Docs Helmcode serves eight open-weight models behind one OpenAI-compatible API at https://api.helmcode.com/v1. Every model works the same way: only the model id changes. Seven of them are covered by your plan; GLM 5.2 is a per-key add-on and has its own section below. Use the id exactly as shown. For limits and quotas, see Rate limits.
glm-5.2 sota flagship 744B MoE · 40B active · 500K context The strongest open model we serve: long-horizon agentic coding over a real repo, plus competition-level maths. Sold per key at €150 a month rather than covered by your plan, with its own section below.
deepseek-v4-flash sota flagship 284B MoE · 21B active · NVFP4 · 1M context Flagship for complex, agentic work: long documents, deep codebases, and multi-step tool use.
qwen3.6 35B MoE · 3B active · NVFP4 · 256K context Built for throughput: speculative decoding delivers ~2× tokens per second.
gemma4 26B MoE · 4B active · NVFP4 · 256K context Google's open architecture, sized to run efficiently — vision and reasoning built in.
qwen3-embedding 8B · 4096 dimensions · Float32 · MMTEB 70.58 Multilingual semantic embeddings for retrieval and search.
rerank Qwen3-Reranker-8B · BF16 · /v1/rerank Cross-lingual reranking: reorders retrieved passages by actual relevance.
kokoro 82M params · <1s latency · 67 voices Real-time text-to-speech with sub-second latency.
whisper-large-v3 99+ languages · 3.2% WER (Spanish) Speech-to-text with automatic language detection.
On Dedicated and On-premise plans you can also run custom or fine-tuned models on hardware reserved for you. See Examples for how to call each family.
add-on · €150 per key a month · runs in the EU
The strongest open-weight model we serve, and the one the catalogue above does not describe well: it runs on our own hardware in the EU with zero logs, like everything above, but a plan does not cover it. You add as many keys with GLM 5.2 as you need, at €150 a month each, on top of any enterprise plan.
| Price | €150 per key a month, on top of any enterprise plan |
| Requests per minute | 100 rpm |
| Concurrency | 5 concurrent requests |
| Tokens per minute | 3M tpm |
| Context | 500K per request |
The model takes 1M of context. We serve 500K, which is what keeps concurrency high and holds the price at €150.
Volume is capped per organisation and shared across every key on the account, separately from the plan quota that covers the models above:
| Plan | GLM 5.2 tokens a month |
|---|---|
| Starter | 2,000M |
| Growth | 6,500M |
| Scale | 18,000M |
When the allowance runs out, GLM 5.2 pauses until the next billing cycle and everything else keeps answering. There are no overage charges.
Why we serve it, what it costs against paying per token, and where it is not your best option
separate offer · paid from credit · runs outside the EU
The same API also serves nine frontier models from three US providers. This is a separate offer from the open-weight catalogue above, and they differ in the two things that decide a model for most buyers: who runs the hardware, and what pays for the tokens.
Nothing above changes. Our own models keep running on hardware we operate in the EU, and a monthly plan keeps covering them at a flat rate.
They run on the provider’s infrastructure, in the United States
Each of these models runs on Anthropic, Google or OpenAI. Your messages, attachments and the replies are sent to whichever of them serves the model you pick, and processed on that provider’s infrastructure outside the EU. Our own models run on hardware we operate in the EU and nothing leaves it. The console shows the same notice next to the model you are choosing, because that is the moment the decision gets made.
Paid by the token, out of prepaid credit
No monthly plan covers them, at any tier, with any allowance left. An organisation with a plan and no credit still sees them and can select one, and the request is refused until the balance is topped up. Your plan allowance is left untouched and keeps working on our own models.
Provider list price, no markup on inference
On input, output and cache reads we pass each provider’s list price through with nothing added. What we publish is exactly what we debit, which is what lets you reconcile a statement against the rate card line by line. Our revenue on these models is the fee charged when you buy credit, not your consumption.
Same catalogue and same key as everything above. They are returned by the model list endpoint alongside our own, and the id is the only thing that changes.
| Model | Provider |
|---|---|
claude-fable-5 | Anthropic US |
claude-opus-5 | Anthropic US |
claude-sonnet-5 | Anthropic US |
claude-haiku-4-5 | Anthropic US |
gpt-5.6-sol | OpenAI US |
gpt-5.6-terra | OpenAI US |
gpt-5.6-luna | OpenAI US |
gemini-3.6-flash | Google US |
gemini-3.5-flash-lite | Google US |
The rates are not published here, deliberately. A rate lives in exactly one place, and that place is the thing that charges it: your console, under Credits, on the Private models tab. There you get the full price list, the four rates per model (input, output, cached input and cache write), and what each model has cost you so far.
Open the price list in the console
Those rates move, and it is worth knowing why, so that a change does not read as a whim of ours. We pass each provider’s list price through with no margin, which means we are not the ones setting it: when a provider reprices, ours follows the same day. And because we charge in euros while the providers publish in dollars, the euro amount also follows the euro to dollar reference rate. Both of those sit outside our control, which is exactly why the number belongs in one always-current place rather than on a page that can quietly go stale between deploys.
A prompt cache lets a model re-read a long prefix (a system prompt, a document, a codebase) without paying the full input rate for it again. OpenAI and Google charge a reduced rate to read one and nothing at all to create it. Anthropic charges a premium to write and a discount to read, which makes the write a fourth billed class rather than a discount, and that is why the price list carries a cache write column that is empty for the other six models.
We publish and debit one Anthropic cache write rate, whatever the lifetime of the entry. Anthropic itself bills less for a short-lived cache, so on those writes our rate sits above theirs. We publish it instead of folding it into the input rate, because a rate we charge and do not show would be a markup in everything but name. Your statement lists all four quantities separately, so the arithmetic stays checkable: quantity times published rate, for each of the four, adds up to what was charged.
Same base URL, same key, same request body. Only the model id changes.
curl https://api.helmcode.com/v1/chat/completions \
-H "Authorization: Bearer $HELMCODE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "claude-sonnet-5",
"messages": [{"role": "user", "content": "Hello"}]
}'If the balance runs out the API answers 402 credits_exhausted and only these models refuse: everything your plan covers keeps answering, so one refusal is not an outage. The balance never goes negative. When a request costs more than what is left we charge what was there and absorb the difference.
How credit works: buying it, the purchase fee, running out and reconciling it