OpenRouter

Atallah previously co-founded the NFT marketplace OpenSea with Devin Finzer in 2018 and served as its chief technology officer; he stepped down in July 2022 to "build something from zero to one," remaining on OpenSea's board [12]. He then co-founded OpenRouter with Louis Vichy [12]. Per its own About page, the company "started in early 2023 as the first LLM marketplace" [3]. The founding thesis, per later investor framing, was that AI stacks were fragmenting and needed unifying infrastructure [12].

OpenRouter acts as a proxy: it exposes one endpoint (https://openrouter.ai/api/v1) behind one API key, implements the OpenAI API spec for /completions and /chat/completions, and normalizes responses from all providers into OpenAI-style JSON, so any OpenAI-compatible SDK works by repointing the base URL [5][7][8]. It passes through underlying provider pricing while pooling uptime across providers [7]. Providers accessible through the platform include OpenAI, Anthropic, Google, Meta, xAI, and DeepSeek [9][10][12]. The platform supports text, images, and PDFs, and by its 2026 Series B had expanded into image, audio, speech, transcription, embedding, and video models [7][4].

Routing involves two decisions: which model answers a request and which provider serves that model [5]. The default provider-routing rule prioritizes providers with no significant outages in the last 30 seconds, then selects among stable providers weighted by the inverse square of price, with the rest serving as fallbacks [5][6]. The company describes this as "a reliability mechanism first and a cost mechanism second" [6]. An optional Auto Router (openrouter/auto) delegates model selection, classifying each prompt into a task type and ranking candidates by aggregate community spend over a trailing seven-day window, at no surcharge [5]. Model variants such as :nitro (throughput-optimized) and :floor (price-optimized) act as shortcuts for provider sorting [5][7].

Failover operates at two layers: automatic provider-layer failover within a model, and opt-in model-layer fallbacks via a models array [5][6]. Customers are billed only for the model that completes a request [5][6]. Sacra describes an edge-based architecture that adds roughly 25 milliseconds of overhead [8].

On privacy, OpenRouter says it logs request metadata but performs "zero logging of your prompts/completions, even if an error occurs, unless you opt-in," with a 1% usage discount for opting in [7]. A bring-your-own-keys option lets customers keep existing provider agreements; pay-as-you-go accounts include $25,000 per month of such usage before a 5% charge applies [7][5].

OpenRouter charges a fee on credit purchases: 5.5% ($0.80 minimum) via Stripe and 5% via crypto payments, with no markup on underlying provider pricing [7]. Sacra characterizes this as roughly a 5% commission on inference spend flowing through the platform [8]. Sacra identifies three main user types: indie hackers, product teams using it as a drop-in OpenAI replacement, and enterprises applying organization-wide policies [8]. Customers "range from early-stage startups to large multinationals" [12], and integrations include Microsoft VSCode, Zapier, and Cloudflare [12]. A partnership with OpenAI gave OpenRouter early access to frontier models such as GPT-4.1 [12].

Sources give two accounts of the early rounds. Sacra reports a $12.5 million seed led by Andreessen Horowitz, followed by a $28 million Series A led by Menlo Ventures in April 2025 at a $500 million valuation [8]. The Block and TechCrunch report a single $40 million round announced in late June 2025, co-led by Andreessen Horowitz and Menlo Ventures with participation from Sequoia; The Wall Street Journal put the valuation at $500 million, while TechCrunch, citing PitchBook, reported an estimated $547 million post-money [9][12].

In May 2026, OpenRouter raised a $113 million Series B led by CapitalG, Alphabet's independent growth fund, with participation from NVentures (NVIDIA), ServiceNow Ventures, MongoDB Ventures, Snowflake Ventures, and Databricks Ventures, alongside existing investors Andreessen Horowitz and Menlo Ventures [4][9][10]. The New York Times reported the valuation "landed at about $1.3 billion post-money" [9][11], bringing total funding to over $153 million [13]. Atallah said of the round: "Running inference at scale is fundamentally a multi-model problem. The era of picking a single model is over" [10].

In August 2026, Atallah appeared on the 20VC podcast, where the central topic was reports that OpenRouter was in an acquisition process with Stripe for $10 billion; no deal terms or confirmation appear in the source [13].

As of the May 2026 Series B, OpenRouter reported processing 25 trillion tokens per week (100 trillion per month), a fivefold increase from six months earlier, with more than 8 million developers, and said it was "on pace to process over a quadrillion tokens this year" [4][9][10]. Its About page displays higher figures, 200 trillion-plus monthly tokens and 10 million-plus users [3]. The company processed over $100 million in annualized inference spend by May 2025, up from roughly $19 million at the end of 2024 per Sacra (The Block reports $10 million) [8][12]. Sacra estimates annualized revenue reached $50 million by March 2026 [8].

Sacra names Portkey, Martian, and Not Diamond as routing competitors; Together AI, Fireworks AI, and Groq as vertically integrated inference providers; and the cloud hyperscalers (AWS Bedrock, Google Vertex AI, Azure OpenAI Service) as "the most significant competitive threat" [8]. The open-source LiteLLM offers a self-hosted alternative, while OpenRouter is "a router you call: a managed service with a single endpoint in front of 400+ models" [5]. Sacra describes OpenRouter's advantage as data network effects from its token volume, while noting relatively low technical barriers to entry, since core functionality is API proxying and load balancing [8].

In August 2025, a roughly 50-minute database incident "took the whole service down, fallbacks included"; the company acknowledges failover "can't route around OpenRouter itself" and is "a layer of resilience rather than an SLA guarantee" [5][6]. Users have reported edge cases where HTTP 429 errors consumed credits and partial outputs were billed despite errors, and aborting a stream does not stop billing on some providers, including Bedrock, Groq, Google, and Mistral [5][6]. OpenRouter states that no quality score exists to sort providers by output quality, though some serve more heavily quantized variants that underperform [5]. Sacra lists structural risks: dependence on major providers continuing to allow third-party API access, commoditization pressure from open-source and cloud-bundled alternatives, and the possibility that consolidation toward fewer, more capable models reduces the need for multi-provider routing [8]. Under the company's terms, unused credits may expire one year after purchase, platform fees are non-refundable, and cryptocurrency payments are never refundable [7].