[bayuh85]

Summary

This PR adds AtmoRouter (https://atmorouter.dev) as provider atmorouter. AtmoRouter is a multi-model relay. This entry models only its OpenAI-compatible Chat Completions API; no [api.messages] tables are added.

• Endpoint: https://atmorouter.dev/v1. Authentication: ATMOROUTER_API_KEY. npm: @ai-sdk/openai-compatible.
• The catalog covers 27 coding/agent chat models across the Claude, GPT, Gemini, GLM, Kimi, Qwen, DeepSeek, MiniMax and Grok families.
  • Each model references existing lab metadata through base_model (schema 2).
  • Each model file contains only real overrides plus the relay's reasoning surface.
• logo.svg is AtmoRouter's bolt mark: currentColor, square 0 0 32 32 viewBox, no fixed dimensions.

OpenCode consumes this registry, so no OpenCode source change is needed.

Sources and scope

Model IDs, context/output limits and input kinds come from AtmoRouter's public catalog at GET https://atmorouter.dev/v1/models, fetched on 2026-10-25. The authenticated list had 167 entries. Lab facts and release dates are inherited through base_model, not guessed.

• Pricing is omitted on purpose. This follows the precedent of relays whose prices are not fixed per model.
  • AtmoRouter bills each request at the rate of whichever offer tier serves it, and tier prices change continuously.
  • By default, a request can route to a tier priced above the published "From" rate unless the account sets a budget.
  • Putting the "From" price in cost would understate what users pay. The reason is recorded in provider.toml and in each model's header comment.
• Limits. A limit is overridden only where AtmoRouter publishes a lower value than the lab. When the lab's limit.input would then exceed the new context, base_model_omit = ["limit.input"] drops it.
  • GPT-5.6 Sol/Luna/Terra and GPT-6 Sol: context 272k.
  • GPT-6.1 Sol: context 500k.
  • Gemini 3.1 Pro: context 1M.
  • GLM-5.3 and GLM-5.3 Flash: output 48k.
  • Kimi K3 and K3.1 Pro: context 1M, output 32k.
  • MiniMax M3: context 1M, output 48k.
  • Where AtmoRouter lists a higher value than the lab, the lab value is kept, because it was not verified:
    • GPT-5.3 Codex: 1M listed vs the lab's 400k.
    • Sonnet 4.6: 128k output listed vs 64k.
    • DeepSeek V4.1 Flash and V4 Pro: 393,216 output listed vs 384,000.
• Modalities are the lab's input MIME types filtered to the kinds AtmoRouter lists for each model. AtmoRouter lists no PDF/file kind, so PDF is dropped where the lab has it. Output is text.
• Reasoning. AtmoRouter is a same-surface relay that takes top-level reasoning_effort. Each model copies the reasoning controls of its lab entry (effort levels, toggle, budget). A toggle or budget is kept only when a live call showed AtmoRouter forwards that field (see "Reasoning controls" below); no universal ladder is invented.
  • kimi-k2.7-code is supported = true with no effort list, like its lab entry.
  • minimax-m3 copies the lab's toggle. Through AtmoRouter, thinking is on when thinking is omitted and off with thinking.type = disabled, so the default is recorded as on.
  • The lowest and highest effort of every list was sent live and accepted (see below).
  • claude-sonnet-5 leaves out the lab's toggle because thinking.type = disabled still returned reasoning through AtmoRouter.
  • system_updates, effort_updates and tools.updates are left unset, because AtmoRouter does not document them.
• Live testing. Every model was live-tested before inclusion. moon/kimi-k3 returned HTTP 502 in testing and was left out in favour of cb/kimi-k3, which passed.

Docs: AtmoRouter, docs, models.

Validation

On Linux with Bun 1.4.2, against dev at 7cd0423:

• bun install --frozen-lockfile: passed.
• bun validate: passed.
• cd packages/sdk && bun run test: passed, 25 tests, 0 failures.
• git diff --cached --check: passed.
• bun run v1:check: exits 1, and every reported change is an addition.
  • It reports the 27 new atmorouter/* models as "Added to v1" and the new provider's fields (id, env, npm, api, name, doc) as added. It reports "v1 output changed in 3 of 22 views".
  • No existing provider or model changes.
  • Prime Inference (feat: add Prime Inference provider #9497) was merged the same way, with its models in v1. If you would rather mark these v2_only, I'm happy to change that.

Live test

Run on 25 Oct 2026, about 16:50–17:10 WIB (UTC+7), against https://atmorouter.dev/v1 with a real AtmoRouter key. The key is not in this PR or its logs.

• Model list. Authenticated GET /v1/models returned HTTP 200 in 1.14s with 167 models. All 27 IDs below are listed.
• Chat. Each model got Reply with exactly: pong at temperature 0.2 and max_tokens 256, once non-streaming and once streaming.
  • Every call returned HTTP 200 with pong, and every stream ended with [DONE].
  • The cb/gemini-3.1-pro non-streaming call was slow in this single sample; its stream finished in about 5s.
• Effort probe. The same prompt was sent with reasoning_effort set to the lowest and the highest value in each model's effort list. Every call returned HTTP 200 (the lists were revised after review; the changed boundaries were re-sent at about 17:25 WIB).
  • This shows the relay accepts every listed boundary value.
  • It does not show that each level changes the upstream behaviour.

Model	base_model	Non-stream	Stream total / first token	Effort probe (lowest / highest)
cc/claude-sonnet-4-6	anthropic/claude-sonnet-4-6	2.15s	2.39s / 2.38s	low 200, max 200
cc/claude-opus-4-7	anthropic/claude-opus-4-7	1.64s	1.71s / 1.7s	low 200, max 200
cc/claude-opus-4-6	anthropic/claude-opus-4-6	2.76s	2.91s / 2.89s	low 200, max 200
cc/claude-opus-4-8	anthropic/claude-opus-4-8	1.83s	1.62s / 1.55s	low 200, max 200
cc/claude-haiku-4-5	anthropic/claude-haiku-4-5	2.46s	1.85s / 1.83s	n/a (budget only, like the lab)
cc/claude-sonnet-5	anthropic/claude-sonnet-5	1.54s	1.69s / 1.69s	low 200, max 200
cc/claude-opus-5	anthropic/claude-opus-5	2.10s	2.09s / 1.67s	low 200, max 200
cc/claude-sonnet-5-5	anthropic/claude-sonnet-5-5	2.21s	2.98s / 2.97s	low 200, max 200
cc/claude-opus-5-5	anthropic/claude-opus-5-5	1.77s	2.17s / 2.16s	low 200, max 200
cx/gpt-5.6-sol	openai/gpt-5.6-sol	4.98s	4.01s / 3.9s	none 200, max 200
cx/gpt-5.6-luna	openai/gpt-5.6-luna	7.03s	6.68s / 6.56s	none 200, max 200
cx/gpt-5.6-terra	openai/gpt-5.6-terra	7.78s	6.98s / 6.89s	none 200, max 200
cx/gpt-6-sol	openai/gpt-6-sol	13.44s	7.66s / 7.56s	none 200, max 200
cx/gpt-6.1-sol	openai/gpt-6.1-sol	7.18s	6.98s / 6.88s	low 200, max 200
cb/gpt-5.3-codex	openai/gpt-5.3-codex	3.16s	4.65s / 4.49s	none 200, xhigh 200
cb/gemini-3.1-pro	google/gemini-3.1-pro-preview	24.85s	5.11s / 4.78s	low 200, high 200
cbcn/glm-5.3	zhipuai/glm-5.3	3.72s	2.77s / 2.65s	low 200, max 200
cbcn/glm-5.3-flash	zhipuai/glm-5.3-flash	2.51s	2.45s / 2.39s	low 200, max 200
ali/kimi-k2.7-code	moonshotai/kimi-k2.7-code	2.07s	1.62s / 1.59s	n/a (no effort list)
cb/kimi-k3	moonshotai/kimi-k3	4.09s	4.13s / 4.12s	low 200, max 200
cb/kimi-k3.1-pro	moonshotai/kimi-k3.1-pro	5.04s	6.78s / 6.78s	low 200, max 200
ali/qwen3.8-max	alibaba/qwen3.8-max	2.20s	1.90s / 1.87s	low 200, xhigh 200
ali/qwen3.8-flash	alibaba/qwen3.8-flash	1.34s	2.31s / 2.27s	low 200, xhigh 200
ali/deepseek-v4.1-flash	deepseek/deepseek-v4.1-flash	1.92s	3.85s / 3.85s	low 200, max 200
cbcn/deepseek-v4-pro	deepseek/deepseek-v4-pro	2.77s	2.09s / 2.07s	low 200, max 200
cb/minimax-m3	minimax/MiniMax-M3	1.88s	2.48s / 2.46s	n/a (toggle, no effort list)
grok/grok-4.7	xai/grok-4.7	2.01s	2.27s / 2.23s	low 200, xhigh 200

Reasoning controls (live, about 17:25–17:30 WIB)

A multi-step arithmetic prompt was sent to check which lab controls AtmoRouter actually forwards. reasoning_tokens comes from usage.completion_tokens_details.

Model	Control sent	Result
cbcn/deepseek-v4-pro	thinking.type = disabled	0 reasoning tokens (default: 1087)
ali/deepseek-v4.1-flash	thinking.type = disabled	no reasoning (default: 3266)
ali/qwen3.8-max	enable_thinking = false / thinking_budget = 200	no reasoning / exactly 200 reasoning tokens (default: 1728)
ali/qwen3.8-flash	enable_thinking = false / thinking_budget = 200	no reasoning / exactly 200 reasoning tokens (default: 4625)
cb/gpt-5.3-codex	reasoning_effort = none	200, 129 reasoning tokens (low: 238, xhigh: 413)
cb/kimi-k3, cb/kimi-k3.1-pro	reasoning_effort = none	still reasoned (316 / 211 tokens), so none was dropped
cb/minimax-m3	thinking.type = disabled / adaptive / enabled	0 reasoning tokens for all three (omitted: ~860, twice), so toggle with default on
cc/claude-sonnet-5	thinking.type = disabled	still reasoned (112 tokens), so no toggle
cc/claude-haiku-4-5, cc/claude-sonnet-4-6	thinking.budget_tokens = 1024 / 12000	200 both; Anthropic treats the budget as a soft cap, so this shows only that the field is accepted

[github-actions]

Action items

• [high] [violation] providers/atmorouter/models/cc/claude-haiku-4-5.toml:13 - Check: Baseline effort = lab + same-surface peer set, not fixed low/medium/high. Why: Native providers/anthropic/models/claude-haiku-4-5.toml is budget-only with no effort, relay peers are toggle+budget (CrossModel/Merge), budget (Opencode) or reasoning=false (Requesty); the only effort peer (LLMGateway) is 5-level low/medium/high/xhigh/max. No baseline has 3-level low/medium/high. Action: Replace with the host-verified surface (budget/toggle if Messages is forwarded, or peer-supported effort with docs/live proof each level has distinct effect) and cite it.
• [high] [violation] providers/atmorouter/models/cb/gpt-5.3-codex.toml:14 - Check: Relay must copy underlying controls from lab + same-surface peers. Why: Both providers/openai/models/gpt-5.3-codex.toml and providers/openrouter/models/openai/gpt-5.3-codex.toml are none/low/medium/high/xhigh; PR drops none to 4 levels with only “lowest/highest accepted” and no proof the host rejects none. Action: Restore none or provide host docs/live proof none is rejected/ignored on this host.
• [high] [violation] providers/atmorouter/models/cbcn/deepseek-v4-pro.toml:9 - Check: DeepSeek V4 baseline is toggle + low/high/max per AGENTS.md. Why: First-party providers/deepseek/models/deepseek-v4-pro.toml is toggle=true, effort=[low,high,max] with xhigh→high mapping; PR uses effort=[none,high,xhigh] with no toggle — invents distinct xhigh, drops low/max, collapses off to none without proof none disables thinking. Same pattern in ali/deepseek-v4.1-flash.toml. Action: Change to toggle=true + effort=[low,high,max] with leading Toggle/Effort wire comments, or provide host docs/live proof each of none/high/xhigh has distinct meaningful effect and no toggle exists.
• [medium] [possible mistake] providers/atmorouter/models/cb/kimi-k3.toml:9 - Check: Do not invent off where lab says always-on. Why: providers/moonshotai/models/kimi-k3.toml and kimi-k3.1-pro.toml are effort=[low,high,max] with “no off toggle”; PR adds none (none/low/high/max, same in cb/kimi-k3.1-pro.toml) with only “accepted live” — acceptance does not prove none disables reasoning. Action: Remove none or prove live that none returns no reasoning content vs ignored.
• [medium] [possible mistake] providers/atmorouter/models/ali/qwen3.8-flash.toml:14 - Check: Copy toggle/budget when lab/peers expose them unless host lacks the field. Why: providers/alibaba/models/qwen3.8-flash.toml is toggle+effort=[low,medium,xhigh]+budget and CrossModel peer is toggle+effort=[low,medium,xhigh]; PR drops to effort-only with no proof AtmoRouter lacks enable_thinking/thinking_budget (same in ali/qwen3.8-max.toml). Action: Add toggle (+budget if forwarded) with wire comments, or cite host docs/live showing those fields absent/rejected.
• [medium] [possible mistake] providers/atmorouter/models/cc/claude-opus-4-7.toml:14 - Check: Extra levels beyond lab/peers need host docs or live meaningful effect. Why: Native is low/medium/high/xhigh/max (no none); Requesty peer is none/low/medium/high/max (no xhigh); PR unions to 6-level none/low/medium/high/xhigh/max with only “lowest/highest accepted” — no proof none disables or xhigh differs from high. Same union in cc/claude-opus-4-8.toml, cc/claude-opus-5.toml, cc/claude-sonnet-5.toml. Action: Align to one baseline or provide per-level docs/live proof of distinct effect.
• [low] [possible mistake] providers/atmorouter/provider.toml:1 - Check: Model [api.*] tables are the complete API list and replace the provider default. Why: Header/body advertise OpenAI Chat Completions plus Anthropic Messages-compatible API, but all 27 models list no [api.*] and therefore expose only the default https://atmorouter.dev/v1 chat-completions route. Action: Add [api.messages] tables where supported (with base_url/sdk) or correct the comment/body to chat-completions only.

[github-actions]

Action items

• [high] [violation] providers/atmorouter/models/cb/minimax-m3.toml:9 - Check: Relay reasoning controls must copy the lab + same-surface peers; reasoning = { supported = true } alone means affirmatively always-on, not uncertainty. Why: First-party providers/minimax/models/MiniMax-M3.toml exposes toggle = true (thinking.type = adaptive|disabled) and peers expose controls (providers/opencode/models/minimax-m3.toml has toggle, providers/requesty/models/minimax-m3.toml has effort), but this file claims always-on with only "documents no reasoning control" and no live disabled-returns-no-reasoning test like the DeepSeek/Qwen/Claude files. Action: Live-test thinking.type = disabled|adaptive on AtmoRouter Chat Completions; if it disables thinking add toggle = true with a leading # Toggle: wire comment, otherwise replace the docs-only note with the dated live result.

[github-actions]

No actionable findings.