GLM-5.3 for Claude Code: the two-model mapping

Z.ai released GLM-5.3 on 2026-08-14. It uses the same base model as GLM-5.2, and Z.ai says every gain comes from post-training. It takes text in and returns text out. GLM-5.3-Flash, released on 2026-08-26, is the multimodal member of the family.

The claude-multi GLM template is now two models. The current template maps glm-5.3[1m] to primary, opus, sonnet, and fable, and glm-5.3-flash[1m] to haiku. That is the mapping in Z.ai’s Claude Code example. Version 0.11.0 put GLM-5.3 in both the opus and sonnet slots. Version 0.12.0 moved sonnet to Flash. Version 0.13.0 moved Flash to haiku and returned GLM-5.3 to sonnet. Version 0.13.2 added the Fable slot.

Z.ai no longer serves glm-5.2 or glm-5.1 on the Coding Plan endpoint. Requests for either ID route to glm-5.3. Requests for glm-4.7 route to glm-5.3-flash. The plan lists GLM-5.3 and GLM-5.3-Flash as its models, so those are the only two IDs the template needs.

The details are in Z.ai’s GLM-5.3 announcement, the GLM-5.3 docs, and the GLM-5.3-Flash announcement.

What changes in the template

The output cap stays at 128,000 tokens, up from 64,000 in 0.10.0. Z.ai documents 128K max output for GLM-5.3 and GLM-5.3-Flash, and the template sets CLAUDE_CODE_MAX_OUTPUT_TOKENS to 128000. The endpoint stays at https://api.z.ai/api/anthropic.

The haiku slot gets Flash. Z.ai says Flash carries three times the plan quota of GLM-5.3 and accepts image input, which matters for the short calls and subagents that read screenshots. Primary, opus, sonnet, and fable run on GLM-5.3 for long refactors and hard debugging.

Thinking stays on, and it is no longer optional. GLM-5.3 rejects thinking.type: "disabled". reasoning_effort accepts low, high, or max, defaults to max, and Z.ai recommends max for coding. The template sets CLAUDE_CODE_EFFORT_LEVEL to high, adds CLAUDE_CODE_ALWAYS_ENABLE_EFFORT=1 so Claude Code sends it for the custom model ID, and uses 8,000 thinking tokens as a client-side hint. Thinking bills as output at the 24x multiplier, so max is a quota decision as much as a quality one.

Context is simpler now. Both models support 1M, so the template adds the [1m] suffix to both and sets CLAUDE_CODE_AUTO_COMPACT_WINDOW to 1000000. Unsuffixed glm-5.3 defaults to roughly 200K in Claude Code. One global window fits both slots.

If you already have a GLM instance, run claude-multi doctor fix to sync the new template. It preserves your API key and any custom env vars.

Credits and plans

The Coding Plan meters usage in credits:

credits = (input x input_mult + cached_input x cached_mult + output x output_mult) / 10,000

Z.ai’s table lists these two models. Older IDs do not have their own line: glm-5.2 and glm-5.1 route to GLM-5.3, and glm-4.7 routes to Flash.

GLM-5.3 consumes exactly three times the credits of Flash: 6.9 against 2.3 on input and 24 against 8 on output. Flash also carries three times the plan quota, so the two effects compound if you use it for everyday work.

Off-peak calls cost 50% of the standard credit rate. Peak is Monday to Friday, 14:00-18:00 UTC+8. Nights, weekends, and weekday mornings are off-peak. Plans start at $18 per month for Lite, with 20% off quarterly billing and 30% off yearly billing. Z.ai has the current Pro and Max prices on its plan page.

This credit system replaced the older peak/off-peak multiplier plan, which Z.ai discontinued for new users on 2026-07-30. The plan update notice has the migration details.

Benchmarks

Every number below is vendor-published from Z.ai’s announcement tables. The comparison labels are Z.ai’s. There is no SWE-bench Verified score for GLM-5.3 in the announcement, so this is the available data, not an independent result.

On Z.ai’s in-house Code Bench, GLM-5.3 completed 34.5% of tasks at roughly 75K output tokens per task, against GLM-5.2’s 23.4% at roughly 96K. At High effort it posted 31.4% at about 50K tokens, ahead of Claude Opus 4.8’s 29.5% at about 120K. It is still behind Claude Fable 5, which reaches 39.5% at Max effort. Z.ai claims open-source SOTA on Terminal-Bench 3.0 and ALE-CLI, and the best result on CyberGym.

The baseline matters. GLM-5.3 shares its base model with GLM-5.2, so all of this comes from post-training, and none of it is independently verified.

Availability and upgrades

At launch, GLM-5.3 was available on the Coding Plan and ZCode only. The pay-as-you-go API followed, and Z.ai now lists GLM-5.3 on its model API. claude-multi uses the Coding Plan endpoint.

If you have an older instance, nothing breaks today: the endpoint routes glm-5.2 and glm-5.1 to glm-5.3. But your env may still name the old models and cap output at 64000. Run claude-multi doctor fix to sync the current template and get Flash in the haiku slot.

Getting started

claude-multi add glm

Paste your Z.ai API key when prompted. The template fills in the endpoint, the model mappings, thinking, and the 128K output cap. The instance is ready.

See the GLM provider page for the full spec table, the GLM-5.2 post for the older mapping, or compare against Kimi K3.1 Pro and MiniMax M3.1 Flash. The CHANGELOG has the full diff.