFace, so the model you benchmark is the model you can self-host.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Node.js in 2026
Node.js developers in 2026 can adopt Atlas, the terminal-native AI coding agent, for secure, efficient development. Leverage local embeddings, AST indexing, and robust safety features.
Atlas for PyTorch: Terminal-Native AI Coding for nn.Module, Devices, and Autograd in 2026
Atlas is a terminal-native AI coding agent for PyTorch in 2026, where device placement, autograd, and DataLoader worker counts cause most bugs and most slowness.
Diagnose a Hanging or Long-Running Command with Atlas in 2026
How to diagnose a hanging command with Atlas in 2026: the bash tool races every command against a timeout and tells you whether it is slow or blocked on input.
Atlas for Expo: Terminal-Native AI Coding for expo-router and Config Plugins in 2026
Atlas is a terminal-native AI coding agent for Expo apps in 2026, covering expo-router file routes, config plugins, and EAS build profiles with diff-first review.
Atlas for Groovy: A Terminal-Native AI Coding Agent for Gradle, Spock, and Jenkins in 2026
Atlas is a terminal-native AI coding agent for Groovy in 2026. It reads build.gradle closures and Jenkinsfiles, writes Spock specs, runs ./gradlew test, and applies Spotless.
Atlas for OCaml: A Terminal-Native AI Coding Agent for dune and opam Projects in 2026
Atlas is a terminal-native AI coding agent for OCaml in 2026. It reads dune stanzas and .mli signatures, runs dune runtest behind a prompt, and finishes with ocamlformat.
Atlas vs Qodo: Choosing Your AI Coding Agent in 2026
Comparing Atlas, the terminal-native AI coding agent, with Qodo 2.0, the multi-agent PR reviewer, for developers in 2026. Evaluate features, pricing, and workflow.
Atlas for Astro: Islands, Content Collections, and Zero JS by Default in 2026
Atlas is a terminal-native AI coding agent for Astro in 2026. It reads astro.config.mjs, src/pages, and content collection schemas, drops needless client:load directives, and runs astro check.