Every LLM's price, context window and capabilities behind one REST API — and a cost estimator that shows its work.
Picking a model means checking four things at once: what it costs, how much context it takes, what it can do, and whether it is about to be deprecated. That lives in a 1.7 MB JSON file you have to parse yourself. This service reads it once at boot and serves 3,214 models from 127 providers out of memory, with no database — filtered, sorted, compared — and prices a single call down to the exact rate key it used.
curl https://api-llm-specs.axium-lab.com/v1/models/claude-sonnet-4-5
Ask for a model and you get everything the catalog knows about it — every rate, including the ones that only apply past a threshold or to a cache write, plus the context window and the capability flags:
{ // 29 attributes in total, trimmed here
"deprecation_date": "2026-09-29",
"cache_creation_input_token_cost": 0.00000375,
"cache_creation_input_token_cost_above_1hr": 0.000006,
"cache_read_input_token_cost": 3e-7,
"input_cost_per_token": 0.000003,
"input_cost_per_token_above_200k_tokens": 0.000006,
"output_cost_per_token_above_200k_tokens": 0.0000225,
"provider": "anthropic",
"max_input_tokens": 200000,
"max_output_tokens": 64000,
"mode": "chat",
"output_cost_per_token": 0.000015,
"supports_prompt_caching": true,
"supports_vision": true,
"prompt_cache_min_tokens": 1024,
"id": "claude-sonnet-4-5"
}
Or ask a question of the whole catalog — a provider, a mode, a price ceiling, a context window — and project just the fields you care about:
curl 'https://api-llm-specs.axium-lab.com/v1/models?provider=anthropic&mode=chat&sort=input_cost_per_token:desc&fields=id,input_cost_per_token,output_cost_per_token,max_input_tokens&limit=3'
{
"total": 26,
"limit": 3,
"offset": 0,
"data": [
{ "id": "claude-3-opus-20240229", "input_cost_per_token": 0.000015, "output_cost_per_token": 0.000075, "max_input_tokens": 200000 },
{ "id": "claude-4-opus-20250514", "input_cost_per_token": 0.000015, "output_cost_per_token": 0.000075, "max_input_tokens": 200000 },
{ "id": "claude-opus-4-1", "input_cost_per_token": 0.000015, "output_cost_per_token": 0.000075, "max_input_tokens": 200000 }
]
}
total
is the size of the filtered set before the page is cut, and 126 more providers' worth of models are one
provider=
away. Every filter, sort and projection →
Rates are one thing; knowing which of those four Anthropic cache keys applies to your request is another.
That is what POST /v1/estimate
is for — and it answers with a receipt, not a number. Every line carries
rate_key
, the literal key it was billed with:
input.text 190000 x 0.000006 = 1.14 [input_cost_per_token_above_200k_tokens]
output.text 4000 x 0.0000225 = 0.09 [output_cost_per_token_above_200k_tokens]
cache_read.text 20000 x 6e-7 = 0.012 [cache_read_input_token_cost_above_200k_tokens]
cache_write.1h 8000 x 0.000012 = 0.096 [cache_creation_input_token_cost_above_1hr_above_200k_tokens]
------
1.338 USD
That prompt crossed the 200k threshold, so the whole request was repriced — output included — and the 1 hour
cache TTL composed with it into that quadruple key. Both are decisions the estimator makes explicitly and
reports back in resolution
⚠️ Hosted instance — not live yet.https://api-llm-specs.axium-lab.com
is the address the free instance will answer on; it is not deployed at the time of writing. Until then, run it locally or deploy your own — it is onedocker run
away, and the dataset ships inside the image.
📖 Full API reference at llm-specs.axium-lab.com — every endpoint, parameter, response shape and error, with the numbers taken from real responses.
Auditable, not magic. Every cost line names the dataset key it used (rate_key
) and the quantity it multiplied. A total you cannot check against the source is a total you cannot trust.Decimal arithmetic, never The catalog ships floating point noise already serialized (number
.1.2999000000000001e-07
) and rates as small as1.3e-10
. Everything accumulates inDecimal
.A missing rate is never billed as 0. 258 models legitimately declareoutput_cost_per_token: 0
, so "free" and "unknown" have to stay distinguishable: unknown usage goes tounpriced[]
and never inflates or deflates the total silently.It boots without network. The dataset is versioned in the repo and baked into the image. At boot the instance revalidates it upstream withIf-None-Match
; on a304
it transfers0 bytes, and if upstream is down it serves the local copy and says so in/health
.Errors you can branch on. Every failure is RFC 9457application/problem+json
with a stabletype
slug —model-not-found
,ambiguous-model
,model-not-priced
,limits-exceeded
,invalid-query
.
Requires Bun.
git clone https://github.com/axium-lab/llm-specs-api.git
cd llm-specs-api
bun install
bun start # http://localhost:8080
Or straight from the published image — nothing to build, and the dataset is already inside:
docker run --rm -p 8080:8080 ghcr.io/axium-lab/llm-specs-api:latest
To build it yourself instead:
docker build -t llm-specs-api .
docker run --rm -p 8080:8080 llm-specs-api
Then ask it something:
curl 'localhost:8080/v1/models?mode=chat&min_input_tokens=1000000&sort=input_cost_per_token:asc&fields=id,provider,input_cost_per_token&limit=5'
curl 'localhost:8080/v1/models/bedrock/us.anthropic.claude-3-5-haiku-20241022-v1:0'
| Method | Path | Description |
|---|---|---|
GET |
||
/health |
startup_error
.GET
/v1/models
GET
/v1/models/*
/
, :
and *
.GET
/v1/models/by-id?id=
GET
/v1/compare?ids=a,b,c
GET
/v1/providers
GET
/v1/modes
GET
/v1/attributes
GET
/v1/meta
POST
/v1/estimate
Filters for /v1/models
: provider
, mode
, q
, any of the 37 supports_*
keys, min_input_tokens
,
max_input_cost
, sort=field:asc|desc
, fields
, limit
, offset
Everything is optional — the service runs out of the box.
| Environment variable | Default | Description |
|---|---|---|
PORT |
||
8080 |
||
| Injected by Cloud Run. | ||
DATASET_PATH |
||
data/model_prices_and_context_window.json |
||
| The source of truth. The sidecar path is derived from it. | ||
UPSTREAM_URL |
||
LiteLLM's litellm_internal_staging branch |
||
| Revalidation target. See | ||
FETCH_TIMEOUT_MS
30000
DEFAULT_LIMIT
50
/v1/models
.MAX_LIMIT
500
limit
; a larger value is clamped, not rejected.Positive integers only — a malformed value fails the boot instead of being silently ignored.
Every release ships a docker-compose.yml
with the image pinned to that exact version:
curl -LO https://github.com/axium-lab/llm-specs-api/releases/latest/download/docker-compose.yml
docker compose up -d
curl localhost:8080/health
Images are published for linux/amd64
and linux/arm64
at ghcr.io/axium-lab/llm-specs-api, tagged
X.Y.Z
, X.Y
and latest
. A pre-release is never tagged latest
.The compose file declares a named volume over /app/data
, and it matters: the dataset travels inside the image and the service rewrites it when upstream has something newer, so without the volume every restart throws that update away and starts again from the copy baked in at build time. It has to be a named volume — Docker seeds one from the image on first use, whereas an empty bind mount would hide the dataset and leave the service with nothing to serve.
gcloud run deploy llm-specs-api --source . --region europe-west1 \
--min-instances 1 --cpu-boost --allow-unauthenticated
--min-instances 1
is a latency choice, not a correctness one: a cold start revalidates the baked-in copy with
If-None-Match
and transfers nothing on a 304
, and a boot while GitHub is down serves the local dataset instead of failing. Keep it if you care about cold-start latency, drop it if you care about idle cost.
If you mount a Cloud Storage bucket, do not mount it overdata/
: GCS FUSE hides whatever sits below the mount point, just like any Linuxmount
, and you would lose the file baked into the image. Use a separate path (/mnt/dataset
) and pointDATASET_PATH
at it.
The estimator is honest about what it does not price. These are accepted inputs or dataset features that currently produce no cost line, and they are worth knowing before you trust a total:
| Limitation | Effect |
|---|---|
usage.web_search and usage.search_results |
|
Accepted and validated, but never billed. search_context_cost_per_query (282 models) is declared in the catalog and never read. They do not even show up in unpriced[] . |
|
options.tier_policy: "marginal" |
|
| Does not compute per-band pricing. It falls through to the base rate, so the result is the same as a request that never crossed a threshold. | |
tiered_pricing |
|
Not read. The 21 models that price only through it answer 200 with a total of 0 and all usage in unpriced[] . |
|
| Databricks DBU rates | The input_dbu_cost_per_token / output_dbu_cost_per_token keys (47 models) are described in the catalog but no usage field maps to them, so a DBU total is never produced. |
| Malformed JSON, or a body over 256 kB | Answers 500 internal-error instead of 400 / 413 . |
The catalog is LiteLLM's model_prices_and_context_window.json, MIT licensed, vendored into
data/
and redistributed normalized:
src/data/parse.ts
renames
litellm_provider
to provider
on the way in, and that is the only place where upstream's shape is adjusted. The prices are the ones LiteLLM publishes — this project does not source, correct or negotiate them, and flags the implausible onesrather than fixing them.
bun run dev # start with file watching
bun test # HTTP surface, pricing engine, dataset resolution, catalog completeness
bun run typecheck # tsc --noEmit
Stack: Bun + TypeScript + Express 5. No build step, no database, no
state: src/data/
owns the dataset, src/pricing/
owns the cost engine, src/routes/
owns the HTTP surface.
Issues and PRs are welcome — see CONTRIBUTING.md. The one hard rule: every endpoint and every pricing rule keeps a test that proves it.