Add Arcee as an inference provider

[pranav4501]

Following the prices/README.md guidance, I’m opening an issue before adding a new provider.

Motivation

Arcee serves multiple models through its inference API with Arcee-specific pricing and context limits. Adding Arcee as a distinct provider will allow genai-prices to resolve and price requests made to:

https://api.arcee.ai/api/v1

Arcee exposes an OpenAI-compatible Chat Completions API and usage format.

Proposed scope

Add an arcee provider containing these exact model IDs:

• deepseek/deepseek-v4-flash-latest
• deepseek/deepseek-v4-pro
• deepseek/deepseek-v4-pro-0813
• moonshotai/kimi-k3
• thinkingmachines/inkling-small
• trinity-large-thinking
• zai-org/glm-5.2

The provider would include:

• input, cached-input-read, and output token prices
• provider-specific context windows
• exact model matching to avoid collisions with the models’ original providers
• Chat Completions usage extraction for prompt, cached-prompt, and completion tokens
• provider resolution through the Arcee API URL and provider name

Sources

Public pricing page:

https://docs.arcee.ai/get-started/pricing

Authenticated models endpoint:

https://api.arcee.ai/api/v1/models

The public pricing page currently lists six models. Pricing for deepseek/deepseek-v4-pro-0813, along with the context-window values, was verified through the authenticated /models endpoint.

Cache-write prices and maximum output-token limits would be omitted because Arcee does not currently publish them.

Implementation

I have an implementation prepared with generated Python and JavaScript data, provider-resolution tests, price-calculation tests, and usage-extraction tests:

https://github.com/pranav4501/genai-prices/tree/agent/add-arcee

Please let me know if there are any concerns with adding Arcee as a distinct provider or including the model whose pricing is currently available only through /models.

[pydanty]

pydanty runtime failure

Run df_run_54a18210e14849cf9d09 started from the GitHub webhook, but the Dogfooding runtime
failed before pydanty completed the workflow. The pydanty status comment above
(if any) is from a completed run and still stands.

Next step: maintainer attention needed; rerun after fixing the runtime failure.

Failure detail (tail)

Pydanty produced a triage bundle, but the Worker rejected or could not apply it: GitHub POST /repos/pydantic/genai-prices/issues/570/comments failed (403): {"message":"Resource not accessible by integration","documentation_url":"https://docs.github.com/rest/issues/comments#create-an-issue-comment","status":"403"}. No issue mutation or PR dispatch will be assumed complete without intervention.

[pydanty]

• Relationship resolved
• Existing-work coverage resolved
• Investigation inputs sufficient (not applicable)
• Supported production reachability established (not applicable)
• Contract and historical intent reconciled (not applicable)
• User/project relevance established (not applicable)
• Immediate relief assessed (not applicable)
• Long-term change is net-positive (not applicable)
• Change boundary and invasiveness understood (not applicable)
• Success criteria and integration-level regression tests defined (not applicable)

Decision: existing work

Legitimate, well-formed new-provider request: the reporter followed prices/README.md's issue-first gate, and it is distinct — the first Arcee provider ask anywhere in the repo. Its entire scope (arcee provider, 7 exact-match model ids, input/cache-read/output prices, context windows, usage extraction, URL/provider resolution) is already implemented by the reporter's own open PR #570, verified item-for-item against #570's head SHA. Existing work is complete, not partial: route to #570, open no competing PR. Residual work is process, not code — the maintainer decision to accept Arcee as a provider, re-verification of each number against Arcee's pricing page before merge, and optionally JS-side provider-resolution tests added to #570 at review time.

Why it matters: A second Arcee implementation PR would duplicate #570 and split review effort — coordination must happen on #570. The only real gate is the maintainer acceptance decision the README requires, and precedent (#549 Baseten, #557 QuickSilver, both unmerged with no maintainer response) shows contributor provider PRs wait by design. Arcee is a niche endpoint with no mis-pricing of existing data, so urgency is normal backlog (review #570 and decide), not a data-integrity risk; the one data caveat to check at review is the model priced solely via the authenticated /models endpoint.

Evidence: Issue #569 (2026-08-18, contributor pranav4501) is the README-required discussion opener for adding an arcee provider; no maintainer has engaged in 8+ days. Relationship: distinct — first and only Arcee provider ask in the repo; model-level overlaps (kimi-k3 via #467/#552, glm-5.2 via #409, DeepSeek V4 per #571, arcee-ai/trinity-large-thinking under openrouter) are the designed multi-provider data shape, not duplication. Existing work: PR #570 (same author, head agent/add-arcee, 1 commit, OPEN/unmerged, 0 comments/reviews) implements the full scope item-for-item: prices/providers/arcee.yml (+106) with all 7 exact-match model ids, api_pattern, provider_match, Chat-Completions usage extractor, prices_checked: 2026-08-17, and a price_comment flagging deepseek/deepseek-v4-pro-0813 as priced only from the authenticated /models endpoint; generated artifacts regenerated correctly (data.py, data.ts, v2 data.json/data_slim.json, usages.json, README provider list) with no schema.json change; Python tests present in test_price_calc.py (7-way price test, API-URL resolution, anti-spoof LookupError) plus test_extract_usage.py additions. Nothing on main covers any part of the scope; the shape follows maintainer precedent #562 (Modal). Not verified: the numbers against Arcee's own pricing page (arcee.ai links assessed untrusted; re-verification belongs to merge review), and #570 carries no dedicated JS-side provider-resolution tests.