// just use open models

Your AI bill can drop 90%.
Without racking a single GPU.

Open models do most of the work you send to a frontier API, at a tenth of the price. Helmcode runs them in the EU, flat rate, zero logs.

see_pricing how Revolut did it

The same team, the same usage. Per-token spend grows with every good use you find; a flat rate per API key is the same line every month.

-90% on the AI bill

∞ tokens, flat rate

EU infrastructure

0 logs

// 01 · why it grows

You pay per token. The counter only goes up.

Every good use you find adds slope: a chat, then retrieval, then agents that burn ten times the tokens. So you put a cap on it, and the cap eats the return.

Tokens billed over a year, cumulative, order of magnitude. Each new use is steeper than the last; the budget cap stops the counter, and everything above it is the part of the job the model was not allowed to do.

// 02 · the proof: Revolut

At Revolut, the default AI models are already open. Claude and GPT are the exception.

2,000 engineers at a regulated EU bank. Since April 2026 they work on GLM, an open model, by default; Claude Code and Codex stay for the few tasks they do best. The bill is a tenth, and the work got faster.

1/10

of the bill

-60%

coding time

-70%

lead time

$0.30

per M tokens

Lead time and coding time across 2,000 engineers, indexed. Three steps: Gemini, Claude Code with budget caps, GLM as the default. Redrawn from Revolut's chart: shapes, not measurements. watch the talk →

"We keep them as the exception."
David Abitbol, AI operations at Revolut, on Claude and Codex. AI Engineer NYC, 2026.

// 03 · the catch

You do not need to buy GPUs to host your models. We serve the open models for you.

Revolut racked 56 B200s and spent five months tuning them. With 2,000 engineers that pays for itself; with 20, it does not. On Helmcode the same models are two lines away.

self-host

What it took Revolut

56 × NVIDIA B200to serve one model to 2,000 engineers

April to August tuning: TTFT 10 s → 1.6 s
NVFP4 · TP8 · cache hit > 90%
sandbox · telemetry · regulator
a platform team

helmcode

What it takes here

two lines

client = OpenAI(
  api_key="sk-...",
  base_url="https://api.helmcode.com/v1",
)
r = client.chat.completions.create(
  model="glm-5.3",
  messages=messages,
)

GLM 5.3 · DeepSeek V4.1 Flash · Qwen 3.6 · Gemma 4
tuned on B200 · OpenAI-compatible
flat rate per API key
EU · zero logs

// 04 · the numbers

Same workload. Four invoices.

Scenario: 5B tokens a month · 80% input, 20% output · official prices, November 2026

Provider Relative monthly spend Monthly cost

Anthropic claude-opus-5-5

~$36,000/mo

OpenAI gpt-6.1-sol

~$18,000/mo

Google gemini-4-argon

~$18,000/mo

Helmcode deepseek-v4-flash · Starter

5B tokens/month cap, inside the plan

€399/mo

Starter is €399 a month and includes 5B tokens of DeepSeek V4.1 Flash, with Qwen and Gemma unlimited on top. GLM 5.3, the model Revolut runs, is an add-on at €150 per key.

// 05 · helmcode

Helmcode. The European AI cloud. Open models, flat rate, zero logs.

One OpenAI-compatible API on B200 GPUs in EU data centres. You pay per API key, not per token; the request below never leaves the EU and is never stored.

client Your app any OpenAI-compatible SDK

entry Helmcode API TLS 1.3 · API key

inference Helmcode GPU in memory · no logs open weights loaded DeepSeek V4.1 FlashGLM 5.3Qwen 3.6

plans

Starter

€399/mo

For small teams starting to integrate AI into their workflows.

5 API Keys
Unlimited tokens · Qwen and Gemma
5B tokens/month · DeepSeek V4.1 Flash and GLM 5.3 Flash
OpenAI API compatible
Zero logs · EU data

get_started →

// get started

START BURNING TOKENS

Skip the AI infra work. Deploy your first private inference endpoint today.

Flat rate. EU data. OpenAI API compatible.

book_a_call

Your AI bill can drop 90%.Without racking a single GPU.