AI NewsInfrastructureAnnouncement
Antirion is an open-source LLM gateway in Rust with budgets and guardrails
Antirion is a single-binary LLM gateway that routes calls to 16 providers with budgets, guardrails and logs. The repository has 474 stars.
Image: GitHub
Why it mattersA team running a model-backed feature can replace three separate tools with one. Antirion covers routing, budgets, guardrails and observability.
A team that puts a model in a product usually ends up gluing three tools together. One routes calls between providers. One tracks the budget. One checks the input and the output for problems. A new open-source project called Antirion does all three from one binary. The GitHub repository has 474 stars thirteen days after launch, under an Apache 2.0 licence.
The author is Elias Biondo. The project was first opened on 17 October 2026. The star count is a total since launch, not a weekly gain.
What it covers
The README names four jobs. Routing covers 16 providers. The list names OpenAI, Azure OpenAI, Anthropic, Amazon Bedrock, Google Vertex AI, Gemini and Mistral. It also names Cohere, Groq, Together, Fireworks, DeepSeek, xAI, Cerebras, DeepInfra and Perplexity. The gateway translates between the OpenAI and Anthropic dialects so one caller can reach every provider.
Budgets run in dollars or tokens, daily to yearly, with soft, hard or approval-held limits. The README says the budget can be calendar-based or rolling. Guardrails cover PII, secrets, keywords, regular expressions, JSON schemas, tool permissions, topics, moderation and model-based judges. Observability logs the route taken and every attempt that failed. It also records the token count, the cost, the cache status and the guardrail verdict.
How resilience is built
The README lists five resilience features. Fallback chains, weighted splits, routes by least cost, least latency or least busy, and content-based classifiers. Retries, hedging and circuit breakers sit on top. The author says these are the parts an in-house gateway usually gets last. They arrive after the first outage.
The gateway runs as a single Rust binary. The author lists Docker, Kubernetes and a direct binary install as deployment options. There is no hosted service. The user pays only for the model calls that pass through the gateway.
What the author measures
The README carries a benchmark against another open-source gateway called Bifrost. The author ran 5,000 requests a second on a 4-vCPU machine. The author says Antirion added 0.53 milliseconds at the median, where Bifrost added 19.7 milliseconds on the same hardware. The author also says Antirion sustained 14,000 requests a second as a ceiling, where Bifrost reached 3,500. These are the author's own numbers against a named competitor. A reader comparing the two should run their own test on their own hardware. A vendor benchmark against a named rival is marketing.
Who it is for
A reader who already runs a single-provider API client and never needs failover does not need this. A team running a model-backed feature in production probably does. The common need is three features, routing, budgets and logs, and the gateway covers them from one binary.
Source
eliasbiondo/antirion, on GitHub.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
