Move GLM-5.3 onto the live API, add Claude Opus 5, align the model lists

[hiyouga]

GLM-5.3's API launched, so the model moves off the pre-launch footing it shipped on in 0.4.2. Bundles two smaller model-list fixes that were in the same files.

GLM-5.3 general availability

llmsdk_docs/glm5_3/ was re-snapshotted as raw markdown from docs.z.ai, replacing the hand-written summary taken from the Chinese pre-launch pages at docs.bigmodel.cn. The snapshot grew from 2 pages to 7 — chat-completion.md, concept-param.md, migrate-to-glm-new.md, stream-tool.md, and thinking-mode.md all changed for 5.3 upstream, so the folder is now self-contained rather than deferring to glm5_2/. stream-tool.md came back byte-identical to the 5.2 copy but is included so the snapshot stands alone.

No client change was needed. The launched docs confirm the contract the client was already written against: forced thinking, reasoning_effort restricted to low/high/max, default max. concept-param.md adds one fact the pre-launch snapshot lacked — max_tokens defaults to 65536 with a 131072 maximum.

Pricing is now published, so glm-5.3 gets $1.4 input / $0.26 cached / $4.4 output per MTok — the same rates as 5.2 and 5.1.

Claude Opus 5 and gpt-5.6-sol

claude-opus-5 was missing from the registry entirely; added for both the official API and OpenRouter at $5 / $0.50 / $25, from Anthropic's pricing page (the $0.50 is the cache-hit rate, matching how the other Claude entries store cached_tokens). Routes through the existing claude-5 client.

The official GPT-5.6 entry was the bare gpt-5.6 alias; it now names gpt-5.6-sol, matching the OpenRouter entry and its terra/luna siblings. Pure rename — OpenAI's published pricing confirms sol is $5 / $0.50 / $30, exactly what the entry already carried. The bare gpt-5.6 spelling still routes.

Playground model list (closes #178)

The dropdown listed models from several generations back (gpt-5.5, gemini-3.5-flash, claude-opus-4-7, kimi-k2.6, glm-5.1, …). It now lists exactly the 11 official-provider AVAILABLE_MODELS entries, verified programmatically: same model set, and each dropdown entry constructs the same client class the E2E entry does.

text-embedding-3-large was the one model whose id cannot route on its own, so options can now declare data-client-type, read by a new getSelectedClientType(). Default model moves gpt-5.5 → gpt-5.6-luna, in both the rendered dropdown and the server-side fallback.

Verification

Live GLM-5.3 capture under api_captures/glm5_3/ (git-ignored): a streaming tool-call round plus the replay round that re-sends the assistant turn. Round 2 was accepted with content + reasoning_content + tool_calls and the tool result — exactly the shape the client emits and replays — and reported cached_tokens: 128. clear_thinking: false is still accepted on 5.3, which was the open question from the pre-launch implementation.

• Python E2E -k glm-5.3: 8 passed, 6 skipped
• TypeScript E2E -t glm-5.3: 14 passed
• Offline suites, ruff, eslint, tsc: green in both languages

Z.AI throttled this key hard during testing (429 code 1302; the capture needed ~15 min of backoff before its first request landed). Four Python E2E tests failed on 429 in one run and all four passed on retry — no assertion failures at any point.

One protocol mode stays on glm-5.2

Probing the three Z.AI protocol-client modes turned up a real incompatibility:

mode	glm-5.3
openai-chat	works
openai-responses	works
ant-messages	400, code 1210

Z.AI's Anthropic-compatible gateway defaults to thinking-disabled when a request carries no thinking config, and glm-5.3 rejects that: "This model always engages in thinking and cannot be disabled; please use low, high, or max". Confirmed it is specific to 5.3 and to the missing config — glm-5.2 works there with no thinking level, and glm-5.3 works there with thinking_level LOW/HIGH/MAX; only NONE and "unset" fail.

The generic ant_messages client is model-agnostic by design and correctly omits the thinking config when the caller asks for nothing, so teaching it about GLM-5.3 would be the wrong fix. That E2E entry stays on glm-5.2 instead, with the reason in a comment.

Live E2E rate-limit retry

CI failed twice on Z.AI 429s (code 1302) against glm-5.3 — real, not noise: no other rate-limit failure appears in the last 25 Python runs, and the one prior failure was a genuine UnboundLocalError. It is not a concurrency bug either; conftest.py already pins each model's tests to one xdist worker. The heavy E2E bodies simply cross a throttling window that a plain sequential probe does not (24/24 clean locally against both 5.2 and 5.3).

Both suites now retry only rate-limited responses:

• Python — pytest-rerunfailures added to the dev extra; the workflow gained --reruns 5 --reruns-delay 30 --only-rerun RateLimitError --only-rerun 429. test_client.py is untouched; the policy sits beside the existing -n 16 --dist loadgroup.
• TypeScript — the retry lives in the single modelTest wrapper every model test already funnels through, on the same 5 × 30s policy. No test body, assertion, prompt, or helper changed, and nothing branches on a model name. Each test's jest deadline is widened to cover the retry budget.

Verified it does not mask real bugs: a throwaway pair of tests — one raising RateLimitError, one a plain assert 1 == 2 — showed the rate-limit case running 3 times and passing while the assertion case ran once and failed.

Green runs since then report 296 passed, 113 skipped, 2 rerun and 296 passed, 113 skipped, 6 rerun, so the retry demonstrably fires and clears rather than the throttle merely being absent.

This masks sustained throttling rather than curing it. If secrets.ZAI_API_KEY is a lower-tier key than a normal dev key, raising its quota remains the better root-cause fix.

OpenRouter moves to the OpenAI Responses protocol

Third-party gateways now prefer openai-responses. All 17 OpenRouter entries that used the generic openai-chat client were switched, each verified against the live endpoint one model at a time rather than assumed:

anthropic/claude-fable-5, anthropic/claude-opus-5, anthropic/claude-opus-4.8, anthropic/claude-opus-4.7, anthropic/claude-sonnet-5, google/gemini-3.5-flash, minimax/minimax-m3, nvidia/nemotron-3-ultra-550b-a55b:free, openai/gpt-5.6-sol, openai/gpt-5.6-terra, openai/gpt-5.6-luna, openai/gpt-5.5, qwen/qwen3.6-35b-a3b, stepfun/step-3.7-flash, tencent/hy3, x-ai/grok-4.5, xiaomi/mimo-v2.5 — 17/17 served.

OpenRouter's docs don't state Responses coverage and an unauthenticated probe 401s before validating the model, so this was worth checking: the long tail (nvidia, stepfun, tencent, x-ai, xiaomi, qwen) was the risk, and it all works. The probe re-ran any failure over openai-chat as a control so a temporarily-down model could not be misread as an unsupported protocol; that path never had to fire.

Left alone deliberately:

• SiliconFlow's 3 entries stay on openai-chat — it serves Chat Completions only.
• OpenRouter entries with a dedicated client (glm-5.3, kimi-k3, kimi-k2.6, deepseek-v4, openai-embedding) keep it. The preference is about which generic protocol to reach for, not about giving up model-specific handling.

README.md and both skills/*/reference/models.md previously told readers to use openai-chat for OpenRouter, which would now contradict the registry; they state the preference instead. The OpenRouter qwen3.6 E2E entry moved to openai-responses so the suite exercises what the registry recommends, while the SiliconFlow one stays on chat.

README examples follow the registry

section	was	now
Supported-models table	claude-sonnet-5	claude-opus-5
Anthropic example + heading	Claude Sonnet 5	Claude Opus 5
OpenRouter example + heading	GLM-5.1	GLM-5.3

OpenRouter GPT-5.6 repricing

The three OpenRouter GPT-5.6 entries were stale against GET https://openrouter.ai/api/v1/models and are now corrected:

entry	was	now
openai/gpt-5.6-sol	$5 / $0.50 / $30	$2.5 / $0.25 / $15
openai/gpt-5.6-terra	$1 / $0.10 / $6	$2 / $0.20 / $12
openai/gpt-5.6-luna	$0.1 / $0.01 / $0.6	$0.2 / $0.02 / $1.2

sol had been carrying OpenAI's list price rather than OpenRouter's; terra and luna had been carrying their :batch rates. The official entries were already correct and are unchanged — note that gpt-5.6-sol (official, $5/$30) and openai/gpt-5.6-sol (OpenRouter, $2.5/$15) genuinely differ. openai/gpt-5.5 was checked and already matches.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.