Add GreenPT provider

[Casburggraaf]

Adds GreenPT (https://greenpt.ai) as a provider: 24 chat models plus 2 speech-to-text models, served through an OpenAI-compatible endpoint at https://api.greenpt.ai/v1.

This supersedes #2963, which a stale bot closed while I was preparing an update. GitHub would not let me reopen it, so this is the same branch with all review feedback applied. Every model that reuses existing metadata goes through base_model and carries only GreenPT-specific overrides.

Pricing

Costs are USD per million tokens, converted from GreenPT's EUR list price at 1.14 USD/EUR (rate captured 2026-07-24). Every EUR-sourced file carries that rate and its sources in a top-of-file comment.

Speech-to-text is billed per hour and published here as the standard EUR 0.23/hour pre-recorded rate (USD 0.00437/minute). A 50% promotion runs through 31 August 2026, but the durable standard rate is published so the catalog stays correct after it ends. Live and multilingual tiers are priced higher and cannot be represented in the single-rate field.

Context limits

Provider-specific limit.context overrides where GreenPT serves a smaller window than the base metadata:

model	base	GreenPT
gemma-3-27b-it	128k	40k
devstral-2-123b-instruct-2512	262,144	200k
llama-3.3-70b-instruct	128k	100k
qwen3-coder-30b-a3b-instruct	262,144	128k

Modalities

Each entry declares exactly what the GreenPT endpoint serves, per the capability tags on the model cards. Both directions occur:

model	override	model card
gpt-oss-120b	image added	"Large open model with vision, function calling..."
green-r, green-r-raw	image added	Text, Images, Documents, Multilingual
green-l, green-l-raw	image added	Text, Images, Documents, Multilingual
mistral-small-3.2-24b-instruct-2506	image added	Chat, Vision, Function calling
kimi-k2.6, kimi-k2.6-fast	video dropped, image kept	Chat, Reasoning, Agentic coding, Vision
kimi-k2.7-code	image + video dropped, text-only	Chat, Reasoning, Coding, Agentic tasks (no vision)
qwen3.6-35b-a3b	video + audio dropped, image kept	Chat, Code, Vision
qwen3.5-397b-a17b	image + video + audio dropped, text-only	Chat, Code generation, Agentic tasks, Logical reasoning (no vision)
llama-3.3-70b-instruct	attachment = false	Chat, Instruction following, Multilingual (text-only)

kimi-k2.7-code and qwen3.5-397b-a17b are deliberately text-only: GreenPT serves them as text endpoints and their model cards carry no vision capability, unlike their kimi-k2.6 and qwen3.6 siblings.

Reasoning controls

The documented effort enum is declared only on the GreenPT-hosted models whose control is documented first-party and implemented in the proxy itself:

# gemma4, green-r, green-r-raw
reasoning_options = [{ type = "effort", values = ["none", "minimal", "low", "medium", "high"] }]

Thinking is enabled by default, "none" disables it, and the OpenAI-spec levels are accepted for compatibility.

The other eleven reasoning entries are third-party pass-through endpoints: GreenPT forwards reasoning_effort upstream unchanged and does not verify accepted values per model, so they declare reasoning_options = [] rather than an assumed enum.

Also

Renamed gemma-4-26b-a4b-it to gemma4 to match the id the API actually serves. The upstream weights are still referenced via base_model.

Rebased onto current dev. bun validate is green. bun test shows no failures beyond the 3 pre-existing packages/sdk/src/snapshot.js module-resolution errors that also fail on a clean dev checkout.

Evidence

Claim	Source
Token prices (EUR list, converted at 1.14)	https://docs.greenpt.ai/pricing and https://docs.greenpt.ai/model-cards
Context limits and max output	https://docs.greenpt.ai/model-cards, per-model cards
Modalities / vision per endpoint	https://docs.greenpt.ai/model-cards, per-model capability tags (table above)
reasoning_effort values on GreenPT-hosted models	https://docs.greenpt.ai/chat-completion (request body and Notes)
Speech-to-text hourly rates and promotion window	https://docs.greenpt.ai/pricing, speech-to-text section
Served model ids (including gemma4)	https://docs.greenpt.ai/models

[github-actions]

Action items

• [high] [possible mistake] providers/greenpt/models/gemma4.toml:3 - Check: reasoning_options must reflect GreenPT’s verified request surface per model (audit-reasoning-options: do not copy a full effort enum; require explicit evidence for none and minimal; OpenAI-compatible acceptance alone is weak). Why: The same ["none", "minimal", "low", "medium", "high"] set is applied to every reasoning entry (including holo2-30b-a3b, mistral-medium-3.5-128b, gpt-oss-120b / green-r*, Kimi, GLM, Qwen, MiniMax). Other providers document model-specific subsets for the same weights (e.g. gpt-oss often excludes none). Overstated controls mislead clients. Action: Cite GreenPT docs or API tests for the exact request field and each value per model (or family). Drop unverified values; use reasoning_options = [] when no control is verified. Optionally note the request syntax next to the option (e.g. # API: {"reasoning_effort": "<value>"}).
• [medium] [possible mistake] providers/greenpt/models/green-s.toml:15 - Check: Catalog cost should be durable provider pricing, not a short-lived promotion. Why: green-s / green-s-pro publish promotional EUR 0.12/hour (input = 0.00228) through 31 Aug 2026; standard EUR 0.23/hour (~$0.00437/min) is only in comments. After the promo, the API will understate STT cost. Action: Set input to the standard USD/minute rate; keep the promo as a top-of-file note if desired.
• [medium] [violation] providers/greenpt/models/green-s.toml:9 - Check: Non-USD source prices must be converted to USD with rate and date in a leading top-of-file comment (AGENTS.md Cost Schema). Why: STT comments document EUR→USD at 1.14 with no rate date, and sit under [cost] (not above the first key). Chat costs look EUR-scaled (e.g. 0.228, 0.57) but have no conversion notes. Action: Add a top-of-file comment on every EUR-sourced file with rate, date, and what it applies to; move STT notes there.
• [low] [possible mistake] .pr-review/pull-request.json:4 - Check: Data PRs should cite first-party sources mapped to claims (pricing, context overrides, modalities, reasoning). Why: The body describes production-catalog reconciliation but does not clearly map URLs to each material claim (token prices, 40k/100k/128k/200k contexts, vision/STT modalities, effort enum). Action: Add an Evidence section with direct GreenPT pricing/docs/API links and state what each supports.

[Casburggraaf]

Thanks, this was a useful pass. All four items are addressed in c004f2c.

1. [high] reasoning_options overstated. Fair challenge, and you were right that OpenAI-compatible acceptance is not evidence. I have scoped the effort enum to the three GreenPT-hosted models whose control is documented first-party, where the proxy itself implements the contract (gemma4, green-r, green-r-raw), and annotated them with the source:

# reasoning_effort: thinking is enabled by default, "none" disables it, and
# minimal/low/medium/high are accepted. See https://docs.greenpt.ai/chat-completion

The other eleven entries are third-party pass-through endpoints: GreenPT forwards reasoning_effort upstream unchanged and does not verify the accepted values per model. Rather than assume a set, those now declare reasoning_options = [], including gpt-oss-120b, holo2-30b-a3b, mistral-medium-3.5-128b and the GLM / Kimi / Qwen / MiniMax entries. I would rather under-claim here than publish a control we have not confirmed. If we verify per-model behaviour later I will submit them separately with the test output.

2. [medium] Promotional speech-to-text price. Agreed, and changed. cost.input is now the standard EUR 0.23/hour, USD 0.00437/minute. The promotion is recorded in the header instead, so the catalog stays correct after 31 August 2026.

3. [medium] Missing conversion comment. Fixed on every EUR-sourced file, above the first key as the AGENTS.md cost schema requires:

# Cost: converted from GreenPT's EUR list price at 1.14 USD/EUR (rate captured 2026-07-24).
# Sources: https://docs.greenpt.ai/pricing and https://docs.greenpt.ai/model-cards

4. [low] Evidence. Added an Evidence section to the PR description mapping each class of claim to its source.

bun validate is green.

[github-actions]

Action items

• [medium] [possible mistake] providers/greenpt/models/kimi-k2.7-code.toml:4 - Check: Provider modality overrides must match what GreenPT actually serves. Why: This entry forces attachment = false and modalities.input = ["text"], while kimi-k2.6 / kimi-k2.6-fast keep image input and the PR body says all three Kimi entries are vision-only (video dropped). Base metadata is multimodal. Action: Verify against GreenPT model cards/API; if vision is served, set attachment = true and input = ["text", "image"]; if text-only is correct, keep the TOML and correct the PR description.
• [medium] [possible mistake] providers/greenpt/models/qwen3.5-397b-a17b.toml:4 - Check: Multimodal base models should not silently lose image input without evidence. Why: This strips to text-only (attachment = false, input = ["text"]), but sibling qwen3.6-35b-a3b keeps image and the PR body only claims dropping inherited video/audio for both Qwen entries. Base alibaba/qwen3.5-397b-a17b is multimodal. Action: Confirm GreenPT’s served modalities; restore image (+ attachment = true) if supported, or document that this deployment is text-only.
• [medium] [possible mistake] providers/greenpt/models/gpt-oss-120b.toml:4 - Check: Image modality claims for gpt-oss-derived IDs. Why: gpt-oss-120b, green-r, and green-r-raw set attachment = true and modalities.input = ["text", "image"], but models/openai/gpt-oss-120b is text-only. That is a large capability jump for the same base weights. Action: Verify GreenPT accepts image parts for these exact model IDs; if not, drop the image overrides and leave text-only inheritance.
• [low] [possible mistake] .pr-review/pull-request.json - Check: PR body should match the final catalog data and map citations to claims. Why: The body still describes promotional STT pricing as published and full reasoning_options on every reasoning model, but the final TOMLs publish the standard STT rate and scope effort controls to gemma4 / green-r / green-r-raw only. Action: Update the PR body to the final prices/options and note which first-party pages support modalities, limits, and reasoning controls.

[Casburggraaf]

Thanks, the fourth point was a real miss on my side and the first three were worth asking about. I have rewritten the PR description to match the final data exactly, including a per-model modality table with the supporting model-card capability tags.

On the three modality questions, the TOMLs are correct and the old description was the inaccurate part, so I kept the data and fixed the text:

kimi-k2.7-code text-only. Correct as published. Its model card lists Chat, Reasoning, Coding, Agentic tasks with no vision capability, unlike kimi-k2.6 / kimi-k2.6-fast, which do list Vision and therefore keep image input. The old description wrongly lumped all three together as "vision only"; that sentence is gone.

qwen3.5-397b-a17b text-only. Also correct. Its card lists Chat, Code generation, Agentic tasks and Logical reasoning with no vision capability, whereas qwen3.6-35b-a3b explicitly lists Vision and keeps image. The old description only mentioned dropping video/audio and did not state that image was dropped for this one; now spelled out.

Image on the gpt-oss-derived ids. GreenPT serves these with image input even though the base metadata is text-only. The gpt-oss-120b model card reads "Large open model with vision, function calling, and long-context reasoning support", and the green-r / green-r-raw cards list Text, Images, Documents, Multilingual. That is the deployment difference the override is meant to express.

I also went the other way where our deployment is narrower than the base: llama-3.3-70b-instruct now sets attachment = false, since its card is text-only while the base metadata sets attachment = true.

Sources for each claim are in the Evidence table at the bottom of the description. bun validate is green.