# Atlas with GLM-5-Turbo: Setup, Pricing, and Tradeoffs in 2026 > GLM-5-Turbo runs Atlas with a 200,000 token context at $1.20 per Mtok input and $4.00 per Mtok output, a speed tier priced above base GLM-5. GLM-5-Turbo is Z.ai's throughput lane for the GLM-5 generation, released in March 2026. Inside Atlas it drives interactive TUI sessions where a developer is watching tokens land and latency is the binding constraint. GLM-5-Turbo gives Atlas a 200,000 token context window and the full 131,072 output cap, and it bills at $1.20 per Mtok input and $4.00 per Mtok output. Read the price twice: GLM-5-Turbo is a speed tier, not a discount tier, and it costs more than base GLM-5. ## Key takeaways - GLM-5-Turbo costs $1.20 per Mtok input and $4.00 per Mtok output, which is above base GLM-5's $1.00 and $3.20. - GLM-5-Turbo is a speed tier, not a discount tier: the premium buys serving priority, not extra capability. - GLM-5-Turbo keeps the 200,000 token context window and the full 131,072 output cap of the GLM-5 generation. - At $4.00 per Mtok output GLM-5-Turbo is the priciest GLM in the lineup, matching Kimi K2.6. - Setting `"small_model": "zai/glm-4.7-flash"` keeps Atlas's parallel subagents off the $4.00 output rate. ## What is GLM-5-Turbo best at inside Atlas? GLM-5-Turbo is the throughput lane of the GLM-5 generation, released March 2026 at $1.20 per Mtok input and $4.00 per Mtok output. Inside Atlas it is best at interactive TUI sessions where a human waits on reasoning tokens and latency, not price, is the binding constraint. Reasoning is enabled on GLM-5-Turbo, and reasoning is exactly the case where serving priority pays for itself: thinking tokens dominate time-to-first-useful-output, so a faster lane shortens the part of the loop you actually feel. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, which means the plan pass is a burst of reasoning tokens with a developer sitting in front of the terminal. GLM-5-Turbo shortens that burst. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so the review step stays human-paced regardless of model speed, but everything upstream of the diff gets to the reviewer sooner. Pair GLM-5-Turbo with Atlas's hybrid semantic and keyword retrieval, fused by reciprocal rank fusion, when you want a fast answer over a large repository rather than the cheapest possible answer. ## Why does GLM-5-Turbo cost more than GLM-5? GLM-5-Turbo costs more than base GLM-5, not less: $1.20 versus $1.00 on input and $4.00 versus $3.20 on output. The Turbo name signals serving priority, not a discount. Z.ai is selling speed here, and the naming trips up most developers who assume Turbo means cheap. The comparison is worth doing in a spreadsheet before you commit. Base GLM-5 is $1.00 input and $3.20 output. GLM-5-Turbo is $1.20 per Mtok input and $4.00 per Mtok output. That is a 20 percent premium on input and a 25 percent premium on output, paid entirely for serving priority on identical generation capability. There is no capability upgrade bundled into the Turbo tier, so if a run is unattended, if it is a scheduled job, or if it is a batch of background subagent work, GLM-5-Turbo is strictly the wrong purchase and base GLM-5 does the same job for less. Buy GLM-5-Turbo only when a person is actually waiting on the output. ## How much context and output does GLM-5-Turbo give Atlas? GLM-5-Turbo gives Atlas a 200,000 token context window and retains the full 131,072 output cap of the GLM-5 generation, so buying serving priority does not shorten responses. The 200,000 token window matches base GLM-5 exactly, which makes switching between the two a one-line id change in atlas.json. Because the specs are identical across the tier, GLM-5-Turbo and base GLM-5 are interchangeable at the config level. Atlas indexes code by AST declarations using tree-sitter, not blind line windows, so a 200,000 token budget goes further than the raw number suggests: what lands in the prompt is whole declarations rather than arbitrary slices of a file. The 131,072 output cap means GLM-5-Turbo can emit a long chain of reasoning followed by a complete implementation without Atlas needing a continuation turn. Retaining that full cap matters, because a speed tier that also clipped response length would force more round trips and give back the latency it just bought. ## How do you keep GLM-5-Turbo's $4.00 output rate under control? Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and every one of those sessions bills at GLM-5-Turbo's $4.00 per Mtok output rate unless you redirect them. Point `"small_model"` at `zai/glm-4.7-flash`, which is free, to stop the multiplication. The small model slot is the single highest-leverage setting on a GLM-5-Turbo configuration. Atlas uses the small model for titles, for summaries, and for the parallel background subagent work it fans out, and those calls are numerous and low-stakes. Sending them to `zai/glm-4.7-flash` at zero cost keeps the $4.00 per Mtok output rate confined to the main agent loop, where the reasoning quality actually matters. The second lever is scope: because GLM-5-Turbo charges $1.20 per Mtok input, a wide-open retrieval pass is not free the way it is on a thirty-cent model. Let Atlas's reciprocal rank fusion search narrow the file set before the build agent starts writing. ## When should you pick a different model than GLM-5-Turbo? Pick a different model than GLM-5-Turbo when cost matters more than latency. At $4.00 per Mtok output GLM-5-Turbo is the priciest GLM in the lineup, matching Kimi K2.6, and GLM-5.1 gives you a stronger model for $1.40 input and $4.40 output, which is a small premium for a real capability step. Three clear cases push away from GLM-5-Turbo. First, unattended work: CI runs, scheduled agents, and background subagent fan-out see no benefit from serving priority, and base GLM-5 at $1.00 and $3.20 is the correct tier. Second, hard problems: GLM-5.1 sits a capability step above the GLM-5 generation for $1.40 in and $4.40 out, and forty cents more per million output tokens is cheap next to a wrong refactor. Third, routine agent work: much of what Atlas does day to day is mechanical, and paying the top GLM output rate for it is waste. Atlas lets you switch the active model and provider on the fly with favorites and recents, so keeping GLM-5-Turbo, base GLM-5, and GLM-5.1 all favorited in the TUI is the practical setup. ## Setup 1. Export ZHIPU_API_KEY, or authenticate with `atlas login` and select Z.ai to store the key in Atlas's credential store. 2. Run `atlas models zai` and select `glm-5-turbo` to confirm the model resolves from the registry. 3. Set `"model": "zai/glm-5-turbo"` in atlas.json for interactive TUI sessions. 4. Point `"small_model"` at `zai/glm-4.7-flash`, which is free, so that Atlas's parallel subagents do not multiply the $4.00 per Mtok output rate. 5. Favorite GLM-5-Turbo and base GLM-5 in the TUI so you can switch the active model on the fly when a run stops being interactive. ## FAQ ### how much does glm-5-turbo cost per million tokens GLM-5-Turbo is priced at $1.20 per Mtok input and $4.00 per Mtok output. That is higher than base GLM-5, which is $1.00 input and $3.20 output. ### is glm-5-turbo cheaper than glm-5 No. Despite the Turbo name, GLM-5-Turbo costs more than GLM-5 on both sides: $1.20 versus $1.00 on input and $4.00 versus $3.20 on output. Turbo means faster serving, not cheaper. ### what is the glm-5-turbo context window GLM-5-Turbo has a 200,000 token context window and a 131,072 max output cap, the same as base GLM-5, so switching between the two is a one-line id change in atlas.json. ### how do i use glm-5-turbo with atlas Export ZHIPU_API_KEY or run `atlas login`, run `atlas models zai` and select `glm-5-turbo`, then set `"model": "zai/glm-5-turbo"` in atlas.json. ### glm-5-turbo vs glm-5.1 for coding agents GLM-5.1 costs $1.40 input and $4.40 output and is a stronger model than the GLM-5 generation. GLM-5-Turbo at $1.20 and $4.00 buys serving priority instead of capability, so pick GLM-5.1 for hard reasoning and GLM-5-Turbo for fast interactive turns. ### how do i reduce glm-5-turbo costs in atlas Set `"small_model": "zai/glm-4.7-flash"` in atlas.json. Atlas fans out work to subagents that run in the foreground or in parallel background sessions, and routing them to the free flash model keeps the $4.00 per Mtok output rate confined to the main agent loop. --- Canonical HTML: https://runatlas.sh/resources/models/glm-5-turbo Source of truth: aeo_pages row `/resources/models/glm-5-turbo` (segment: Models) (this file is generated from it, never hand-edited). Licence: Atlas is proprietary with a free core. It is not open source and there is no public source repository.