GLM 5 for coding: GLM-5.3, Flash, price, the Coding Plan and Claude

GLM 5 is the model family from Z.ai, the international brand of China's Zhipu AI. The latest are GLM-5.3, released August 18, 2026 with 753 billion parameters and a 1M-token context, and GLM-5.3-Flash, released August 26 with 320 billion parameters (18 billion active) under the MIT license. On Z.ai's API, GLM-5.3 costs $1.40 per million input tokens and $4.40 output, and Flash $0.15 and $0.50. On Vals AI's independent Terminal-Bench 4.0 run, GLM-5.3 scored 38.9% and Flash 37.4%, against 65.2% for Claude Opus 5.5. Z.ai's GLM Coding Plan, from $18 a month, runs GLM-5.3 inside Claude Code, Cursor, Cline and other agents.

Key takeaways

• Versions: GLM-5 (Feb 12), 5.1 (Apr 7), 5.2 (Jun 16), 5.3 (Aug 18) and 5.3-Flash (Aug 26), all 2026.
• Price per million tokens: GLM-5.3 $1.40 / $4.40, cached input $0.26; Flash $0.15 / $0.50.
• Long agent tasks: GLM-5.3 38.9%, Flash 37.4%; Claude Opus 5.5 65.2%, Sonnet 5.5 64.1% (Vals AI).
• Bug fixes: GLM-5.3 95.4% on Vals AI's SWE-bench Verified, sixth.
• Licenses: GLM-5, 5.2 and 5.3-Flash are MIT; GLM-5.3 has its own license with a security review for the very largest AI services.

Sources: Z.ai's release notes, pricing and Coding Plan docs, the GLM-5.3 and GLM-5.3-Flash model cards, Vals AI and LMArena, opened October 29, 2026.

What is GLM 5?

Zhipu AI's fifth-generation models, sold internationally under the Z.ai name. The family so far:

• GLM-5 (February 12, 2026): 744 billion parameters, 40 billion active, about 200K context, MIT.
• GLM-5.1 (April 7) and GLM-5.2 (June 16): GLM-5.2 grew to 753 billion parameters and a 1M-token context, still MIT.
• GLM-5.3 (August 18): "uses the same base model as GLM-5.2," with a 1M context and up to 128K tokens of output. Z.ai doesn't publish its active parameters.
• GLM-5.3-Flash (August 26): "320B total parameters and just 18B active parameters," MIT.

All of them are downloadable from Hugging Face. At these sizes they need a multi-GPU server; Ollama offers GLM-5.3 only as a cloud model.

How good is GLM 5 at coding?

Independent results:

• Vals AI, Terminal-Bench 4.0 (long agent tasks, October 29): GLM-5.3 38.89% (15th of 58), GLM-5.3-Flash 37.37% (18th). Claude Opus 5.5 65.15% and Sonnet 5.5 64.14% lead.
• Vals AI, SWE-bench Verified (archived September 1): GLM-5.3 95.40% (sixth), Flash 92.00% (tenth). The leaders there are within a few points of each other.
• LMArena WebDev (people voting on web apps, October 27): GLM-5.3 at max effort 27th (1623), Flash 32nd (1607). Opus 5.5 is first (1810).

So GLM-5.3 is near the top on well-defined bug fixes and well behind Claude on long, open-ended agent work. Among models you can download today it's the strongest on Terminal-Bench, level with Qwen 3.8 Max (the best open-weight models for coding), and Flash scores almost as well for a tenth of the price.

How much does GLM 5 cost?

Z.ai's international API, per million tokens: input, cached input, output (Z.ai):

• GLM-5.3, 5.2 and 5.1: $1.40, $0.26, $4.40.
• GLM-5.3-FlashX: $0.37, $0.075, $1.25.
• GLM-5.3-Flash: $0.15, $0.03, $0.50.
• GLM-5: $1.00, $0.20, $3.20.

Storing cached input is free for now. None of the GLM-5 models has a free API tier; Z.ai's free models are older GLM-4 Flash versions. For comparison, Claude Sonnet 5.5 is $2 and $10 and Haiku 5.5 $0.10 and $0.50 (Claude API pricing).

What is the GLM Coding Plan?

A subscription for using GLM in coding agents instead of paying per token (Z.ai). Since July 30, 2026 it's counted in credits:

• Lite: 2,000 credits per five hours, 10,000 a week.
• Pro: 12,000 per five hours, 60,000 a week.
• Max: 28,000 per five hours, 140,000 a week.

Z.ai says plans start "at just 18 USD per month"; its own subscribe page loads the higher tiers' prices only in the browser. Requests outside peak hours, 14:00 to 18:00 Singapore time on weekdays, use half the credits. The plan runs GLM-5.3 and 5.3-Flash, and older model names are sent to 5.3.

It works only in supported tools: Z.ai's own ZCode, Claude Code, Codex, OpenCode, Cursor, Cline, Kilo Code, Roo Code, Crush, Goose and others. Account sharing is banned and refunds "are not supported."

Can you use GLM with Claude Code?

Yes. Claude Code is on the Coding Plan's list of supported tools, and Z.ai documents setting it up to send requests to GLM instead of Anthropic. It bills your Z.ai plan or API balance, not a Claude subscription, and Anthropic doesn't support the setup. How Claude Code itself is billed: Claude Code pricing.

What do GLM's licenses allow?

GLM-5, GLM-5.2 and GLM-5.3-Flash are MIT-licensed: use, change and sell them freely. GLM-5.3 ships under its own GLM-5.3 License, which is permissive except that a company running a model-as-a-service business with more than $10 billion in revenue over 12 months "must pass Z.AI's security review" first (Hugging Face).

Where does GLM API data go?

Z.ai's API is provided by a Singapore company, and its data processing terms say customer data "is generally processed in Singapore" and that it does "not store any of the content the Customer or its End Users provide or generate" (Z.ai). The terms don't say anything explicit about training. Running the open weights yourself keeps code off Z.ai's servers entirely.

GLM 5 vs Claude

• Hard agentic coding: Claude Opus 5.5 or Sonnet 5.5, ahead by more than 25 points on Terminal-Bench 4.0.
• Cheap, high-volume work: GLM-5.3-Flash at $0.15 and $0.50, close to GLM-5.3's scores; Claude Haiku 5.5 costs less per token and scores 36.4% on the same test.
• Self-hosting: GLM-5.3-Flash, MIT-licensed. Claude can't be downloaded.

Other open models compared: DeepSeek V4 vs Claude, Kimi K3 vs Claude, Qwen Coder.

Claude, GPT, Grok and Muse in one room

Poly (usepoly.co) doesn't offer GLM. It runs Claude, GPT, Grok and Muse for a team in one shared browser room, each member picking their own model, with a free model from each lab. What is Poly?