Add Z.ai GLM-5.3 and GLM-5.3-Flash to the model registry

[adambalogh]

Summary

Z.ai shipped two new models since the registry was last updated:

• GLM-5.3 (released 2026-08-14) — Z.ai describes it as a post-training-only
  refresh of GLM-5.2 (same base model, no new pretraining run), with sharply
  better coding/agent numbers. Same per-token sticker as GLM-5.2: $1.40/$4.40
  per MTok.
• GLM-5.3-Flash (released 2026-08-26) — a new low-cost, natively
  multimodal (vision) sibling. $0.15/$0.50 per MTok.

Both are registered directly against Z.ai's own OpenAI-compatible Model API
(provider="zai"), using the zai_http_client / ZAI_BASE_URL wiring already
in place for glm-image. This is a deliberate deviation from how GLM-5.2 is
served: GLM-5.2 routes through a BytePlus ModelArk deployment endpoint
(ep-…), but provisioning one of those requires manual, console-only setup in
the BytePlus Ark Console (no creation API — see chat-api#261's "Still blocked
on" section for the same limitation) that I have no way to do here. Z.ai's own
API already serves both models directly, needs no new provisioning, and is an
already-integrated provider, so that's what these route through.

Pricing / parameter citations

• Pricing (both models): Z.ai's own pricing docs (docs.z.ai/guides/overview/pricing),
  cross-checked against aggregator pages (Together AI, OpenRouter, requesty.ai).
• Parameters: Z.ai's GLM-5.3 guide (docs.z.ai/guides/llm/glm-5.3) — thinking
  is now mandatory (thinking.type: "disabled" no longer works, reasoning_effort
  must be low/high/max), but temperature is still accepted (Z.ai's
  documented default is temperature=1.0, top_p=0.95), so no
  supports_temperature=False / force_temperature flag is needed on either
  model.

Changes

• tee_gateway/model_registry.py: GLM_5_3 and GLM_5_3_FLASH enum entries
  under the Z.ai section, plus _MODEL_LOOKUP aliases (glm-5.3,
  glm-5.3-flash). GLM_5_2 is untouched — it keeps routing through
  ModelArk.
• tests/test_pricing.py: resolve + cost tests for both models, mirroring the
  existing GLM-5.2/HY3 tests.
• CLAUDE.md / README.md: updated the Z.ai provider row.

Testing

• uv run python -m pytest tests/test_pricing.py -q — 146 passed (was 144).
• make lint — ruff format/check + mypy clean.
• Full suite (tests/ + tee_gateway/test/, excluding tests/test_server.py
  which needs real provider keys in .env): 457 passed, 6 skipped. One
  pre-existing, unrelated failure
  (test_hash_dict_preserves_multimodal_user_content) reproduces identically
  on main before this change.

Companion PR

chat-api PR adding these to the served catalog:
OpenGradient/chat-api#289

---

🤖 Generated with Claude Code

https://claude.ai/code/session_01Qon6s5f24upx8znGandm1w

[adambalogh]

Scheduled model-catalog check (2026-09-14): re-scanned OpenAI, Anthropic, Google, xAI, ByteDance/ModelArk, OpenRouter and Z.ai for releases not yet in the registry. The only gap found is GLM-5.3 / GLM-5.3-Flash — already covered by this PR and its companion, chat-api#289. No other new models to add this run.

Flagging a blocker before either merges: this PR directly conflicts with #157 ("Remove the Z.ai provider"). #157 removes the zai provider, zai_http_client, ZAI_BASE_URL and ZAI_API_KEY wiring entirely on the premise that "Z.ai is no longer used"; this PR re-establishes direct Z.ai routing for two new chat models on that same wiring. Whichever merges first will break the other's premise. Leaving this for a human call on whether Z.ai is being kept or dropped — not resolving it automatically.

---

Generated by Claude Code