Feature Request: Mid-Tier GLM Model for Agentic Workloads — A "Sonnet" Between 4.7 and 5.2

[roman-ryzenadvanced]

Summary

Introduce a mid-tier GLM model purpose-built for agentic coding and tool-use workloads — positioned between GLM-4.7 and GLM-5.2 in both cost and capability. This model would serve as the go-to option for sustained agent usage (Hermes, OpenClaw, AutoClaw, Claude Code, etc.) without burning through quota at GLM-5-Turbo/5.2 rates.

---

Problem Statement

The current GLM model lineup for agentic tasks presents a gap in the middle:

Model	Tier Analogy	Agentic Performance	Cost	Quota Burn Rate
GLM-4.7	—	Adequate but inconsistent; users report unsatisfactory results on agent-heavy tasks	Low	Slow
GLM-5 Turbo	—	Excellent agentic performance; purpose-built for agents	High	~2× GLM-4.7
GLM-5.2	"Opus"	Top-tier reasoning and generation	Very High	Fastest
Mid-tier	"Sonnet"	Does not exist yet	?	?

Users running agents at scale — especially those working with Hermes, OpenClaw, AutoClaw, Claude Code, and similar agent frameworks — face a frustrating trilemma:

• GLM-4.7 is affordable but produces inconsistent agent-quality output, leading to more retries, wasted turns, and ultimately higher effective cost.
• GLM-5 Turbo delivers strong agentic results but burns through quota at roughly double the rate of 4.7, causing users to hit weekly limits quickly under heavy usage.
• GLM-5.2 is the flagship reasoning model but is overkill (and overpriced) for routine agent loops that don't need maximum depth.

This forces users into an all-or-nothing choice: either economize with a model that underperforms on agents, or overspend on a model that's more powerful than many tasks require. There is no sweet spot — a model that is "good enough" for most agent tasks at a sustainable price.

---

Proposed Solution

A New Mid-Tier Model ("GLM Sonnet")

Release a model explicitly positioned and marketed as the agentic workhorse: not the cheapest, not the most powerful, but the best value-per-token for agent loops.

Target profile:

• Performance: Noticeably stronger than GLM-4.7 on tool-use, multi-step reasoning, instruction-following, and code generation — closer to GLM-5 Turbo in agentic quality.
• Cost: Meaningfully cheaper than GLM-5 Turbo/5.2 — ideally 30–50% less per token than 5 Turbo.
• Quota efficiency: Designed so that heavy agent users can run full multi-hour sessions (dozens of agent dispatches, hundreds of tool calls) without hitting weekly caps.
• Optimization focus: Fine-tuned or post-trained specifically for agentic patterns: structured output, JSON tool calls, multi-turn context retention, and code editing — rather than broad conversational or creative tasks.

Why This Model Tier Matters

1. Agent loops are high-volume by nature. A single agent task may consume 50–200k tokens across multiple turns and tool calls. At GLM-5 Turbo pricing, a power user can exhaust a weekly quota in 2–3 focused sessions. A mid-tier model extends that to 5–7 sessions.

2. GLM-4.7's weaknesses compound in agents. In a multi-step agent workflow, a model that occasionally misinterprets tool outputs or loses context mid-chain creates cascading failures. Users end up spending more total tokens on retries than they would have spent on a slightly better model in one pass. A mid-tier model that gets it right the first time actually saves quota.

3. The "Opus / Sonnet / Haiku" tiering pattern works. Anthropic's Claude model family demonstrates that a clear three-tier structure (flagship / mid-tier / budget) helps users self-select the right model per task. Z.ai currently has the top (5.2) and bottom (4.7) but is missing the critical middle layer that most users would use most of the time.

4. System load distribution. If many users currently default to GLM-5.2 for tasks that don't require it (simply because 4.7 isn't good enough), a mid-tier model would offload those requests to a lighter, faster model — reducing load on the flagship tier and improving latency for everyone.

---

Suggested Positioning

          Reasoning Power
               ▲
               │
          5.2  │  ★ Flagship ("Opus")  — deep analysis, complex code, research
               │
     MID-TIER  │  ◆ Workhorse ("Sonnet") — agents, tool-use, general coding  ← REQUESTED
               │
          4.7  │  ● Economy ("Haiku")  — lightweight tasks, fast responses
               │
               └────────────────────────────────────► Cost per Token

---

Use Cases This Would Serve

• Hermes / OpenClaw agent loops: Sustained multi-turn sessions with frequent tool dispatching.
• AutoClaw workspace tasks: Code generation, file editing, and project scaffolding without quota anxiety.
• Claude Code / ZCode subagent dispatch: Background agents that need reliability but not maximum depth.
• CI/CD agent pipelines: Automated code review, testing, and deployment agents running at scale.
• Power users on weekly quotas: Anyone who has had to stop working mid-session because they ran out of quota on GLM-5 Turbo.

---

Source

Discord user: juvenciusbr
Discord thread: https://discord.com/channels/1346756824233148527/1529900533547274291

---

Regards,
Roman
http://www.rommark.dev

VIbe Coders Telegram Channel: https://t.me/VibeCodePrompterSystem

[github-actions]

👋 感谢你的反馈，我们已经收到。

• 维护者会在 3 个工作日内 给出首次响应。
• 状态保持为 status: 待评估，你可以随时补充信息。
• 信息不全时我们会打上 needs: 更多信息 标签并 @ 你。

---

👋 Thanks — we've received your issue.

• A maintainer will respond within 3 business days.
• Status stays at status: 待评估 (Triage); feel free to add context.
• If we need more details, we'll add needs: 更多信息 and ping you.