OrcaRouter launched free bring-your-own-key routing on July 24, letting developers connect their model-provider accounts while retaining OrcaRouter's adaptive routing and automatic failover. According to OrcaRouter's July 24 release, developers pay no additional platform charge when requests run through their own keys.
The release attributes OrcaRouter's product case to FY Tian, whose title is not disclosed. "Developers should not have to choose between controlling their provider accounts and accessing production-grade routing infrastructure," Tian said. An OrcaRouter technical report published in May lists Fengya Tian among six authors, alongside Zhenghua Bao, Chris Zhang, Zhenjun Chen, Xile Ma and project lead Yi Shi. OrcaRouter has not publicly identified its founder or CEO.
OrcaRouter appears to be the first public infrastructure project from Continuum AI, a research lab that says it was founded in 2026 and remains in stealth. Continuum lists six projects, only two of them public, and describes OrcaRouter as its first output in inference infrastructure. Its published thesis is direct: model selection should move out of application code, allowing each prompt to choose an appropriate model before generation begins.
The price wedge
BYOK matters because the developer keeps the direct account, credits, negotiated pricing and rate limits associated with a model provider. OrcaRouter supplies the layer above those accounts: one API, credential management, routing rules and fallback paths.
A request can start on a developer-owned key and move to another configured key when the first provider returns an error or rate limit. Developers can also place OrcaRouter-managed capacity in the fallback chain. That hybrid setup gives OrcaRouter a chance to become part of a production stack without first asking a team to abandon existing provider contracts or centralize all inference spending on OrcaRouter. The capability and the absence of a BYOK platform fee are OrcaRouter's claims; independent usage or customer data is not public.
The launch is aimed squarely at OpenRouter, the larger model gateway OrcaRouter names in its own announcement. OpenRouter's official materials say BYOK traffic is free within an allowance and carries a 5% fee above it. OpenRouter's pages currently describe that allowance in different ways: its BYOK documentation cites one million requests per month, while its pricing table lists $25,000 of monthly list-price inference for pay-as-you-go accounts. OrcaRouter is offering a simpler headline: no platform fee for BYOK.
That pricing attack arrives after OpenRouter raised a $113 million Series B led by CapitalG on May 28. OpenRouter said at the time that it served more than 8 million developers across over 400 models and had increased weekly volume from 5 trillion to 25 trillion tokens in six months. Those company-reported figures establish the scale OrcaRouter is challenging and the resources available to the incumbent.
Routing is the technical bet
OrcaRouter's documentation describes one OpenAI-compatible endpoint spanning providers including OpenAI, Anthropic, Google, DeepSeek, xAI, Alibaba and Moonshot. Existing OpenAI SDK integrations can switch by changing the base URL, while fallback chains and an automatic router choose among available models. OrcaRouter says developers pay published provider token prices without a markup.
The underlying router is more specific than the usual gateway promise to retry a failed request. OrcaRouter's six-author report describes a contextual-bandit system that evaluates features of an incoming prompt, chooses a model and can learn from feedback after deployment. The system starts with an offline training phase and can continue updating the selected model's routing policy as requests are served.
In a May 20 RouterArena submission, the authors reported that OrcaRouter-Adaptive ranked second with a score of 72.08, 75.54% accuracy and a cost of $1 per 1,000 queries. Those results came from OrcaRouter's own technical report. The report says the deployed submission was not fitted on RouterArena prompts, though initialization relied partly on a curated internal prompt set that has not been released. That limits outside scrutiny of how well the published benchmark represents customer traffic.
BYOK broadens that routing work from choosing a model to choosing a credential and billing path. For an engineering team, a successful route can now depend on task quality, cost, provider health and which account should absorb the charge. OrcaRouter is betting that this decision layer belongs in shared infrastructure rather than scattered retry code inside each application.
Free routing needs a paid control plane
OrcaRouter says its revenue comes from optional subscriptions instead of token markups. OrcaRouter's public pricing lists a free individual tier with three API keys and a Team plan at $499 per month, which adds up to 10 seats, compliance controls, unlimited API keys and priority support. Free BYOK fits that structure: routing attracts workloads, while governance and organizational controls become the product OrcaRouter sells.
The open question is whether that split can finance a reliable gateway. Routing and failover create infrastructure costs even when OrcaRouter collects no token margin, and OrcaRouter has not disclosed revenue, customer count, funding or paid conversion. Continuum AI's public site says its team is growing, while OrcaRouter's LinkedIn page lists San Francisco as its primary location. Neither source establishes who controls OrcaRouter or how much capital supports it.
For Tian and the researchers attached to OrcaRouter, free BYOK is an efficient distribution move. It removes a visible fee, preserves developers' existing provider relationships and puts OrcaRouter in the request path where its routing system can prove its value. The harder job begins after installation: turning free routing into enough paid governance business to support production infrastructure against a heavily funded incumbent.