GLM pricing — GLM-5.2 vs GLM-5-Turbo
← index | → cost-model | → deepseek-pricing
What this is
Odjeljak naslovljen „What this is”Per-token pricing and capability reference for z.ai’s GLM-5 series models as shipped via the GLM Coding Plan (same plan that backs zread MCP). Covers GLM-5.2 (flagship) and GLM-5-Turbo (speed variant). The table below is the single source of truth for cost prediction when routing tasks to ZAI models.
Not to be confused with the DeepSeek pricing page — different provider, different pricing model (no peak-valley mechanism here).
Pricing table
Odjeljak naslovljen „Pricing table”All prices in USD per 1M tokens. No cache-hit discount (z.ai does not expose per-request cache-hit/miss splits in the same way DeepSeek does).
| Role | Flagship reasoning model, 753B total / 40B active MoE |
|---|---|
| Context window | 1M tokens |
| Speed | 173 tok/s, TTFT 1.47s |
| Input | $1.40 |
| Output | $4.40 |
GLM-5-Turbo
Odjeljak naslovljen „GLM-5-Turbo”| Role | Speed/cost-optimized variant of the GLM-5 family |
|---|---|
| Context window | 200k tokens |
| Speed | Faster (quantitative figure pending from z.ai docs) |
| Input | $1.20 |
| Output | $4.00 |
The real tradeoff: context + depth, not price
Odjeljak naslovljen „The real tradeoff: context + depth, not price”Turbo is only ~9–14% cheaper per token than 5.2. The decision between them is driven by:
| Criterion | Pick GLM-5.2 | Pick GLM-5-Turbo |
|---|---|---|
| Context length | Need 200k–1M tokens | Under 200k is fine |
| Task complexity | Deep multi-step SWE | Routine coding / chat |
| Agent mode | long-horizon agents | Interactive loops, high QPS |
| Latency budget | Tolerant (1.47s TTFT) | Need snappier response |
| Cost sensitivity | Secondary concern | Marginal savings (~9–14%) |
For Colibri’s cost-mode routing, the meaningful comparison is not turbo vs 5.2 — it’s either GLM vs DeepSeek. DeepSeek V4 Pro at cache-hit prices ($0.0036/M input) is ~390× cheaper than GLM-5.2 input ($1.40/M). GLM’s pull is the 1M context window and the ZCode harness, not the token price.
Z.ai’s official coding harness for GLM-5.2, shipped Jun 2026 (current: 3.x).
Persistent workspace, multi-agent, long-running tasks — positioned against
Claude Code / Cursor / Copilot. Free client; revenue via the GLM Coding Plan
($16.20 Lite → $144 Max). The same plan that provides ZAI_API_KEY for
pi --provider zai.
ZCode is relevant to the PI_TUI_PROVIDER=zai lane as the upstream harness
target — the model GLM-5.2 is ZCode’s official recommended model.
Where this lives in code
Odjeljak naslovljen „Where this lives in code”- Colibri daemon:
PI_TUI_PROVIDER=zai/PI_TUI_MODEL=GLM-5.1set in build.cfg and propagated via the seed partition. Autospawned zot agents use this for the z.ai JSONL transport path. →build.cfg,live/operator-session/seed/build.cfg - Clawdie-iso:
doc/LLM-PROVIDER-HARNESS.md— matrix of provider/harness verify status;glm-5-turbodocumented for Aider lane via LiteLLM. →doc/LLM-PROVIDER-HARNESS.md - Skills:
.agent/skills/coding-agent/SKILL.mdand.agent/skills/ollama/SKILL.mdreference GLM-5.1 as current production model. - zread MCP: z.ai repo comprehension tool (search_doc, get_repo_structure,
read_file). Uses the same GLM Coding Plan key. Config at
~/.config/zread/zread.envon domedog.
See also
Odjeljak naslovljen „See also”- cost-model — byte-stable prefixes and cache-hit metering
- deepseek-pricing — DeepSeek V4 peak-valley pricing (the cheaper option for most workloads)
- hive-routing — cost-aware task routing across providers
Figures verified against public sources 02.jul.2026: