api.girard-davila.net - now in beta

Blazing-Fast, Ultra-Cheap LLM Infrastructure
Built for Pi Agents

Get your key - free in beta Browse the models

pi - gd/deepseek-v4.1-flash

$ pi --model gd/deepseek-v4.1-flash ◆ using model "gd/deepseek-v4.1-flash" · api.girard-davila.net

> summarize the git log of this repo in one line

◆ Last release: 42 commits - auth rework (SSO + virtual keys), perf: batch embeddings, docs pass. No breaking changes. (1.2s · 384 tok · $0.0009)

Why agents on this endpoint

Fast where it counts. Self-hosted routes are one model on a dedicated GPU - no shared queue, no multi-tenant noise, MTP speculative decoding on top. Cloud routes go straight to DeepInfra’s fastest inference tier.
Honest pricing. The price in the catalog is the price you’re billed - no struck-through list prices, no fine-print margin. Self-hosted routes are at hardware cost-plus.
Cache-aware. Prompt caching at the proxy layer means long, repetitive agent contexts stop paying full price for the prefix.
Yours to lose. Self-hosted, small, and boring on purpose. Your key, your spend, your rate limits - all visible in the dashboard.
Talk to the human. One person runs this infrastructure. Want to explore ideas, shape what comes next, or help keep the infra maintained? Reach Alexandre Girard directly - alexgirard.com.

Join in three steps

1. Create an account

Sign in with Google at api.girard-davila.net. The SSO round-trip issues your personal virtual key automatically - no email verification dance, no manual key creation.

2. Connect Pi

pi install git:github.com/alx/pi-provider-gd

Then inside Pi:

# Sign in - a browser opens
# then log in with Google - your key is created + stored for you
/login gd
/model gd/deepseek-v4.1-flash

3. Top up (when billing launches)

Pick a credit pack at checkout; your existing key keeps working, spend limits included. Nothing to re-install, nothing to re-authenticate.

Model catalog

Budget - High-volume agent loops, drafts, tool calls

Model	Context	Input $/1M	Output $/1M	Features
`llama-3.1-8b-turbo` Tiny, instant, dirt-cheap. Summaries, classification, high-volume agent loops.	131K	0.024	0.048
`gpt-oss-20b` OpenAI's open-weight MoE (3.6B active) with reasoning. Fast answers on a budget.	131K	0.036	0.168	reasoning
`ling-3.0-flash` Token-efficient agentic model tuned for low-latency tool-calling workloads.	131K	0.072	0.216	prompt cachereasoning
`gemma-3-27b-it` Google Gemma 3 27B with vision. Light multimodal work at commodity prices.	131K	0.096	0.192	vision

Mid - Everyday coding and reasoning work

Model	Context	Input $/1M	Output $/1M	Features
`gpt-oss-120b` 117B OpenAI MoE (5.1B active), high reasoning for the price. Workhorse for agent steps.	131K	0.0444	0.204	reasoning
`deepseek-v4-flash` DeepSeek 284B MoE (13B active), 1M context. Fast, strong reasoning, long agentic sessions.	1M	0.108	0.216	prompt cachereasoning
`glm-5.3-flash` Z.ai multimodal flash: efficient coding + long-horizon agent tasks, 1M context, prompt cache.	1M	0.18	0.6	visionprompt cachereasoning

Flagship - Hard problems, long agentic sessions

Model	Context	Input $/1M	Output $/1M	Features
`gemini-2.5-flash` Google thinking model, 1M context, vision. Complex problems with structured reasoning.	1M	0.36	3	visionreasoning
`deepseek-v4.1-flash` DeepSeek 552B MoE, 1M context, vision + reasoning. The all-round flagship.	1M	0.24	0.72	visionprompt cachereasoning
`qwen3.5-27b` Largest dense Qwen3.5: near-frontier reasoning and coding, 262K context, vision.	262K	0.312	3.12	visionreasoning
`kimi-k2.6` Moonshot agentic flagship: long-horizon coding, autonomous execution, swarm tasks.	262K	0.9	4.2	visionprompt cachereasoning

Coding - Agentic code generation

Model	Context	Input $/1M	Output $/1M	Features
`qwen3-coder-480b` 480B MoE (35B active) agentic coder: SWE-bench class coding, browser-use, 262K context.	262K	0.36	1.2	prompt cache

Self-hosted (beta) - Our own GPUs - contact for access

Model	Context	Input $/1M	Output $/1M	Features
`qwen3.8-27b` Dedicated single-tenant GPU - RTX 3090 24 GB (standard) or RTX 5090 32 GB (premium) - IQ4_XS GGUF, MTP speculative decoding. Internal exploration team (beta).	131K	contact

Models tagged prompt cache bill cached input tokens at a deep discount. Self-hosted routes are for the internal exploration team during beta. Prices are USD per 1M tokens and refresh automatically from DeepInfra. ⚠️ The self-hosted GPU is dedicated, but under load Qwen3.8 requests overflow to rented third-party VAST.ai capacity. Catalog generated 2026-10-14T10:25:50Z.

Self-hosted GPU (beta)

A dedicated GPU running qwen3.8-27b for you alone - on-demand, hourly cost-plus, automatic shutdown when idle. Limited beta slots.

You'll get a per-hour quote by email (2× hardware cost, standard or premium tier). The GPU starts on demand and shuts down automatically after 30 minutes idle.

How token billing works

Buy a credit pack (Stripe) - e.g. $5 → 500k tokens at the pack’s fixed rate. Packs never expire while the account is in good standing.
Spend is metered per token on your virtual key - input + output, at the published rate, with caching discounts applied automatically.
You see everything in the dashboard: spend per model, per key, per day. Hard spend caps are enforced, not advisory.

Beta: while we’re in beta, new accounts get free credits - no card required. Stripe top-ups go live with the public launch.

Terms & privacy

By using the service you agree to our Privacy Policy. Terms of Service (B2B and B2C), the Art. 28 Data Processing Agreement, the sub-processor list and the Acceptable Use Policy are being finalised with counsel and will be published here before B2B launch - until then the service is offered on a best-efforts beta basis. Data handling in one line: no prompt or output bodies are logged at our layer - metadata only (model, token counts, timestamps, key ID). Upstream inference (DeepInfra) is zero-data-retention by contract; the rented-GPU tier is not for sensitive data (see above).

Blazing-Fast, Ultra-Cheap LLM InfrastructureBuilt for Pi Agents