Inference gets cheaper the longer you leave it running.

Every other provider prices a static stack. Ours is rewritten by an agent that forms one hypothesis at a time, measures it against a frozen quality gate, and reverts anything that does not win. Fourteen days in, the same model costs a tenth of what it did on day one.

The gate is public. So is every kept commit.

START FREERead the methodology →
LLAMA-3.3-70B · $ PER MILLION INPUT TOKENS
Provider A
0.94
Provider B
0.86
Provider C
0.58
Boole
0.09
COLD START p99
380ms
TTFT p50
62ms
METHODOLOGY: FROZEN OBJECTIVE, ONE MEASUREMENT PER ITERATION2,057 ITERATIONS · 217 KEPTSOC 2 TYPE II
01 — THE MECHANISM

One hypothesis at a time, against a gate that never moves.

The agent proposes a change to the kernel, the scheduler, or the quantisation scheme. It runs the eval suite once. If quality drops below the frozen threshold, the change is reverted and the hypothesis is recorded as dead.

Nothing ships on a hunch. Roughly one in ten attempts survives.

BOOLE LOG · LLAMA-3.3-70BLAST 6 ITERATIONS
#2052fuse rmsnorm + qkv projection-0.3% qualREVERTED
#2053page kv-cache at 64-token granularity+0.0% qualKEPT
#2054int4 weights on layers 0–12-1.9% qualREVERTED
#2055speculative draft depth 5 → 7-0.1% qualREVERTED
#2056batch admission by token budget+0.0% qualKEPT
#2057ternary kernel for mlp down-proj-4.2% qualREVERTED
02 — THE BILL

The same open weights. A tenth of the invoice.

Prices below are live and move down, never up. The rightmost column is what the identical workload cost on the cheapest alternative this week.

MODELIN / MTOKOUT / MTOKTTFT p50CHEAPEST ELSEWHERE
Llama 3.3 70B Instruct$0.09$0.1462ms$0.58 in
Qwen 2.5 72B Instruct$0.11$0.1671ms$0.62 in
Mixtral 8x22B$0.13$0.1958ms$0.65 in
DeepSeek V3$0.08$0.1277ms$0.41 in
Whisper Large v3$0.0244ms$0.11 /hr
03 — COLD START

Scale to zero without the first request paying for it.

Weights stay resident in a shared page cache and the runtime snapshots after warm-up, so a cold container resumes rather than boots. Every stage below is measured on every cold request and published.

scheduler admit
24ms
snapshot restore
137ms
weight page-in
104ms
first token
92ms
COLD START p99 · TRAILING 14 DAYS
1.04s · JUL 27380ms · TODAY
WHAT YOU DO NOT PAY FOR
Idle replicas held open for headroom.Reserved GPU-hours bought in advance.The container's boot time.
04 — MOVING OVER

Change the base URL. That is the migration.

The API is OpenAI-compatible, including streaming, tool calls, structured outputs, and the batch endpoint. Existing SDKs work unmodified.

Full API reference →
PYTHONNODECURL
from openai import OpenAI

client = OpenAI(
    base_url="https://api.boole.dev/v1",
    api_key=os.environ["BOOLE_API_KEY"],
)

resp = client.chat.completions.create(
    model="llama-3.3-70b-instruct",
    messages=[
        "role": "user", "content": "Summarise this ticket.",
    ],
    stream=True,
)

Run the first million tokens on us.

No card, no sales call. If the numbers on this page do not hold for your workload, the log will tell you why.

GET AN API KEYTalk to an engineer →