[sashapervykh]

Summary

• Point Bothub api / doc at bothub.chat
• Refresh ECO USD prices (ELITE API rate, 2 d.p.) for existing Bothub models
• Add current text models: Gemini 3/4, Claude, OpenAI (+ o3), Grok (+ lab xai/grok-4.20-multi-agent), Qwen 3.8 and Qwen3 Coder, Kimi K2.7 Code / K3 / K3.1 Pro, dated DeepSeek V4 and V4.1, GLM 5.3
• Provider catalog is now 62 Bothub model files

Pricing source

https://bothub.chat/models, ECO mode, checked 2026-11-12

v1 changes (Bothub)

Provider

• api: https://openai.bothub.ru/v1 → https://openai.bothub.chat/v1
• doc: https://bothub.ru/models → https://bothub.chat/models

Existing model costs (USD / 1M)

Updated on 2026-11-12

Other

• gpt-5.6-luna: added long-context tiers (cost.tiers / context_over_200k, >272k)
• deepseek-v4-pro-0813: reasoning effort now includes low (none|low|high|max)
• free models (gemma-4-31b-it:free, nemotron-3-ultra-550b-a55b:free): still 0
• New Pro and Fast SKUs that share a lab model publish their own name (for example GPT-5.6 Luna Pro, Claude Opus 5 (Fast), Claude Opus 4.8 (Fast)). gpt-5.5-pro already inherits GPT-5.5 Pro from its lab file.

Notes / assumptions

• grok-4.20 → base_model = "xai/grok-4.20-0309-reasoning" (xAI alias); undated grok-4.20-multi-agent gets a new lab
• Reasoning controls follow the lab set. grok-4.20 and kimi-k2.7-code stay reasoning = { supported = true } with no caller controls. Qwen 3.8 is low|medium|xhigh. grok-4.3 is none|low|medium|high|xhigh. grok-4.5 is low|medium|high|xhigh.
• tools.freeform is set only where a 2026-11-13 chat-completions probe showed the grammar was enforced: text|lark|regex on gpt-5.6-sol, gpt-5.6-terra, gpt-6-astra, gpt-6-luna, gpt-6-sol, gpt-6.1-sol, gpt-6.1-terra; text only on gpt-5.5. Other new GPT files omit it.
• legacy_interleaved = true stays on the 8 models already on origin/dev and is not set on new files.

[github-actions]

Action items

• [high] [violation] providers/bothub/models/deepseek-v4-flash.toml:13 - Check: Relay reasoning baseline must copy lab + same-surface peers (DeepSeek V4 is toggle + low/high/max). Why: Native providers/deepseek/models/deepseek-v4-flash.toml:38 and deepseek-v4-pro.toml:38 are toggle=true, effort=["low","high","max"]; new deepseek-v4-flash.toml/deepseek-v4-pro.toml invent ["none","high","xhigh"] — drops low/max, invents xhigh, drops toggle with no Bothub wire evidence. Same for deepseek-v4-pro-0813.toml change context. Action: Restore low/high/max shape (map off via documented Bothub field with wire comment) or cite Bothub docs/live probe that xhigh works and low/max do not.
• [high] [violation] providers/bothub/models/gemini-2.5-flash.toml:8 - Check: Do not invent GPT-style low/medium/high on labs with different controls. Why: Natives are toggle+budget (providers/google/models/gemini-2.5-flash.toml:20, gemini-2.5-flash-lite.toml:32) and budget-only (gemini-2.5-pro.toml:19); Bothub gemini-2.5-flash.toml, gemini-2.5-flash-lite.toml, gemini-2.5-pro.toml all invent effort=["low","medium","high"] with no Bothub wire evidence. Action: Copy lab controls (toggle/budget if Bothub forwards enable_thinking/thinking_budget-like fields) or provide Bothub docs/live proof that reasoning_effort=low|medium|high has graded effect on these IDs.
• [high] [violation] providers/bothub/models/qwen3.7-flash.toml:11 - Check: Qwen baseline is lab-specific toggle/budget/effort, not generic L/M/H. Why: Natives are toggle+budget with no effort (qwen3.7-flash.toml:21, qwen3.7-plus.toml:34, qwen3.7-max.toml:27) or toggle+["low","medium","xhigh"]+budget (qwen3.8-flash.toml:21, qwen3.8-max-0902.toml:17); Bothub invents ["low","medium","high"] on qwen3.7-flash/plus/max and qwen3.8-flash, and ["minimal","low","medium","high","xhigh"] / ["none","low","medium","xhigh"] on qwen3.8-max-0902/max-prime/3.8-27b — adds minimal/high/none, drops toggle/budget, swaps xhigh/high. Action: Author lab-matching controls for what Bothub actually exposes (with top-of-file wire paths) or cite Bothub docs/live runs proving each invented level.
• [high] [violation] providers/bothub/models/kimi-k2.6.toml:6 - Check: Copy Moonshot native controls. Why: Natives are toggle-only (kimi-k2.6.toml:33), always-on (kimi-k2.7-code.toml:32), and effort=["low","high","max"] with no none (kimi-k3.toml:34, kimi-k3.1-pro.toml:27); Bothub invents ["low","medium","high"] on k2.6/k2.7-code and adds none on k3/k3.1-pro. No Bothub wire proof that medium exists or none disables K3. Action: Restore toggle-only / always-on / low|high|max unless Bothub docs/live probe proves the new sets.
• [high] [violation] providers/bothub/models/glm-5.toml:7 - Check: Copy Zhipu native controls. Why: Natives are toggle-only (glm-5.toml:21, glm-5-turbo.toml:25, glm-5.1.toml:25) or toggle+["high","max"] (glm-5.2.toml:26); Bothub invents ["low","medium","high"] on glm-5/turbo/5.1 and ["none","high","xhigh"] on glm-5.2 — drops toggle/max, invents low/medium/none/xhigh. Action: Restore toggle / toggle+high|max or cite Bothub wire proof for the new levels.
• [high] [violation] providers/bothub/models/grok-4.20.toml:17 - Check: Do not invent controls on always-on lab; lab file must hold host-independent facts. Why: Native providers/xai/models/grok-4.20-0309-reasoning.toml:30 is reasoning={supported=true} with no controls; Bothub invents ["low","medium","high"] and adds [limit] context=2_000_000/output=1_800_000 vs lab models/xai/grok-4.20-0309-reasoning.toml:20-21 (1_000_000/30_000). New lab models/xai/grok-4.20-multi-agent.toml:20-21 also claims 2_000_000/1_800_000 vs xAI grok-4.20-multi-agent-0309.toml:42-44 (1_000_000/30_000) with no first-party citation. Action: Remove invented effort or prove Bothub exposes it; correct lab/provider limits to 1_000_000/30_000 or cite xAI source for 2M/1.8M.
• [high] [violation] providers/bothub/models/gpt-5.5.toml:6 - Check: GPT-5+ additions must assess freeform per API with header evidence. Why: OpenAI natives carry tools.freeform=["text","lark","regex"] (e.g. providers/openai/models/gpt-5.5.toml:21, gpt-6-astra.toml:16); all 17 new Bothub OpenAI files (gpt-5.5/pro, gpt-5.6-*, gpt-6-*/o3) omit freeform and header probe notes, so v1/v2 tool capability is unverified. Action: Probe Bothub chat-completions for custom tools (restrictive regex/lark grammar) and set freeform where enforced, or document that custom tools are rejected; add header evidence comment.
• [medium] [possible mistake] providers/bothub/models/gemini-2.5-flash.toml:12 - Check: Provider [limit] deltas must be real and sourced. Why: gemini-2.5-flash/lite set output=65535 vs lab models/google/gemini-2.5-flash.toml:52 / gemini-2.5-flash-lite.toml:52 (65_536); kimi-k2.6/k2.7-code set 235929 vs lab 262_144; qwen3.7-flash sets 65536 vs lab 131_072; grok-4.5/4.6/4.7 set 450000 vs lab 500_000; headers cite only pricing, no limit source. Action: Fix typo (65535→65536 if not a real cap) and cite Bothub limit source or drop non-deltas to inherit from base_model.
• [medium] [violation] providers/bothub/models/qwen3-coder-flash.toml:10 - Check: legacy_interleaved must not be hand-added, and never on non-reasoners. Why: AGENTS.md says sync writes legacy_interleaved; both non-reasoning files (qwen3-coder-flash, qwen3-coder-next inherit reasoning.supported=false) add [api.chat-completions] legacy_interleaved=true, publishing interleaved:true for models with no reasoning. Action: Remove the table/flag from non-reasoners; remove hand-added legacy_interleaved elsewhere unless preserving an existing synced value.
• [medium] [possible mistake] providers/bothub/models/grok-4.5.toml:16 - Check: Context-tier threshold and completeness must match authoritative pricing. Why: xAI natives use when={context={gte=200_000}} (providers/xai/models/grok-4.5.toml:46); Bothub grok-4.20/multi-agent/4.3/4.5/4.6/4.7 use gt=200000, shifting the 200k boundary; grok-4.3 drops native xhigh (none|low|medium|high vs native none|low|medium|high|xhigh), grok-4.5 drops xhigh (low|medium|high vs low|medium|high|xhigh). Action: Confirm Bothub bills at > vs >=200k and that xhigh is rejected on those IDs, or restore gte and missing xhigh.

[github-actions]

Action items

• [high] [violation] providers/bothub/models/glm-5.toml:12 - Check: Relay reasoning controls must copy the underlying model's lab/peer control set (AGENTS.md → Reasoning controls, audit skill Step 2). Why: effort = ["low","medium","high"] is invented for 13 models whose lab and same-surface peers expose no graded effort: glm-5/glm-5.1/glm-5-turbo (zhipuai first-party, OpenRouter and 302ai peers: binary toggle = true only), glm-5.2.toml:12 (zhipuai: toggle + effort = ["high","max"]; crusoe live-probed that low/medium/high are accepted but do not scale reasoning, so it encodes a toggle), kimi-k2.6.toml:13/kimi-k2.7-code.toml:13 (moonshotai first-party: toggle only / no caller control; OpenRouter and 302ai peers: toggle only), gemini-2.5-flash.toml:13/gemini-2.5-flash-lite.toml:13/gemini-2.5-pro.toml:12 (google first-party: toggle+budget / budget only; OpenRouter peer: toggle only), qwen3.7-flash.toml:12/qwen3.7-plus.toml:12/qwen3.7-max.toml:12 (alibaba first-party and 302ai peer: toggle + budget, no effort levels), and grok-4.20.toml:13 (xai grok-4.20-0309-reasoning first-party: no caller control at all). A bothub page listing alone doesn't meet the audit skill's evidence bar when it contradicts both lab and peers. Action: Replace these sets with the lab/peer shape (e.g. toggle = true with a leading top-of-file wire comment when reasoning_effort = "none" is the off switch and graded values don't scale), or keep graded effort only with host-doc/live evidence that the levels have distinct effects, noted in the header.
• [high] [possible mistake] providers/bothub/models/deepseek-v4-pro.toml:12 - Check: DeepSeek V4 effort levels must match the lab surface (reasoning_effort = low|high|max; minimal→low, medium/xhigh→high; GET /models reports [low, high, max], per the header of providers/deepseek/models/deepseek-v4-pro.toml). Why: deepseek-v4-pro.toml:12 and deepseek-v4-flash.toml:12 author effort = ["none","high","xhigh"] — xhigh is a DeepSeek alias of high, not a distinct level, and low is dropped. Per providers/deepseek, the undated deepseek-v4-pro serves V4-Pro-0813 and deepseek-v4-flash is served by V4.1-Flash-1017, i.e. the same underlying models as this PR's own deepseek-v4-pro-0813 (which the PR just corrected to none|low|high|max) and deepseek-v4.1-flash(-1017) (none|low|high|max), so the same host publishes contradictory sets for the same models. Action: Verify against bothub's live API and align the undated entries with the dated siblings (none|low|high|max), or document live evidence that bothub's xhigh behaves differently from high.
• [high] [possible mistake] providers/bothub/models/grok-4.20.toml:17 - Check: Provider [limit] overrides must be real host deltas, not restatements or values the underlying model cannot serve. Why: grok-4.20.toml and grok-4.20-multi-agent.toml override to context = 2_000_000 / output = 1_800_000, and grok-4.3.toml:18 sets output = 900_000 — 30–60× the documented 30,000-token output cap of these models (docs.x.ai, and the lab values this PR itself set in models/xai/grok-4.20-multi-agent.toml via PATCH 20). 1.8M output on a 2M-context model is not achievable; the bothub page pair was likely misread (the PR first created the lab at 2M/1.8M, then corrected it to 1M/30k, then re-published 2M/1.8M as host overrides — internally contradictory sourcing). Also grok-4.3.toml:17 restates context = 1_000_000, identical to models/xai/grok-4.3.toml (redundant override). Action: Live-verify an actual >30k-token completion on bothub (as done for muse-spark-1.3-contributor) or drop the output overrides and the redundant context line.
• [medium] [violation] providers/bothub/models/qwen3.8-flash.toml:12 - Check: Effort lists must be the model's actual levels, never a near-full enum dump. Why: qwen3.8-flash has ["low","medium","high"], but the Qwen3.8 generation's native (and 302ai peer) set is ["low","medium","xhigh"] — this PR's own qwen3.8-27b.toml:12 correctly uses ["none","low","medium","xhigh"]. qwen3.8-max-0902.toml:12 and qwen3.8-max-prime.toml:12 use ["minimal","low","medium","high","xhigh"] against the native ["low","medium","xhigh"], adding non-native minimal/high (full-enum-dump anti-pattern). grok-4.3.toml:13 drops the native xhigh that bothub's own grok-4.6/grok-4.7 files keep. Action: Copy the lab/peer sets for these four files, or document live evidence for each divergent level.
• [medium] [violation] providers/bothub/models/gpt-5.5.toml:4 - Check: AGENTS.md requires checking which tools.freeform formats each API accepts (host docs or a live probe with a restrictive grammar) when adding any OpenAI GPT-5+ model on any host, and noting the evidence in the header. Why: All 17 new GPT-5.5/5.6/6/6.1 files state "tools.freeform omitted: not live-probed on Bothub this pass" — the required check was explicitly not performed, while bothub only lists [api.chat-completions] and OpenAI documents custom tools on Chat Completions. Action: Probe bothub's chat-completions with a custom tool and a restrictive grammar (e.g. regex ZZ[0-9]{3}) whose output proves enforcement, set freeform for the formats that work, and record the evidence in the headers.
• [low] [violation] .pr-review/pull-request.json (PR body) - Check: AGENTS.md requires v1 data changes to be listed in the PR as clearly sourced fixes for maintainer approval. Why: The body's v1 price table lists the 2026-11-03 values, but the final files carry the 2026-11-12 refresh for every row (e.g. deepseek-v4-flash-0731 0.05/0.34 not 0.07/0.33; glm-5.3 1.2/3.77 not 1.17/3.69; gpt-5.6-luna 0.23/1.35 not 0.22/1.33; muse-spark 0.11/0.23 not 0.11/0.22), and the "Pricing source" line still says checked 2026-11-03 — a maintainer would approve numbers that don't ship. Action: Update the body's v1 table and source date to the final values.
• [low] [possible mistake] models/xai/grok-4.20-multi-agent.toml:16 - Check: New lab files must state facts true on every host. Why: The lab sets top_k = false, but every xAI peer (providers/xai/models/grok-4.20-multi-agent-0309.toml, grok-4.3–grok-4.7) and the grok-4.20-0309 lab files set top_k = true (xAI documents top_k as an API parameter) — the file matches the peer's limits but silently diverges on sampling. Action: Align top_k with the xAI entries or source the divergence in the header comment.

[github-actions]

Action items

• [high] [violation] providers/bothub/models/gemini-2.5-flash.toml:13 - Check: Relay reasoning controls must copy the underlying lab's option set, and levels beyond lab/peers require host docs or a live meaningful effect — never a bare HTTP 200 (AGENTS.md → Reasoning controls; .opencode/skills/audit-reasoning-options/SKILL.md Step 2 / Evidence bar). Why: These new files author effort = ["low","medium","high"] (or similar) for models whose first-party lab entries expose no graded effort at all: gemini-2.5-flash/gemini-2.5-flash-lite (google: toggle + budget), gemini-2.5-pro (google: budget only), glm-5/glm-5.1/glm-5-turbo (zhipuai: thinking.type toggle only), kimi-k2.6 (toggle only), kimi-k2.7-code (no caller control), qwen3.7-flash/qwen3.7-max/qwen3.7-plus (alibaba: enable_thinking + thinking_budget). glm-5.2 claims ["none","high","xhigh"] where the lab documents high|max (xhigh maps to max; max is dropped); deepseek-v4-flash/deepseek-v4-pro claim ["none","high","xhigh"] where the lab and AGENTS.md document DeepSeek V4 as low|high|max (xhigh maps to high); grok-4.20 claims graded effort where xAI's own grok-4.20-0309-reasoning entry has no caller control. The only recorded evidence is reasoning_effort=… → HTTP 200, often probed on a different model ("(family)"), and no peer relay (OpenRouter, W&B) authors L/M/H for these families — while this PR's own sibling files (deepseek-v4-*-0731/-0813, glm-5.3*, gpt-*, claude-*) correctly mirror their labs. Action: Copy each lab's option set (e.g. glm-5.2 → effort = ["high","max"]; undated deepseek v4 → ["none","low","high","max"] like the dated files in this PR; gemini-2.5 → toggle/budget shape or no controls; glm-5/5.1/5-turbo and kimi-k2.6 → toggle), or provide per-model evidence of meaningful effect on this host (e.g. reasoning-token counts that differ per level) plus the bothub page that documents the levels.
• [medium] [violation] providers/bothub/models/gpt-5.5.toml:9 - Check: tools.freeform on GPT-5+ models must come from the host's docs or a live probe with a restrictive grammar whose output proves the grammar is enforced, "not just accepted" (AGENTS.md → tool details). Why: All 17 new GPT files (gpt-5.5, gpt-5.5-pro, gpt-5.6-{luna,sol,terra} and -pro, gpt-6-{astra,luna,sol} and -pro, gpt-6.1-{sol,terra} and gpt-6.1-sol-pro) set tools = { supported = true, freeform = ["text","lark","regex"] } with the header "Live probe 2026-11-12: tools.freeform text|lark|regex → HTTP 200", which only proves the payloads were accepted; a gateway can accept custom-tool requests while ignoring the grammar, which is exactly the failure mode the rule guards against. Action: Re-probe bothub's /v1/chat/completions with a restrictive grammar per format (e.g. regex ZZ[0-9]{3}, a lark grammar) and record that the output conforms, or drop freeform from these files until that evidence exists.
• [medium] [possible mistake] providers/bothub/models/qwen3.8-flash.toml:12 - Check: Effort lists must match the lab set for the model (subset/superset deltas need host evidence). Why: qwen3.8-flash publishes ["low","medium","high"] where the alibaba lab entry documents low|medium|xhigh ("high" is only an accepted alias of xhigh); qwen3.8-max-0902 and qwen3.8-max-prime add minimal and high beyond the lab's low|medium|xhigh (the probe tested exactly those two non-native values, acceptance only); grok-4.3 drops xhigh from the xAI set and its recorded probe was run on grok-4.20 (a different model with a different set); grok-4.5 drops xhigh while grok-4.6/grok-4.7 in this same PR keep it. Action: Align these lists with the lab sets (low|medium|xhigh for the Qwen 3.8s, none|low|medium|high|xhigh for grok-4.3, low|medium|high|xhigh for grok-4.5) or verify each level per model on bothub with meaningful-effect evidence or a bothub page citation.
• [low] [violation] providers/bothub/models/gemini-4-carbon.toml:12 - Check: legacy_interleaved = true is a v1-only marker that AGENTS.md reserves for sync: "Sync writes and keeps it; don't add it by hand." Why: The PR hand-adds legacy_interleaved = true to every newly added reasoning model (~65 files, all new models except the two non-reasoning qwen3-coder-* files). These models have no v1 history to preserve; the 8 pre-existing bothub files carry the field from the v1→v2 migration, and hand-authoring it is what the rule prohibits. Action: Remove it from the newly added files, or get maintainer sign-off to keep v1's interleaved: true consistent across the provider's reasoning models.

[github-actions]

Action items

• [low] [possible mistake] providers/bothub/models/muse-spark-1.3-contributor.toml:26 - Check: New price vs the prior live billing verification. Why: The base file recorded a live usage.cost reading (2026-09-11, exactly $0.10/$0.20 per 1M); this PR overwrites it with page-scraped ECO prices ($0.11/$0.22) and deletes the verification note without reconciling the two sources. If the API still bills the old rate, the catalog now overstates the price. Action: Re-run a small chat completion and read usage.cost (as done on 2026-09-11) to confirm the new values, or note in the header why the page ECO rate supersedes the earlier live billing.
• [low] [possible mistake] providers/bothub/models/gpt-5.6-luna-pro.toml:14 - Check: The gpt-5.6-luna-pro pricing row. Why: Every other Pro SKU added in this PR carries a ~2.2–2.5× premium over its base SKU (gpt-5.6-sol 1.8/9.02 → sol-pro 4.51/22.54; gpt-6-luna 0.05/0.23 → 0.11/0.56; gpt-6.1-sol 0.9/4.51 → 2.25/11.27), but gpt-5.6-luna-pro (0.23/1.35) is identical to gpt-5.6-luna on both scrape dates — the row may have been duplicated from the base model. Action: Confirm on https://bothub.chat/models that the luna-pro row really equals the luna row, and note the absence of a pro premium in the header.
• [low] [possible mistake] providers/bothub/models/claude-opus-5.toml:8 - Check: Evidence for the none effort level on this host (also claude-opus-5-fast.toml:8, claude-opus-4.8-fast.toml:8, claude-sonnet-5.toml:8, claude-haiku-5.5.toml:8, gemini-3-flash-preview.toml:8, gemini-3.1-flash-lite.toml:8, gemini-3.1-flash-lite-preview.toml:8). Why: These effort lists include none, which the Anthropic and Google lab entries do not list (their sets start at low/minimal). The sets match established relays (Vercel/Poe for Claude, OpenRouter/Kilo/Prime for Gemini-3), but the PR body only claims "reasoning controls follow the lab set", and these files — unlike the deepseek/qwen/kimi/glm/grok files in this same PR — carry no Levels: https://bothub.chat/<id> citation or probe note for the host-specific none. Action: Add the bothub-page level citation or a live-probe note to these files (as done for the other families), or state the peer-copied basis in the PR body.
• [low] [possible mistake] models/xai/grok-4.20-multi-agent.toml:7 - Check: Lab description for grok-4.20-multi-agent. Why: xAI's official description ("Grok model for agentic tool use, reasoning, coding, and live assistance") is already carried verbatim in providers/xai/models/grok-4.20-multi-agent-0309.toml:20 and providers/poe/models/xai/grok-4.20-multi-agent.toml:7, and every other models/xai/grok-4.x lab file mirrors its xai counterpart's description; the new lab file invents a different paraphrase. Action: Use xAI's official description from docs.x.ai, as carried by the existing xai entry.