modelping

⚡ Benchmark LLM, STT, TTS, and full voice pipeline latency across every major AI provider.

One tool. Every provider. The metrics that actually matter.

---

---

What It Measures

Category	Metrics
LLM	Time to First Token (TTFT) P50/P95/P99, tokens/sec, cost
STT	Transcription latency, time to first partial transcript
TTS	Time to First Audio Byte (TTFB), realtime factor
Pipeline	Full STT→LLM→TTS end-to-end latency (the hero metric)

All measurements use real streaming requests — latency is captured at the byte level.

---

Installation

pip install modelping

Or install from source:

git clone https://github.com/LatencyGrid/modelping
cd modelping
pip install -e .

---

Quick Start

# Copy and configure API keys
cp .env.example .env
# Edit .env with your keys

# Run the full voice pipeline benchmark (hero feature)
modelping pipeline --stt groq-whisper-large-v3 --llm llama-3.3-70b-versatile --tts cartesia-sonic-2

# Benchmark LLMs head-to-head
modelping run gpt-4o claude-3-5-sonnet-20241022 gemini-2.5-flash

# Benchmark STT providers
modelping stt

# Benchmark TTS providers
modelping tts

---

Sample Output

Pipeline (STT→LLM→TTS)

$ modelping pipeline --stt groq-whisper-large-v3 --llm llama-3.3-70b-versatile --tts cartesia-sonic-2

╭──────────────────────────────────────────────────────────────────────────────╮
│  modelping pipeline  •  3 runs                                               │
╰──────────────────────────────────────────────────────────────────────────────╯

 STT                     LLM                      TTS              STT    LLM    TTS   Total
 ─────────────────────────────────────────────────────────────────────────────────────────────
 groq/whisper-large-v3   groq/llama-3.3-70b        cartesia/sonic-2  182ms  44ms   91ms   317ms

✓ Pipeline tested  •  fastest total: 317ms

LLM

$ modelping run gpt-4o claude-3-5-sonnet-20241022 gemini-2.5-flash llama-3.3-70b-versatile

╭─────────────────────────────────────────────────────────────────────────────────╮
│  modelping  •  5 runs  •  prompt: 64 tokens                                     │
╰─────────────────────────────────────────────────────────────────────────────────╯

 Model                          Provider     TTFT P50   TTFT P95   Tok/s   Cost/1M
 ─────────────────────────────────────────────────────────────────────────────────
 llama-3.3-70b-versatile        groq           42ms       67ms    312.4     $0.79
 gemini-2.5-flash               google        936ms     1100ms     45.2     $0.60
 claude-3-5-sonnet-20241022     anthropic     198ms      234ms     71.1    $15.00
 gpt-4o                         openai        312ms      489ms     82.3    $10.00

✓ 4 models tested  •  12.3s total

Colors: 🟢 green = fastest, 🟡 yellow = mid, 🔴 red = slowest (relative to the tested set).

STT

$ modelping stt --runs 3

 Model                    Provider     Latency P50   Latency P95   Words
 ────────────────────────────────────────────────────────────────────────
 whisper-large-v3-turbo   groq            180ms         210ms        9
 nova-2                   deepgram        240ms         290ms        9
 whisper-1                openai          890ms        1100ms        9

✓ 3 providers tested

TTS

$ modelping tts --runs 3

 Model                    Provider      TTFB P50   TTFB P95   Realtime
 ──────────────────────────────────────────────────────────────────────
 sonic-2                  cartesia         89ms      112ms      14.2x
 eleven_flash_v2_5        elevenlabs      210ms      267ms       8.1x
 aura-asteria-en          deepgram        198ms      245ms       9.3x
 tts-1                    openai          312ms      398ms       6.4x

✓ 4 providers tested

---

Full CLI Reference

# LLM benchmarks
modelping run gpt-4o claude-3-5-sonnet-20241022 gemini-2.5-flash
modelping run --all
modelping run --provider groq
modelping run gpt-4o --runs 10
modelping run gpt-4o --prompt "custom prompt"
modelping run gpt-4o --json
modelping run gpt-4o --csv
modelping run gpt-4o --fail-above-ttft 500

# STT benchmarks
modelping stt
modelping stt groq-whisper-large-v3 deepgram-nova-2
modelping stt --runs 5

# TTS benchmarks
modelping tts
modelping tts cartesia-sonic-2 elevenlabs-flash
modelping tts --runs 5
modelping tts --text "Custom text to synthesize"

# Pipeline benchmark (full STT→LLM→TTS)
modelping pipeline
modelping pipeline --stt groq-whisper-large-v3 --llm gpt-4o-mini --tts cartesia-sonic-2
modelping pipeline --stt all --llm all --tts all
modelping pipeline --runs 3

# List available models
modelping models
modelping models --provider anthropic

---

Supported Providers

LLM

Model	Provider	Input $/1M	Output $/1M
gpt-4o	openai	$2.50	$10.00
gpt-4o-mini	openai	$0.15	$0.60
o3-mini	openai	$1.10	$4.40
claude-3-5-sonnet-20241022	anthropic	$3.00	$15.00
claude-3-haiku-20240307	anthropic	$0.25	$1.25
gemini-2.5-flash	google	$0.15	$0.60
llama-3.3-70b-versatile	groq	$0.59	$0.79
mixtral-8x7b-32768	groq	$0.24	$0.24
accounts/fireworks/models/llama-v3p1-70b-instruct	fireworks	$0.90	$0.90
meta-llama/Llama-3.3-70B-Instruct-Turbo	together	$0.88	$0.88
mistral-large-latest	mistral	$2.00	$6.00
mistral-small-latest	mistral	$0.10	$0.30
command-r-plus	cohere	$2.50	$10.00
command-r	cohere	$0.15	$0.60

STT

Model	Provider
whisper-large-v3	groq
whisper-large-v3-turbo	groq
distil-whisper-large-v3-en	groq
whisper-1	openai
gpt-4o-transcribe	openai
nova-2	deepgram
nova-3	deepgram
best	assemblyai
nano	assemblyai
(default)	gladia

TTS

Model	Provider
eleven_flash_v2_5	elevenlabs
eleven_multilingual_v2	elevenlabs
sonic-2	cartesia
sonic-english	cartesia
tts-1	openai
tts-1-hd	openai
(streaming)	fish-audio
PlayDialog	playht
Play3.0-mini	playht
aura-asteria-en	deepgram
aura-luna-en	deepgram
blizzard	lmnt
aurora	lmnt

---

Configuration

Set API keys in a .env file in your working directory:

# LLM providers
OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...
GOOGLE_API_KEY=AIza...
GROQ_API_KEY=gsk_...
FIREWORKS_API_KEY=fw_...
TOGETHER_API_KEY=...
MISTRAL_API_KEY=...
COHERE_API_KEY=...

# STT providers
DEEPGRAM_API_KEY=...
ASSEMBLYAI_API_KEY=...
GLADIA_API_KEY=...

# TTS providers
ELEVENLABS_API_KEY=...
CARTESIA_API_KEY=...
FISH_AUDIO_API_KEY=...
PLAYHT_API_KEY=...
PLAYHT_USER_ID=...
LMNT_API_KEY=...

modelping auto-detects configured providers and skips (with a warning) any without keys set.

---

CI/CD Example

GitHub Actions

name: AI Latency Check

on:
  schedule:
    - cron: '0 */6 * * *'   # every 6 hours
  workflow_dispatch:

jobs:
  latency-check:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      - name: Set up Python
        uses: actions/setup-python@v5
        with:
          python-version: '3.11'

      - name: Install modelping
        run: pip install modelping

      - name: Run LLM latency benchmark
        env:
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
          GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
        run: |
          modelping run gpt-4o claude-3-5-sonnet-20241022 --runs 3 --fail-above-ttft 1000

      - name: Check voice pipeline latency
        env:
          GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
          CARTESIA_API_KEY: ${{ secrets.CARTESIA_API_KEY }}
        run: modelping pipeline --stt groq-whisper-large-v3 --llm gpt-4o-mini --tts cartesia-sonic-2 --fail-above-ttft 500

      - name: Export JSON results
        env:
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
        run: |
          modelping run --provider openai --json > results.json

      - name: Upload results
        uses: actions/upload-artifact@v4
        with:
          name: latency-results
          path: results.json

---

Roadmap

• LLM benchmarking (TTFT, throughput, cost)
• STT benchmarking (transcription latency)
• TTS benchmarking (time to first audio byte)
• Full STT→LLM→TTS pipeline benchmark
• Community leaderboard — submit anonymous results, see global rankings
• Web UI — run benchmarks from your browser (bring your own keys)
• Self-hosted / open source model endpoints
• Historical tracking and latency alerts

---

Contributing

See CONTRIBUTING.md for instructions on adding a new provider.

PRs welcome for:

• New providers
• New models / updated pricing
• Output improvements
• Bug fixes

git clone https://github.com/LatencyGrid/modelping
cd modelping
pip install -e .

---

License

MIT