chore(sync): update core ZenMux model families

[Alcatraz-Zhang]

Summary

• reconcile ZenMux Claude, GPT, Gemini, Kimi, and DeepSeek entries with the live models-page allowlist
• factor third-party routes through lab metadata, adding only the nine missing complete lab definitions
• refresh USD per-million-token prices and context tiers while omitting per-second/per-request pricing and preserving complex conditional prices
• align reasoning controls with current first-party peers and documented ZenMux wire paths

Sources

• https://zenmux.ai/models
• https://zenmux.ai/docs/api/openai/openai-list-models.html
• https://zenmux.ai/docs/api/anthropic/anthropic-list-models.html
• https://zenmux.ai/docs/api/vertexai/google-list-models.html
• https://zenmux.ai/docs/guide/advanced/reasoning.html

Validation

• bun validate
• bun run compare:migrations
• git diff --check dev...HEAD --
• live allowlist reconciliation: 84 remote / 84 local, 0 missing, 0 extra
• independent pricing/base-model/reasoning/override audit: 0 issues
• 108 changed files, below the reviewer 300-file limit

Test note

The broad bun test command is not green on the current dev baseline because of missing SDK effect/snapshot artifacts and existing unrelated weights-link, Hyper, and DeepInfra expectation failures. The targeted open-weight check confirms neither newly added DeepSeek metadata entry remains in its missing-weights list.

[github-actions]

Action items

• [high] [violation] providers/zenmux/models/google/gemini-3-pro-image.toml:3 - Check: reasoning = true models must keep correct host reasoning_options (lab + first-party peer). Why: Lab google/gemini-3-pro-image and first-party providers/google/models/gemini-3-pro-image.toml are reasoning models with effort low/high. This route forces reasoning = false, so the catalog drops caller controls and mislabels the model. Action: Remove reasoning = false and author the ZenMux-supported effort set (or [] only if this host truly has no control), matching lab/first-party baseline.
• [high] [violation] providers/zenmux/models/google/gemini-2.5-flash-image.toml:3 - Check: Do not override a lab reasoner to non-reasoning without a real host delta. Why: Lab and first-party Google keep reasoning = true with reasoning_options = [] (always-on / no control). Setting reasoning = false contradicts that and breaks the resolved capability. Action: Drop reasoning = false; use reasoning_options = [] if ZenMux exposes no control, or the host’s real controls if present.
• [high] [violation] providers/zenmux/models/google/gemma-4-31b-it.toml:2 - Check: Relay of a lab reasoner must not invent reasoning = false / omit cost. Why: Lab gemma-4-31b-it is reasoning = true; first-party Google and OpenRouter keep reasoning with a toggle. This file sets reasoning = false and has no [cost], so pricing and reasoning controls are both wrong/missing. Action: Restore reasoning + toggle (or documented host controls), add USD/MTok cost from ZenMux pricing, and keep only real deltas.
• [high] [violation] providers/zenmux/models/google/gemma-4-26b-a4b-it.toml:2 - Check: Same Gemma reasoner baseline as peers. Why: Lab is reasoning; OpenRouter uses toggle. Forcing reasoning = false misrepresents the model on this relay. Action: Align with lab/peer reasoning controls instead of disabling reasoning.
• [high] [violation] providers/zenmux/models/deepseek/deepseek-v4.1-flash.toml:1 - Check: Provider models need host cost (and DeepSeek V4 surface fields when applicable). Why: New paid DeepSeek V4.1 Flash entry has only base_model + reasoning options—no [cost], and no interleaved.reasoning_content that ZenMux V4 peers and first-party DeepSeek use. Action: Add ZenMux USD/MTok pricing and the same interleaved/reasoning surface as the other ZenMux DeepSeek V4 routes unless this ID is intentionally free/unpriced.
• [high] [violation] providers/zenmux/models/openai/gpt-transcribe.toml:1 - Check: Resolved provider models should author cost unless request-only / no public price. Why: File is only base_model = "openai/gpt-transcribe" with no cost or other host fields; a catalogued API model without price is incomplete relative to AGENTS provider requirements and the PR’s pricing refresh. Action: Add ZenMux pricing, or document in a leading comment that this route is intentionally unpriced/request-only.
• [medium] [violation] providers/zenmux/models/openai/chat-latest.toml:2 - Check: reasoning override must match the underlying model. Why: base_model = "openai/gpt-5.5-instant" is reasoning = true in lab metadata, but this route sets reasoning = false (and drops reasoning_options). That is a hard capability flip, not a display rename. Action: Keep inherited reasoning + host effort controls, or point base_model at a true non-reasoning lab identity if Chat Latest is a different product.
• [medium] [possible mistake] providers/zenmux/models/deepseek/deepseek-v3.2.toml:17 - Check: Limit overrides must be real host deltas. Why: Lab deepseek-v3.2 output is 64_000; this sets output = 8_000, an 8× cut that looks like a bad default rather than documented ZenMux max. Action: Verify against ZenMux model docs/API and restore 64_000 (or the true served max) if 8k is wrong.
• [medium] [possible mistake] providers/zenmux/models/openai/text-embedding-3-large.toml:9 - Check: Embedding limit.output is vector dimensions, not a second context window. Why: New lab metadata and first-party OpenAI use output = 3_072 for large; this override sets context = 8_192 / output = 8_192, which misstates embedding size. Same pattern on text-embedding-3-small (1_536 expected). Action: Match lab/OpenAI dimension limits (or omit identical values) and only override real ZenMux deltas.
• [medium] [possible mistake] providers/zenmux/models/anthropic/claude-sonnet-5.toml:24 - Check: Output-limit override vs lab. Why: Lab/first-party Sonnet 5 use output = 128_000; this forces 64_000 without a cited host cap. Action: Confirm ZenMux’s max output; remove the override if it matches the lab.
• [medium] [possible mistake] models/google/veo-3.1-generate-001.toml:12 - Check: Date fields must be YYYY-MM or YYYY-MM-DD. Why: last_updated = "2026-01" is valid month form, but paired provider Veo files omit cost while only bumping context to 100_000 against lab 1_024—inconsistent and hard to trust without a mapped pricing/limit citation. Action: Normalize dates if a full day is known; add cost or an explicit top-of-file note that Veo is non-token / intentionally unpriced; justify the 100k context override from ZenMux docs.
• [low] [possible mistake] providers/zenmux/models/deepseek/deepseek-chat-v3.1.toml:3 - Check: Hybrid DeepSeek V3.1 reasoning flag. Why: Lab deepseek-v3.1 is reasoning = true (thinking/non-thinking). This route sets reasoning = false under a generic “DeepSeek V3.1” name, which may be correct for a non-thinking SKU but is easy to misread. Action: Confirm this ID is non-thinking-only; if not, keep reasoning + toggle; if yes, say so in the display name/comment.
• [low] [possible mistake] providers/zenmux/models/openai/gpt-5-codex.toml:4 - Check: Override-only attachment must differ from base. Why: Lab gpt-5-codex has attachment = false; this sets attachment = true (and similar image/pdf modality restates appear on several GPT routes). That may be a real ZenMux delta or accidental copy. Action: Verify multimodal input on ZenMux for each overridden GPT id; drop redundant fields that match the lab.

[Alcatraz-Zhang]

Addressed the automated review in 6921d68. Applied the two substantive fixes: embedding output dimensions now inherit 3072/1536 from lab metadata, and DeepSeek V4.1 Flash now records its reasoning side channel plus the discounted base token rates (with a time-band limitation note). Added source comments for intentional ZenMux host deltas and for non-token/unavailable pricing instead of changing them to contradict the live ZenMux protocol catalogs. Re-ran bun validate, compare:migrations, diff check, and the exact live/local 84-model set comparison; all pass.

[github-actions]

Action items

• [high] [possible mistake] providers/zenmux/models/moonshotai/kimi-k2.8-preview.toml:19 - Check: Provider limit overrides must be real host deltas, not context-window copies. Why: The file sets limit.output = 1_048_576, equal to the model context. Lab metadata has no output cap, and the established Kimi Code peer for this model uses 32_768 output. Treating max completion as the full 1M context is very likely wrong and will mislead clients. Action: Verify ZenMux’s actual max output for kimi-k2.8-preview from the host model list/docs and set limit.output to that value (or omit if it matches an inherited/default you can document).
• [high] [possible mistake] providers/zenmux/models/moonshotai/kimi-k2.8-preview.toml:1 - Check: Relay reasoners that expose thinking side channels should author interleaved like same-host peers. Why: This route adds toggle + effort and cites ZenMux reasoning docs, but omits [interleaved]. On the same provider, kimi-k2.6 and kimi-k3 set field = "reasoning_content", and the Kimi Code peer for K2.8 also does. Missing interleaved understates the host response shape. Action: Add [interleaved] field = "reasoning_content" if ZenMux returns that field, or document why this route differs.
• [high] [possible mistake] providers/zenmux/models/google/gemini-3.1-flash-lite-image.toml:15 - Check: Output limits after base_model must match the host, not invent a middle value. Why: Lab metadata is output = 4_096; Google first-party overrides to 65_536. ZenMux sets 32_768, which matches neither baseline and looks like a confused merge of image/token limits. Action: Confirm ZenMux’s listed max output for this route and align limit.output (or drop the override if the lab value is correct for this host).
• [medium] [possible mistake] providers/zenmux/models/anthropic/claude-opus-4.5.toml:19 - Check: Output-limit overrides need host evidence when they cut lab defaults. Why: Lab/base_model resolve to 64_000 output; this PR overrides to 32_000 with no leading comment explaining a ZenMux-only cap (unlike the documented Sonnet 5 / DeepSeek V3.2 overrides). Action: Cite the ZenMux model-list value that caps Opus 4.5 at 32k, or remove the override.
• [medium] [possible mistake] providers/zenmux/models/openai/gpt-5-pro.toml:13 - Check: Provider limit.output must not silently shrink lab/first-party limits. Why: Lab and OpenAI first-party use 272_000 output; ZenMux overrides to 128_000 without a source comment. Action: Verify against ZenMux’s OpenAI-compatible model list and either keep with a top-of-file citation or drop the override.
• [medium] [possible mistake] providers/zenmux/models/openai/text-embedding-3-large.toml:9 - Check: Embedding context should match lab/first-party token limits. Why: New lab entries and OpenAI’s provider files use context = 8_191 / output = 3_072 (large) and 1_536 (small). ZenMux overrides context to 8_192 and drops the large/small output distinction, which is a common off-by-one vs the established OpenAI limit. Action: Prefer inheriting lab limits unless ZenMux documents 8192; if overriding, cite the host list and keep correct embedding output dims.
• [medium] [possible mistake] providers/zenmux/models/google/gemini-embedding-2.toml:7 - Check: Embedding limit.output should not zero out meaningful lab dimensions without cause. Why: Lab metadata has output = 3_072 (max embedding dims). ZenMux sets output = 0 while Google first-party uses a non-zero sentinel. Zero output is inconsistent with the lab fact and peer pattern. Action: Inherit lab output dims, match Google’s host convention, or document why ZenMux needs 0.
• [medium] [possible mistake] providers/zenmux/models/google/gemini-3.5-flash.toml:12 - Check: Do not invent free cache_write pricing. Why: This PR adds cache_write = 0 while Google first-party omits cache write for the same model and other ZenMux Gemini 3.x Flash entries do not claim free writes. A zero write price is almost certainly incorrect or incomplete. Action: Remove cache_write = 0 unless ZenMux pricing explicitly lists free cache writes; if cache is unsupported, omit the key.
• [low] [possible mistake] providers/zenmux/models/deepseek/deepseek-v4.1-flash.toml:20 - Check: Pricing should be clearly mapped when it diverges from first-party. Why: DeepSeek first-party Flash is 0.15 / 0.6 input/output (off-peak noted there); ZenMux records 0.075 / 0.3 with a time-of-day caveat. That may be correct for a discounted band, but the values are half the lab catalog and need a direct pricing mapping in the PR/file header. Action: In the leading comment, state which ZenMux price band/API field produced 0.075/0.3/0.0015 (and whether peak rates are omitted intentionally).

[Alcatraz-Zhang]

Addressed the second automated review in 0ff0622. Added Kimi K2.8 interleaved.reasoning_content and removed the invalid Gemini Embedding 2 output=0 override. The remaining host limits/prices are exact 2026-09-19 ZenMux catalog values, now recorded in leading source comments: K2.8 max_completion_tokens=1048576, Gemini 3.1 Flash-Lite Image=32768, Claude Opus 4.5=32000, GPT-5 Pro=128000, text-embedding context_length=8192, and Gemini 3.5 Flash 1h cache-write=USD 0/MTok. DeepSeek V4.1 now maps the precise discounted and higher time bands. Validation, migration comparison, diff check, and 84/84 live-set audit pass.

[github-actions]

Action items

• [medium] [possible mistake] providers/zenmux/models/deepseek/deepseek-v4-flash.toml:4 - Check: Override-only provider files must keep only real deltas after base_model, including display name when the lab id is a dated snapshot. Why: The route id is still deepseek-v4-flash, but base_model is now deepseek/deepseek-v4-flash-0731, whose lab name is DeepSeek V4 Flash 0731. With no local name, the catalog will surface the dated snapshot name for the undated host id. DeepSeek’s own first-party alias also no longer points at 0731 (it uses deepseek-v4.1-flash). Action: Confirm whether ZenMux still serves the 0731 snapshot; if the undated id is current Flash/V4.1, retarget base_model accordingly and/or set name = "DeepSeek V4 Flash" (and align effort with the chosen lab peer).
• [medium] [possible mistake] providers/zenmux/models/deepseek/deepseek-v4-pro.toml:4 - Check: Same base_model/name inheritance rule for dated lab targets. Why: base_model = "deepseek/deepseek-v4-pro-0813" inherits lab name DeepSeek V4 Pro 0813 for host id deepseek-v4-pro, while first-party DeepSeek overrides that snapshot to name = "DeepSeek V4 Pro". Action: Keep the 0813 lab target if that is what ZenMux hosts, but add the undated display name (or point at the undated lab entry if that is the intended identity).
• [medium] [possible mistake] providers/zenmux/models/google/gemini-3-pro-image.toml:5 - Check: Relay reasoning should follow the lab + first-party control surface unless this host truly disables thinking. Why: Lab google/gemini-3-pro-image is reasoning = true, and Google first-party authors effort low/high. This PR forces reasoning = false, which drops reasoning controls for a model peers treat as a reasoner. Action: Verify ZenMux’s live route; if thinking is available, restore reasoning inheritance and author matching reasoning_options (with wire comment). Only keep reasoning = false if the host API actually disables thinking.
• [medium] [possible mistake] providers/zenmux/models/google/gemma-4-26b-a4b-it.toml:5 - Check: Do not disable reasoning on a relay when the lab/first-party expose a real control. Why: Lab and Google first-party Gemma 4 IT entries are reasoning = true with toggle. Forcing reasoning = false on ZenMux contradicts that baseline. Same pattern on gemma-4-31b-it.toml. Action: Confirm ZenMux’s protocol flags; if thinking can be toggled, drop the false override and set reasoning_options (at least toggle with a leading wire comment). Keep the override only with host-proof that reasoning is unavailable.
• [medium] [possible mistake] providers/zenmux/models/google/gemini-2.5-flash-image.toml:4 - Check: reasoning = false vs first-party peer shape for the same model. Why: Lab is a reasoner; Google first-party keeps reasoning = true with reasoning_options = [] (always-on / no caller control). ZenMux instead flips reasoning = false, which is a stronger semantic change than “no controls.” Action: Prefer reasoning_options = [] if the model still reasons without caller control; use reasoning = false only if ZenMux truly serves a non-reasoning image route.
• [medium] [possible mistake] providers/zenmux/models/moonshotai/kimi-k2.7-code.toml:1 - Check: Provider-specific interleaved should match this host’s reasoning side channel when peers/lab routes expose it. Why: This PR adds interleaved.field = "reasoning_content" for Kimi K2.6 / K2.8 / K3 on ZenMux, and Moonshot first-party K2.7 Code (and HighSpeed) also set that field, but ZenMux kimi-k2.7-code and kimi-k2.7-code-highspeed still omit it while keeping reasoning_options = []. Action: Add the same interleaved block on both K2.7 Code routes if ZenMux returns reasoning_content, or document why these two hosts differ from the other ZenMux Kimi reasoners.

[Alcatraz-Zhang]

Addressed the third automated review in aedca5a. Added interleaved.reasoning_content to both ZenMux Kimi K2.7 Code routes. Confirmed the undated ZenMux DeepSeek slugs explicitly display the 0731/0813 snapshots, so their dated base models/names are intentional and now documented. Strengthened the four Google host-delta comments with exact live protocol evidence: page supports_reasoning=0, Google thinking=false, and where present OpenAI/Anthropic capabilities.reasoning=false. I did not flip those routes to reasoning=true because that would contradict all live ZenMux list APIs. Validation, migration comparison, diff check, and 84/84 live-set audit pass.

[github-actions]

No actionable findings.

[github-actions]

No actionable findings.