Bench

Numbers, signed and replicated.

Public benchmarks across price, latency and throughput. Same prompts, same hardware tier, same time window — measured weekly from fra1. Reproduction scripts live in github.com/tomoul/bench.

Latest run May 11, 2026 · 06:00 UTCWindow 7-day rollingRegion Frankfurt (fra1)Live · refreshes weekly
$0.18
Avg cost per 1M tokens
−42% vs. category median
28ms
Median TTFT · chat
−51% vs. category median
184ms
p99 TTFT · chat
−18% vs. category median
99.97%
90-day availability
vs. 99.86 category
Model
Workload
Updated 2 days ago

Cost · per 1M input + output tokens

llama-3.1-70b-instruct · 1k in / 256 out · USD · lower is better
tomoul.aifra1 · serverless
$0.18
together.aius-east · serverless
$0.26
groqus-east · serverless
$0.29
fireworks.aius-east · serverless
$0.35
aws bedrockeu-central-1 · on-demand
$0.65
azure aiswedencentral · paygo
$0.83

Latency · TTFT for streaming chat from Frankfurt

tomoul.ai
p50 / p99
28ms · 184ms
Tokens / s
142 tok/s sustained
together.ai
p50 / p99
71ms · 312ms
Tokens / s
118 tok/s sustained
groq
p50 / p99
22ms · 148ms
Tokens / s
280 tok/s sustained
fireworks.ai
p50 / p99
62ms · 284ms
Tokens / s
114 tok/s sustained
aws bedrock
p50 / p99
124ms · 480ms
Tokens / s
81 tok/s sustained

Specialty · Yoruba TTS · seconds of audio per GPU-second (L40S)

afrispeech-tts · 22kHz · higher is better · the kernel we wrote about here
tomoul.ai · fused kernell40s · zig
51.0× RTF
tomoul.ai · referencel40s · onnxruntime
12.1× RTF
replicatea100 · pytorch
9.2× RTF
huggingface inferencea10g · pytorch
6.1× RTF
Methodology

The rules

Same prompts. Same hardware tier where it exists on the vendor's menu. Same 7-day window. Streaming chat measured from request issued to first token rendered. Long-context measured with the same shared 16k-token prompt set across all vendors.

  • 1,000 requests per data point, distributed across the week
  • Measured from a Frankfurt-resident benchmarker on Hetzner
  • p50 and p99 reported alongside the mean; mean alone is misleading
  • Cost calculated at posted list price as of run start, no commits or reserved discounts
  • Failed requests excluded; failure rate reported separately on the status page

Reproducing locally

Bench is open source. Clone the repo, set your vendor keys, run the makefile. The full suite takes about 4 hours of wall-clock and ~$8 of vendor cost.

# clone, configure, run
git clone https://github.com/tomoul/bench
cd bench
cp .env.example .env  # add TOGETHER_KEY, GROQ_KEY, BEDROCK_*, …
make bench   # 4h · writes results/ as JSON + CSV
make report  # renders the table you see above

What is excluded — and why

We do not benchmark OpenAI / Anthropic / Gemini on this page. Different category — closed models, different price model, different residency story. Our quarrel is with the open-weights serving layer, not the frontier labs. If you want a head-to-head on a specific task, write to [email protected]; we will run it and publish.

We do not run on a customer's behalf or to a customer's specification on this page. That is what a sales call is for.

Signatures

Each run produces a signed manifest (ed25519) committed to the tomoul/bench repo with the raw timing samples. The public key is at tomoul.ai/.well-known/bench.pub. Verify with make verify run=2026-05-11.