eigenvector↗Let’s talk
+ 04 / THE ECONOMICS

More tokens.
Fewer dollars.↗

Cost per token is GPU time divided by the tokens it produces. Raise throughput on your workload and every token gets cheaper, on the same hardware and at the same latency target.

THROUGHPUT IS THE PRICE

More tokens per GPU.
Fewer dollars per token.

Cost per token is GPU time divided by the tokens it produces. Raise throughput on your workload and every token gets cheaper, on the same hardware.

GPUS NEEDED14 → 3
COST PER 1M OUTPUT TOKENSUSD
Claude Sonnet 5
$10.00
Claude Haiku 4.5
$5.00
Standard serving on H100
$4.72
Gemini 3.8 Flash
$3.75
Tuned stack on H100
$0.69
Standard serving$0
Tuned stack$0
Saved every year$0

Throughput uplift ×6.8 measured on a creative writing workload. Standard serving: SGLang without speculative decoding, 8K in / 1K out, 64 concurrent requests (SemiAnalysis InferenceX). GPU counts assume 60% average utilisation. List prices as published on 24 September 2026. Public APIs for open-weight models can list lower prices; they are shared services, not tuned to your workload.

Benchmark your own workload
WHERE THE SAVINGS COME FROM

Four workloads.
Four cost curves.

The more your traffic repeats and the more predictable its outputs, the further the cost per token falls.

WORKLOAD PROFILE

Creative writing

Long stories and character dialogue at high temperature. Every reply is new prose, and readers pause long enough between turns for ordinary caches to forget the story.

MODEL
Qwen3.8-Flash-Next
HARDWARE
4× H100
LATENCY TARGET
≥ 40 tok/s per user
×0.0output throughput per GPU
Output throughput↑ 6.8×
STANDARD
760
TUNED
0
tokens / s / GPU
Cost per 1M output tokens↓ 6.8×
STANDARD
$1.46
TUNED
$0.00
USD · at $3.99/GPU-hour
Time to first token, p50↓ 7.9×
STANDARD
1.90
TUNED
0.00
seconds
Prompt tokens from cache↑ 4.1×
STANDARD
22
TUNED
0
% of prompt tokens
COMPARE YOUR MONTHLY BILL

Already paying for an API? Put your current spend next to an estimated monthly fee for reserved, workload-tuned inference capacity.

MAKE THE NUMBERS YOURS

Your workload.
A different equation.

Compare usage-based API spending with an estimated monthly fee for equivalent managed inference and reserved capacity.

$
$
$
YOUR COST SCENARIOUSD / MONTH
1/30*

of your current API spend

Current API spend$30,000
Managed inference scenario$1,000
Potential annual savings$348,000
↘ 96.7%

* Illustrative scenario, not a quote or measured benchmark. Adjust all inputs to your workload. No savings or equivalent performance are guaranteed. Eigenvector operates the servers; reserved capacity is delivered through authenticated inference APIs.

Get a workload-specific assessment
MINIMAX H3 / INTERNAL SERVING-COST MODEL

What would 1/30 require?

A measured 5.17-second video took 93.19 seconds of generation. At the provider’s reference output-only rate of $0.08/second for 768p, the output would cost about $0.4133.

Reaching 1/30 of that requires an effective generation-compute cost of about $0.532 per GPU-hour or less. Explore the assumptions below.

Provider pricing reference
Modelled compute cost / clip$0.0129API / compute ratio31.9×

This models internal inference-serving costs, not a customer GPU rental offer or a Eigenvector service price. Excludes operations, idle/cold-start overhead beyond the selected utilisation, storage, transfer, licences, failed jobs and margin. Local INT8 + Turbo8 quality is not assumed equivalent to the full hosted pipeline.

THE MATH, IN PLAIN SIGHT

Cost per 1M output tokens = GPU cost per hour ÷ (output tokens per second per GPU × 3,600) × 1,000,000. GPUs needed = monthly tokens ÷ (throughput per GPU × seconds per month × utilisation). All amounts are in USD and exclude taxes. Capacity is delivered through Eigenvector’s managed inference APIs; these inputs are not GPU rental prices or server access charges.

COMPARE LIKE WITH LIKE

The details
make the difference.

A useful cost comparison accounts for the entire workload, not just a GPU-hour.

01

Match the output

Use the same model version, quality target and precision. Speculative decoding keeps the target model’s output distribution; quantisation needs its own evaluation.

02

Hold the latency target

Throughput is only comparable at the same per-user speed and time to first token. Batching harder is not the same as serving faster.

03

Measure actual demand

Utilisation matters. Cache hit rates and acceptance lengths come from your traffic, so we measure them on your traffic.

A note on MiniMax H3.

Eigenvector-operated H3-Base inference and the official hosted 2K pipeline have different component boundaries. A fair comparison must specify which generation, preprocessing, reference-input and regeneration costs are included.

See the model provider’s current pricing ↗
THE NEXT MOVE IS YOURS

Bring your workload.
We’ll bring the numbers.

Benchmark your workload

One benchmark.
A whole new cost curve.