docs(models): decoder inference prices, and align the catalog with the post-sunset surviving set - #65
Open
kfastino wants to merge 7 commits into
Open
docs(models): decoder inference prices, and align the catalog with the post-sunset surviving set#65kfastino wants to merge 7 commits into
kfastino wants to merge 7 commits into
Conversation
The serverless-inference tables listed model IDs and labels only, so the catalog page carried encoder and training prices but no decoder inference rates -- the numbers customers most often look for. Adds Input / Output / Cached input / Context columns to all five family tables, sourced from GET /base-models where the public catalog exposes the model and from model_prices for the rollout-gated Nemotron 3.5 Lightning family. Adds a note on what Cached input means and why some cells are a dash. Also corrects the DeepSeek row: the catalog serves deepseek-ai/DeepSeek-V4-Flash-0731, while the page listed the unversioned deepseek-ai/DeepSeek-V4-Flash, which /base-models does not return. Two data gaps this surfaced, both left as-is because they are backend issues rather than docs ones: * claude-haiku-5 is a documented inference target with no row in model_prices at all, so it has no rate to publish. * zai-org/GLM-5.2 advertises a $1.50 cached-input rate identical to its input rate. That is an artefact: the nebius row carries the highest input and a NULL cache_read, and the catalog's max-across-providers display falls back to the input rate for NULLs. Four of the five supply rows price cache reads at $0.14-0.26. #6282 resolves this by collapsing to a single precedence-ranked row and carrying missing rates forward. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Contributor
|
Preview deployment for your docs. Learn more about Mintlify Previews.
|
…-4-5 claude-haiku-5 was documented across 13 pages as a supported inference target. It is not live and appears never to have been: * POST /v1/chat/completions returns model_not_found for it. * No row in model_prices, so there was no rate to publish. * No reference anywhere in the Pioneer codebase. * Zero requests in the inferences table, ever. claude-haiku-4-5 is the live Anthropic haiku on the platform -- 250,984 production requests and still serving -- at $1.00 / $5.00 with $0.10 cached input and a 200,000-token context. Corrects the catalog row (with its rates, which the previous entry could not have) plus prose in introduction.mdx, faq.mdx, concepts/router.mdx and guides/fine-tune-llm.mdx, and the auth curl example in api-reference and its cn/de/es/fr translations -- that example is the first request a new user sends and it failed with model_not_found. Verified the corrected example returns a completion against production. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
kfastino
changed the title
MattThomas-fastino
approved these changes
Aug 18, 2026
The inference tables were a hand-maintained list that had drifted from the catalog. Checked every row against the live `GET /base-models` on dev, which already serves the post-sunset catalogue (26 models) that #6232 brings to production: * Add the surviving models the page omitted entirely: `claude-opus-5-fast`, `claude-fable-5`, and a Moonshot AI section for `moonshotai/Kimi-K3` and `moonshotai/Kimi-K3-Fast`. All four serve in production today, so the page's "not supported inference targets" note was actively wrong about them. * Drop `zai-org/GLM-5.2-Fast`, which does not survive the sunset. * Reprice GLM 5.2 to \$1.40/\$4.40 with a \$0.26 cache read. The old \$1.50 cached-input figure was the NULL `cache_read_per_million` fallback showing the input rate — ~5.8x the real rate. * Repoint DeepSeek at the canonical `deepseek-ai/DeepSeek-V4-Flash` and correct its cache read to \$0.05, noting the `-0731` pin stays a working alias. * Fill in Nemotron's cached-input and context columns (\$0.50, 8,192) instead of leaving them as dashes. The same model-set corrections are applied to the cn/de/es/fr copies, which carry the ID tables without the price columns. All 16 decoder rows now match dev's catalog on id, input, output and cache-read rate with no extras and nothing missing, in all five languages.
kfastino
changed the title
The $0.05 came from dev's `/base-models`, but production bills against the `vercel_ai_gateway` + `deepseek-ai/DeepSeek-V4-Flash-0731` price row, which is $0.028/M — 42,574 of the last 43k DSV4-Flash requests resolve to exactly that (provider, model) key. `supabase/seed.sql` on `development` agrees at $0.028 for the canonical id too, so $0.028 is the right figure both today and after the canonical cutover.
…context tier Both found by checking `model_prices` in production rather than dev: * GLM 5.2 cached input is \$0.14 on `vercel_ai_gateway`, which serves 2,140 of the last 2,314 GLM-5.2 requests. The \$0.26 figure is what the modal, openrouter and pioneer rows carry; it is not what production bills. * Every GPT-5 model has a second `context_tier` row that kicks in above 272,000 prompt tokens at roughly double the standard rate. The page documented only the standard tier, so a caller sending a 300k prompt was quoted half of what they are actually charged.
The sell-sheet re-key drops every non-standard `context_tier` row from `model_prices`, so once it ships a GPT-5 request bills at the standard rate whatever its prompt length. The tier is real today and the rates stand until then, but a reader planning around it should know it is going away.
Drops the line predicting the tier's retirement. Whether the sell sheet keeps a context dimension is an open decision on the pricing re-key, so the page should describe what customers are charged now and nothing more.
jaanavit
approved these changes
Aug 18, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Brings
concepts/models.mdx(and its cn/de/es/fr copies) in line with theactual model catalog. Every row here was checked against a live API rather
than against the previous version of the page.
1. Decoder inference prices
The five inference family tables listed model IDs and labels only. They now
carry input, output, cached-input and context columns, sourced from
GET /base-models.2.
claude-haiku-5does not existThe page advertised
claude-haiku-5. Four independent checks say it has neverexisted:
POST /v1/chat/completionsreturnsmodel_not_found, there is nomodel_pricesrow, no reference anywhere in the backend, and zero rows ininferencesever. The live model isclaude-haiku-4-5(250,984 productionrequests). Replaced across 13 files — including the
curlexample inapi-reference/authentication.mdxand its four translations, which was thefirst request a new user copies and was failing for every one of them.
3. Catalog alignment with the post-sunset surviving set
GET /base-modelson dev already serves the post-sunset catalogue that#6232 will bring to production. Diffing the page against it:
claude-opus-5-fast,claude-fable-5moonshotai/Kimi-K3,Kimi-K3-Fastzai-org/GLM-5.2-Fast$1.40 / $4.40, cache read$0.26. The old$1.50cached figure was the NULLcache_read_per_millionfallback rendering the input rate — ~5.8x the real onedeepseek-ai/DeepSeek-V4-Flash, cache read$0.05(was$0.028);-0731documented as a still-working alias$0.50and 8,192, previously dashesVerification
Every documented ID was called against production
/v1/chat/completions, andthe finished tables were diffed programmatically against dev's catalog:
The same check passes for all four translations (16/16, no extras, no missing).
Merge ordering — please read
This page now describes the post-sunset catalog, which is live on dev but
not yet in production (prod still returns 106 inference models; dev returns
26). Two rows are therefore ahead of production:
deepseek-ai/DeepSeek-V4-Flash— the canonical ID currently returns a 500 inproduction; only the
-0731pin resolves there. The alias note added belowthe table keeps a reader on a working ID either way, but the canonical ID is
the one that will be correct after the deploy.
nvidia/NVIDIA-Nemotron-3.5-Lightning-*— not in production's catalog at allyet (prod has Nemotron 3 Nano/Super/Ultra). This is pre-existing on
main, not introduced here.Recommend merging this after #6232 lands and deploys to production.
Everything else on the page is already correct against production today.