ITBook a call
ProductWhy VenticHow it worksPlatformResellers and providersCost comparisonConsole
Pricing
DeployOn-premiseDeploy a demo
Language: ITBook a call
Cost comparison

How fast,
for how many, at what price.

Open models with Ventic, on rented GPUs or in-house, against APIs and plans. November 2026 prices.

The economics

Same agents,
a fraction of the spend.

Four ways to give AI to a development team, with investment, monthly cost and three-year total.

Developers
Today

Claude Max or ChatGPT Pro plan

Investment
—
Per month
$2,000
Speed per user
89–186 tok/s
Over 3 years
$72,000

Fixed price, with usage caps

Today

Pay-per-use frontier API

Investment
—
Per month
$2,880
Speed per user
151 tok/s
Over 3 years
$103,680

No caps, cost grows with use

Ventic

Open models in the cloud, frontier when needed

Investment
—
Per month
$959
Speed per user
provider's
Over 3 years
$34,510

−52% vs the plan over 3 years, no investment

Ventic

Purchased local servers, frontier when needed

Investment
$32,604
Per month
$708
Speed per user
21–38 tok/s
Over 3 years
$58,100

4 servers · payback in 26 months vs the plan, in 16 months vs the API

Cumulative spend
  • Claude Max or ChatGPT Pro plan
  • Pay-per-use frontier API
  • Open models in the cloud, frontier when needed
  • Purchased local servers, frontier when needed
For the CFO

The economic conclusions, for 20 developers

−52%Lower AI spend, no investment

$34,510 over 3 years against $72,000 for plans and $103,680 for the API. Operating cost, variable with use.

26 monthsLocal servers: an investment that pays back

$32,604 for 4 servers. Payback in 26 months vs the plan, in 16 months vs the API; 3-year ROI 43%. Bought for data that must stay in.

0Budget overruns

Every request reserves its maximum cost before it runs, per team and per key: spend never exceeds the approved budget.

Novendor lock-in

If a provider raises prices, changes or retires a model, one rule moves the traffic. Negotiating power stays with the company.

Recommendation: start with open models in the cloud, with no investment; add local servers for flows with sensitive data.

When it does not pay: against $20 plans or with teams under 5 people the saving disappears. In the cloud, open models cost less than local servers: the server is justified by data sovereignty.

8 M generated tokens per developer per month, 80% input; $100 plan. Sonnet 5.5 $2/$10, DeepSeek V4.1 Flash $0.30/$1.20, Ryzen AI Max+ 495 192 GB server at €7,499 (6 developers each). Ventic options keep 10% of traffic on the frontier and include the licence. November 2026 prices, to be reconfirmed.

Detailed comparison: GPUs, models and speed
Models

Even through the API, open models cost half to a sixteenth.

$ per million tokens (80% input), saving vs Sonnet 5.5.

  • Claude Opus 5.5frontier, closed · $4.00 / $20.00
    $7.20+100%
  • Claude Sonnet 5.5frontier, closed · $2.00 / $10.00
    $3.60baseline
  • Kimi K3.1 Pro2.8T, 104B active, open weights · $2.00 / $10.00
    $3.60same
  • GLM-5.3753B, 40B active, open weights · $1.40 / $4.40
    $2.00−44%
  • DeepSeek V4.1 FlashMoE, open weights · $0.30 / $1.20
    $0.48−87%
  • GLM-5.3 Flash320B, 18B active, MIT · $0.15 / $0.50
    $0.22−94%
  • Qwen3.8 Flashopen version: Flash Next, 125B · $0.15 / $0.47
    $0.21−94%
Hardware

Where each model fits

Minimum GPUs for the weights plus 10% cache. Estimate, not a measurement.

Model8×B300288 GB · 4-bit8×MI300X192 GB · FP88×H10080 GB · FP8495 server192 GB · 4-bit
Kimi K3.1 Pro1.56 TB in MXFP46 GPUno fitno fitno fit
GLM-5.3756 GB in FP82 GPU5 GPUno fitno fit
GLM-5.3 Flash328 GB in FP81 GPU2 GPU5 GPUfits
DeepSeek V4.1 Flash168 GB in FP81 GPU1 GPU3 GPUfits
Qwen3.8 Flash Nextestimate1 GPU2 GPU3 GPUfits
Speed

Alone, faster than Claude and Codex; under load, you pick the trade-off.

Tokens per second seen by one user. More sessions, less speed each.

  • DeepSeek V4.1 Flash · 4×B300 · 1 sessionvLLM measurement
    575 tok/s
  • Kimi K3.1 Pro · 8 GPU · 1 sessionvLLM measurement
    376 tok/s
  • GPT-5.5 medium (Codex)Artificial Analysis
    186 tok/s
  • DeepSeek V4 Flash · 1×MI300X · 1 sessionproduction measurement
    159 tok/s
  • Claude Sonnet 5.5Artificial Analysis
    151 tok/s
  • Claude Opus 5.5Artificial Analysis
    114 tok/s
  • Kimi K3.1 Pro · 8×B300 · 120 sessionsforecast
    100 tok/s
  • DeepSeek V4.1 Flash · 8×B300 · 111 sessionsforecast
    100 tok/s
  • Claude Opus 5.5 · Claude Codemedian over 20,000 turns
    89 tok/s
  • GPT-5 Codex highArtificial Analysis
    78 tok/s
  • Kimi K3.1 Pro · full loadvLLM measurement
    54 tok/s
  • DeepSeek V4.1 Flash · 8×B300 · 2.048 sessionsvLLM measurement
    34 tok/s
  • Claude and Codex (API)
  • Self-hosted measurement
  • Forecast
Capacity

Pick the speed to guarantee

Developers per 8-GPU node at that speed; each generates 30% of the time, node always on.

8×B300 · DeepSeek V4.1 Flash$5.57–9.99 / GPU-hour
Sessions111
Developers370
per developer per month$87.9–158
  • vs Claude Pro · ChatGPT Plus+340% / +688%
  • vs Claude Max 5x · ChatGPT Pro 5x−12% / +58%
  • vs Claude Max 20x · ChatGPT Pro 20x−56% / −21%
8×B300 · Kimi K3.1 Pro$5.57–9.99 / GPU-hour
Sessions120
Developers400
per developer per month$81.3–146
  • vs Claude Pro · ChatGPT Plus+307% / +629%
  • vs Claude Max 5x · ChatGPT Pro 5x−19% / +46%
  • vs Claude Max 20x · ChatGPT Pro 20x−59% / −27%
8×MI300X · DeepSeek V4 Flash$2.55–3.39 / GPU-hour
Sessions16
Developers53
per developer per month$281–374
  • vs Claude Pro · ChatGPT Plus+1307% / +1768%
  • vs Claude Max 5x · ChatGPT Pro 5x+181% / +274%
  • vs Claude Max 20x · ChatGPT Pro 20x+41% / +87%

Kimi K3.1 Pro at work: an 8×B300 node costs as much as Claude Max 5x for 325–583 developers. Below that, use the API; above it, or with data that cannot leave, run it in-house.

Forecasts interpolated between two public vLLM measurements; MI300X from production measurements. Speeds drop with long contexts.

In-house

What runs on the Ryzen AI Max+ 495 server

One session; measurements mostly from Ryzen AI Max+ 395, similar speeds.

  • Claude Sonnet 5.5 (API)baseline
    151 tok/s
  • gpt-oss-120bMXFP4
    55 tok/s
  • Qwen3.8 Flash Next with MTP125B, 6B active
    38 tok/s
  • Qwen3.6 35B-A3BQ8_0
    32 tok/s
  • Qwen3.8 Flash Nextllama.cpp
    21 tok/s
  • GLM-5.3 FlashIQ4_XS, AMD figure
    20 tok/s
  • DeepSeek V4 FlashQ4_K_XL; another source 26
    12 tok/s
  • GLM-5.3 · Kimi K3.1 Prodo not fit
    no fit
Beyond chat

Transcription, time series, classification, speech, embeddings

Same endpoint, same rules. API against self-hosted where a benchmark exists.

Transcription/v1/audio/transcriptions
$ per audio hour
  • Google Chirp$0.960
  • OpenAI Whisper$0.360
  • AssemblyAI$0.150
  • Ventic · H100$0.032–0.068

vs Whisper API: −91% / −81%

Time series/v1/forecast
$ per month
  • TimeGPT Enterprise$12,000
  • Ventic · 1×H100$1,500–3,219

vs TimeGPT Enterprise: −88% / −73%

Classification (System One)/v1/systemone
$ per million decisions
  • Sonnet 5.5$1,050
  • Haiku 5.5$52.5
  • Ventic · MI300X$49.7–66.0

vs Haiku 5.5: −5% / +26%

Speech/v1/audio/speech
$ per million characters
  • ElevenLabs v3$80.0
  • OpenAI tts-1$15.0
  • Venticto be measured

No reliable self-hosted benchmark

Embedding/v1/embeddings
$ per million tokens
  • text-embedding-3-large$0.130
  • text-embedding-3-small$0.020
  • Venticto be measured

No reliable self-hosted benchmark

November 2026 USD prices, to be reconfirmed; euros converted at €0.92 per $1. GPU bands from cheapest to dearest neocloud, spot included. Network, staff, Ventic licence and redundancy excluded.

Get started

Get in touch
to get started.

We help you install Ventic on your infrastructure and start directly from the workload you want to put in production.

Book a technical call