On 17 September 2026 we changed two things at once in our per-token LLM API: we cut prices by up to 66% and we started passing the prompt-cache discount through — until yesterday the house kept it. This article shows the new numbers and, more importantly, where they come from — because "cheaper" without the arithmetic is worth nothing.
⚡ In short
In the Essentials tier, DeepSeek V4 Flash costs R$0.49 per million input tokens and R$1.09 output — DeepSeek's own API charges the equivalent of R$2.27 and R$6.80 for the same model. A cached input token now costs from 2.5% of the input price (10% on most models). Billed per token in reais, no subscription, no 3.5% IOF tax and no card FX spread.
The price, no hedging
Every figure below is per million tokens, in Brazilian reais, and it is what the API actually charges — our pricing page is regenerated hourly from the live catalogue precisely so the advertised number and the charged number never diverge.
Essentials tier — open-weight models
| Model | Input | Cached | Output | Context |
|---|---|---|---|---|
| DeepSeek V4 Flash | R$0.49 | R$0.098 | R$1.09 | 1M |
| Qwen 3.6 35B | R$0.59 | R$0.197 | R$3.49 | 200k |
| Qwen 3.8 27B | R$0.69 | R$0.23 | R$3.99 | 1M |
| GLM 5.3 Flash | R$1.89 | R$0.378 | R$5.90 | 250k |
| GLM 5.2 | R$7.90 | R$2.09 | R$17.90 | 250k |
The floor of the whole catalogue sits in the Confidential tier: Nemotron 3 Nano at R$0.15 input and R$0.60 output per million — running inside a hardware-sealed enclave where not even the machine operator can read your prompt.
The comparison that matters: the same model, at its maker
DeepSeek V4 Flash is an open-weight model, and DeepSeek sells that same model on its own API. So this is apples to apples. Converting their list price at the day's rate (R$5.15 per US dollar):
| DeepSeek V4 Flash | Input / 1M | Output / 1M |
|---|---|---|
| At GPUBrazil | R$0.49 | R$1.09 |
| On DeepSeek's API (peak hours) | R$2.27 | R$6.80 |
| On DeepSeek's API (off-peak) | R$1.13 | R$3.40 |
That is 4.6x less on input and 6.2x less on output against their peak rate — and still 2.3x and 3.1x less against their reduced off-peak rate, which only applies during part of the day. Here there is no peak hour: the price is the same at 3am and at 3pm.
The cache discount stopped staying with the house
When you resend the same system instruction, the same file or the same conversation history, most of those input tokens are not reprocessed — they come from cache. Until yesterday we charged full input price on them and the saving stayed with us. That changed.
A cached token now costs a fraction of the input price, in the same proportion by which it costs us less — 10% on most of the 48 models, down to 2.5% on the most discounted and up to 33% on the least:
| Model | Fresh input | Cached input | Discount |
|---|---|---|---|
| DeepSeek V4 Flash | R$0.49 | R$0.098 | −80% |
| GLM 5.3 Flash | R$1.89 | R$0.378 | −80% |
| Claude Sonnet 5 | R$15.45 | R$1.545 | −90% |
| Gemini 3.8 Flash | R$5.79 | R$0.579 | −90% |
What that is worth in practice depends on your usage pattern, and we measured ours: across 8,812 real calls from customers with a coding-assistant profile, 77.7% of all input tokens came from cache. For that profile the effective bill drops by close to 60% — without changing a line of code.
The number comes back in the API response, so you can audit it without trusting us:
"usage": {
"prompt_tokens": 51204,
"prompt_tokens_details": { "cached_tokens": 39781 },
"completion_tokens": 412,
"total_tokens": 51616,
"cost_brl": 0.0072
}💡 Cache also makes it faster
Measured on our own traffic: on calls with more than 50% cache, time to first token fell from 14.8s to 3.0s — five times faster, even with larger prompts. A big, stable prompt stopped being both a cost problem and a latency problem at the same time.
Today's cuts, model by model
On the Frontier shelf — the branded models — the change was substantial:
| Model | Output before | Output now | Cut |
|---|---|---|---|
| Claude Opus 5 | R$573.52 | R$193.15 | −66% |
| Claude Haiku 4.5 | R$115.29 | R$38.63 | −66% |
| o3 | R$162.38 | R$61.81 | −62% |
| GPT-5.5 Pro | R$2,919.52 | R$1,390.68 | −52% |
| Claude Sonnet 5 | R$89.96 | R$77.26 | −14% |
The cause is less glamorous than a promotion: our cost calculation used a hand-measured routing factor that was up to 2.5x wrong for some models. We re-measured with real calls, replaced the hand-written number with the value the catalogue itself publishes on every read, and the price fell with it. A price that is wrong upward is our problem, not the customer's — so we fixed it and we are saying so.
Where we are not the cheapest
Being straight about the Frontier shelf
On branded models (Claude, GPT, Gemini) our per-token price sits above the maker's list price. We are not going to pretend otherwise: if your only criterion is the lowest price per token on those models, buying direct from the maker costs less.
What you get here is a different thing: billing in Brazilian reais via Pix, with no 3.5% IOF tax on the international purchase and no card FX spread (usually another 2% to 4%, never itemised on the statement), a price that does not move when the dollar does, the same balance as the GPU instances, a Brazilian invoice and Portuguese-language support. For a team without a corporate international card, or one that needs a local invoice for accounting, that solves a problem a few cents of discount does not.
On the open-weight models in the Essentials tier, the price here is a fraction of what the model's maker charges — and that is where most of our volume happens.
Getting started
The API is OpenAI-compatible. Swapping the base URL and the key is enough:
from openai import OpenAI
client = OpenAI(
base_url="https://gpubrasil.com.br/v1",
api_key="gpub_live_your_key",
)
r = client.chat.completions.create(
model="gpub-fast",
messages=[{"role": "user", "content": "Summarise this contract in 5 bullets."}],
)
print(r.choices[0].message.content)
print("Cost of this call: R$", r.usage.cost_brl)No subscription, monthly fee or minimum commitment: you spend what you consume, deducted from the same reais balance as the GPU instances. The in-panel playground lets you test the models first, and the programmatic API is enabled after the first deposit.
See the full table, always current
48 models with per-token prices, regenerated hourly from the live catalogue.
View per-token pricingFrequently asked questions
What is the cheapest LLM API per token?
In our catalogue the floor is Nemotron 3 Nano at R$0.15 input and R$0.60 output per million. For general use with a 1M context window, DeepSeek V4 Flash at R$0.49 / R$1.09 is the best ratio — roughly 6x below what DeepSeek itself charges for the same model.
Is the cache discount passed through?
Yes, since 17 September 2026. A cached input token costs a fraction of a fresh one — 10% on most models, ranging from 2.5% to 33% at the extremes — and the count comes back in usage.prompt_tokens_details.cached_tokens.
Is it cheaper than buying direct from the maker?
On open-weight models, yes, by a wide margin. On the branded Frontier models, no — there the advantage is billing in reais with no IOF and no FX spread, one balance shared with the GPUs, a Brazilian invoice and Portuguese support.
Do I need a subscription?
No. Per-token billing, no monthly fee, allowance or minimum. Creating an account is free.
Is the API OpenAI-compatible?
Yes, streaming included. Base URL https://gpubrasil.com.br/v1, key gpub_live_, and each call's cost returns in usage.cost_brl.
Conclusion
A per-token AI price only means something with the arithmetic shown: which model, which token, compared against whom. On open-weight models we are a fraction of the maker's own price, and the cache pass-through now cuts the bill again for anyone resending large context. On branded models we do not compete on cents, and we would rather say so than invent a table that does not hold up.
Read next: The brain: per-token inference · Per-token API vs a dedicated GPU · AI GPU cost in Brazil 2026