Update GreenPT: cached-token rates, compression variants, kimi-k3

[Casburggraaf]

Refreshes the GreenPT catalog added in #3726. Four things changed on our side since that merged: prompt caching shipped, nine compression ids went GA, one model was withdrawn upstream, and a few list prices moved.

No entries are removed and nothing outside providers/greenpt/ is touched. Costs remain USD, converted from GreenPT's EUR list price at 1.14 USD/EUR; files whose price was re-verified for this update carry the later capture date in their top-of-file comment, untouched files keep the original one.

1. Cached prompt tokens (cost.cache_read)

GreenPT now bills prompt-cache hits at a reduced input rate. cost.cache_write is omitted rather than zeroed: cache writes are not charged at all.

entry	EUR cached	cache_read
glm-5.2 and all nine compression variants	0.275	0.3135
kimi-k3	0.825	0.9405
kimi-k2.6	0.22	0.2508
kimi-k2.7-code	0.165	0.1881
minimax-m2.5	0.055	0.0627

These are the only models that discount cache hits; every other GreenPT entry bills input at its normal rate and is left without the field.

2. Price corrections

Our upstream list prices were re-read from the gateway's own catalogue rather than its marketing page, which had been quoting higher figures:

entry	field	was	now	EUR
kimi-k2.6	input	0.828	0.7524	0.66
kimi-k2.6	output	4.389	4.275	3.75
kimi-k2.7-code	input	0.941	0.9006	0.79
minimax-m2.5	input	0.188	0.1938	0.17

glm-5.2 prices are unchanged; it only gains cache_read. Untouched entries were deliberately not re-rounded, to keep the diff limited to real changes.

3. New: kimi-k3

GA on GreenPT at 3.762 / 18.81, cached 0.9405. Declared through base_model = "moonshotai/kimi-k3", inheriting the 1,048,576 context. GreenPT serves it with text and image input, so the inherited video modality is overridden away, consistent with how kimi-k2.6 is already declared here.

4. New: nine compression variants of glm-5.2

glm-5.2-{caveman,honey,ponytail} each in -lite, unsuffixed and -ultra. They are the same upstream model carrying a built-in output-compression ruleset: caveman compresses prose, ponytail compresses generated code, honey compresses both, and the suffix selects intensity. Same price per token (including the cached rate) and the same context; the saving comes purely from generating fewer output tokens.

All nine are declared through base_model = "zhipuai/glm-5.2", so their cost and limits cannot drift from the base entry.

5. Deprecated: kimi-k2.6-fast

Withdrawn by the upstream provider; GreenPT no longer serves the id and requests for it fail. Marked status = "deprecated" rather than deleted, so existing configurations still resolve. Its prices are untouched.

Reasoning controls

Unchanged from #3726 and applied to the new entries too: the documented effort enum stays limited to the GreenPT-hosted models whose control is documented first-party (gemma4, green-r, green-r-raw). kimi-k3 and all nine variants are pass-through endpoints whose accepted values we have not verified, so they declare reasoning_options = [].

Verification

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. Diffing the resolved catalog against the published models.dev/api.json confirms exactly the 10 additions, the 4 repriced fields, the 15 cache_read values and the one status change, with nothing else moved.

Evidence

Claim	Source
Every cached rate, and which models discount cache hits	https://docs.greenpt.ai/prompt-caching
The nine compression ids, their families and intensities	https://docs.greenpt.ai/compression-models
Token prices, context windows, modalities	https://docs.greenpt.ai/model-cards
Full price table including the cached-input column	https://docs.greenpt.ai/pricing
reasoning_effort values on GreenPT-hosted models	https://docs.greenpt.ai/chat-completion

[github-actions]

No actionable findings.

[rekram1-node]

Are you sure reasoning efforts arent supported at all here? No toggles either? Seems inaccurate to me.