Tradecraft › Library › AI infrastructure
Tradecraft report · · Free sample
OpenRouter: can anyone still compete with the LLM router Stripe just bought?
Short answer
Not head-on. Stripe agreed to acquire OpenRouter on Aug 19, 2026 (reported at $7-8B). OpenRouter makes money from a 5.5% fee on credit purchases, not model markups, routes roughly 70 trillion tokens a week across 457 models, and grows almost entirely through coding-agent integrations and its public rankings. The routing itself is easy to copy and free rivals exist; the openings are quality-based vertical routing, per-customer AI margin tools and regulated routing.
Key facts
- verifiedStripe agreed to acquire OpenRouter on August 19, 2026.
- estimateReported deal value: over $7B (Bloomberg) to about $7.5B (NYT).
- verifiedOpenRouter charges 5.5% on credit purchases (8% on Business) and says there is no markup on inference prices.
- verified457 models are listed in OpenRouter's public models API (October 2026).
- estimateAnnualized revenue was about $160M in August 2026 (Sacra).
- verifiedThe OpenRouter website runs no advertising pixels.
Bottom line
Competing head-on is no longer realistic. Stripe agreed on Aug 19, 2026 to acquire OpenRouter verified, at a reported $7–8B estimate, three months after a $113M Series B at about $1.3B estimate. OpenRouter routes an estimated tens of trillions of tokens a week across 457 models verified, and it is wired into the default settings of the coding agents that drive that volume. Stripe adds payments, fraud handling and billing (it bought Metronome in January estimate).
But the product itself is easy to copy and the fee is under attack. The API is OpenAI-compatible, so switching costs are close to zero inferred. Vercel, Cloudflare and the open-source LiteLLM already offer gateways at 0% markup estimate. OpenRouter's whole take is a 5.5% fee on credit top-ups (8% on Business) verified, and card processing eats a large share of it inferred. The defensible assets are supplier relationships, model breadth on launch day, and the public rankings that make it the place developers look.
The openings are around it, not against it: vertical routers that pick models by task quality (not just price), per-customer AI cost and margin tools for app builders, sovereign/regulated routing, and investing in the independent gateways that are now obvious acquisition targets. Details in section 7.
1. The company
- Founded 2023 by Alex Atallah (CEO, co-founder of OpenSea), Chris Clark and Louis Vichy verified. New York estimate.
- Funding: $12.5M seed led by a16z (Feb 2025) and $28M Series A led by Menlo (Apr 2025), announced together as $40M in June 2025 with Sequoia verified; Series A valuation about $500M estimate. $113M Series B led by CapitalG, May 2026, about $1.3B post-money estimate.
- Exit: Stripe agreed to acquire it on Aug 19, 2026 verified. Bloomberg: over $7B; NYT: about $7.5B including $1.5B for the founders; Axios: over $8B in cash and stock estimate. Earlier WSJ talk of $10B rumour. We found no report that the deal has closed.
- Stripe's framing: "building the economic infrastructure for AI… routing their requests intelligently and spending their tokens efficiently" verified. Forrester calls it Stripe taking "the AI consumption layer" estimate.
2. How it makes money
| Charge | Terms | Tag |
|---|---|---|
| Credit purchases, Standard | 5.5% of the top-up (min $0.80 on non-crypto); crypto 5.0% per the June 2025 fee post | verified |
| Credit purchases, Business | 8%, adds EU/US in-region routing and up to 1,000 workspaces | verified |
| Model prices | Passed through: "There is no markup on inference pricing" (FAQ) | verified |
| Bring your own key (BYOK) | First $25K a month of list-price usage free, then 5%; to be replaced by a fixed monthly subscription (price TBD) | verified |
| Enterprise | Contract pricing, prepaid credits, annual commits, invoicing, fee discounts, SSO/SAML, SLAs, private endpoints | verified |
| Free models | 25+ free models; 50 requests/day, 1,000/day after buying $10 of credits | verified |
Price spot-check: Claude Opus 5, Sonnet 5 and Haiku 5.5 match Anthropic's list prices exactly, including cache reads verified. Gemini 3.8 Flash is about 10% above Google's own page ($0.825/$4.125 vs $0.75/$3.75 per million tokens) verified, probably a different upstream host or tier inferred. So the money is almost all in the top-up fee, and some of it can come from provider choice.
Margin: on a card top-up, roughly 3 points of the 5.5% go to card processing inferred, so the net take on card volume may be only 2–3% inferred. This is one reason a payments company is the natural owner: Stripe pays itself for processing.
3. How big it is
| Date | Tokens | Source | Tag |
|---|---|---|---|
| ~Nov 2025 | ~5T a week | company figure at the Series B | estimate |
| Feb 2026 | ~12.1T a week | third party on X | estimate |
| May 2026 | 25T a week; ~8M users | company, at the Series B | estimate |
| Aug 2026 | more than 10T a day (~70T a week); ~10M developers and companies | Stripe announcement | verified |
| Oct 27, 2026 | top 10 models alone ~118T a week | our sum of the public rankings page | inferred |
Who drives the volume (rankings page, month to Oct 27): Hermes Agent (Nous Research) 55.4T tokens, Claude Code 44.0T, Kilo Code 26.6T, pi 16.8T, Cline 16.7T, Codex 12.6T verified. Coding and agent tools dominate. The top models by tokens are cheap Chinese "Flash" models: DeepSeek V4.1 Flash at 37.3T a week, GLM 5.3 Flash 14.1T, plus free models verified. Claude Opus 5.5 is the top Western premium model at 8.1T a week verified.
- Sacra estimates revenue at ~$50M annualized at end of 2025 and ~$160M in Aug 2026 estimate.
- $160M ÷ 5.5% fee ≈ $2.9B a year flowing through. With some BYOK and crypto at lower rates, ~$3–3.6B inferred.
- Cross-check: ~70T tokens a week ≈ 3.6 quadrillion a year. $3B ÷ 3.6 quadrillion ≈ $0.80 per million tokens, plausible given how much traffic is cheap and free models inferred.
- At a reported $7.5B, Stripe paid roughly 45–50× current revenue inferred, a bet on 10×-a-year volume growth (Stripe's letter) continuing.
4. How it grows: almost no marketing
- No paid acquisition. The site runs only Google Analytics, Cloudflare and Clerk (login): no ad pixels verified. Meta's Ad Library shows no Page advertising to openrouter.ai verified. Paid search is about 2% of keywords estimate.
- Built into developer tools. Roo Code, Kilo Code, Zed and SillyTavern ship it as a provider (OpenRouter writes the guides), Cline per tutorials verified. Its npm package for Vercel's AI SDK had 12.0M downloads in 30 days, its TypeScript SDK 3.6M; the Python package 1.64M verified. 336K GitHub repos mention it in their README verified.
- The leaderboard loop. Apps that send two headers (site URL and name) get a public ranking slot and an app page verified. Developers browse the rankings to choose tools, so builders opt in for free distribution, and OpenRouter becomes the industry's scoreboard inferred. Rankings data is published under CC BY 4.0, which invites press and analysts to cite it verified. The a16z "100 Trillion Token" study (Dec 2025) was built on it estimate.
- Stealth model launches. Labs test unreleased models free on OpenRouter under code names (Quasar Alpha in Apr 2025, then Horizon, Sonoma) verified. Each launch spikes traffic and search: its top non-brand keyword is a stealth model name estimate.
- Traffic: 36.1M visits in Sep 2026 per Semrush, up 54% in two months estimate; Similarweb says ~7M a month and flat estimate. The tools disagree 4–5×, but both show long sessions (10+ minutes) typical of logged-in developers estimate. Similarweb: US 16%, India 11%, China 8% estimate. Discord: ~55,000 members verified.
5. Who already competes
| Competitor | What it is | Fee model |
|---|---|---|
| Vercel AI Gateway | Gateway inside Vercel's platform | 0% token markup, BYOK included estimate |
| Cloudflare AI Gateway | Caching, logs, rate limits at the edge | Free with own keys; 5% only on its unified billing credits estimate |
| LiteLLM | Open-source proxy you host; enterprise tier | 0% (you pay your own servers) estimate |
| Portkey | Governance and observability gateway | Subscription from ~$49/mo, no token fee estimate |
| Helicone | Observability plus gateway | 0% on tokens; plans $79–$799/mo estimate |
| Requesty | Closest like-for-like router | ~5% markup estimate |
| Martian, Not Diamond, Unify | Smart routing by quality and cost | not verified |
| Together, Fireworks, Groq, DeepInfra | Model hosts (also OpenRouter's suppliers) | sell their own tokens |
| AWS Bedrock, Google Vertex, Azure Foundry | Multi-model inside cloud contracts | bundled into cloud spend |
The fee ceiling is now set by free rivals inferred. OpenRouter's response is visible: BYOK is free up to $25K a month and moving to a flat subscription verified.
6. How hard is it to compete?
| Piece | Difficulty | Why |
|---|---|---|
| The API and routing | Easy | OpenAI-compatible, open-source options exist (LiteLLM). A small team can ship a router in weeks inferred |
| Price | Easy, but pointless | Rivals already charge 0%; you can't undercut zero inferred |
| Model breadth on launch day | Hard | Needs contracts and capacity with 80+ providers and labs inferred |
| Being the scoreboard | Very hard | The rankings are the industry reference; they compound with volume inferred |
| Being built into coding agents | Very hard | Cline, Kilo, Roo and others already default to it; Claude Code and Codex traffic flows through it verified |
| Payments, fraud, credits at scale | Very hard now | Prepaid credit fraud is a known problem (the founders cited it); Stripe is the best-placed company on earth to solve it verified |
| Trust for enterprises | Medium | Data-policy and in-region routing are table stakes; OpenRouter now sells them on Business and Enterprise verified |
7. Opportunities
- Vertical routing by quality, not price. OpenRouter routes by price, uptime and speed; it doesn't judge output quality for a specific job inferred. In practice the gap is large: moving an everyday generation task from a premium model to a cheap open model, with a premium model grading the output, can cut cost by an order of magnitude inferred. A router that tests models on a customer's own task, picks the cheapest one that passes, and keeps re-testing as new models land is a product people would pay for. Martian and Not Diamond chase this generically estimate; a vertical version (support email, sales outreach, coaching) is less crowded inferred.
- Per-customer AI cost and margin for app builders. Most AI apps can't see which users are unprofitable. A drop-in tool that shows margin per customer and caps or re-prices unprofitable users is useful. Note this is exactly Stripe's direction (Metronome + OpenRouter), so it's better as a feature or an acquisition target than a standalone company inferred.
- Regulated and sovereign routing. EU/US in-region routing is a Business-plan feature at OpenRouter verified. A router for healthcare, finance or government that guarantees no-retention providers, audit logs and local hosting has buyers that a Stripe-owned general router may serve slowly inferred.
- Investment angle. With Stripe owning OpenRouter and Cloudflare and Vercel building gateways, the independents (Portkey, Helicone, Requesty, LiteLLM's company BerriAI, Martian, Not Diamond) are now acquisition candidates for clouds, payment companies and observability vendors inferred.
8. What would change this view
- The Stripe deal failing to close, or a regulator blocking it.
- Labs moving launch-day access away from OpenRouter, or the big coding agents switching their default provider.
- A top-down move by OpenAI, Anthropic or Google to give one-key access to each other's models (unlikely) or a cloud making multi-model billing truly free and simple.
- Real revenue figures from Stripe, replacing the Sacra estimates.
9. Tracking and tools: full breakdown
What stands out: an unusually bare site. Google Analytics 4 (G-R8YZRJS2XN) through Google Tag Manager, Cloudflare (CDN, bot protection, web analytics), Clerk for login, Next.js. No ad pixels, no affiliate tools, no email or CRO tools. This matches a product-led company that doesn't buy traffic verified.
Scanned 6 pages in a fresh headless browser on 2026-10-29: 842 requests to 8 third-party hosts, 144 scripts, 16 cookies. Everything below was seen in the page code, network traffic, cookies or the Google Tag Manager container verified.
| What they measure | Tools |
|---|---|
| Website traffic measurement | Google Analytics 4, Cloudflare Web Analytics |
| Paid traffic (ad pixels) | none detected |
| Attribution and server-side tracking | Clerk |
| Affiliate and creator traffic | none detected |
Every tool, by category
| Category | Tool | Evidence |
|---|---|---|
| Web analytics and traffic measurement | Google Analytics 4 | request: googletagmanager.com/gtag/js?id=G- cookie: _ga |
| Web analytics and traffic measurement | Cloudflare Web Analytics | request: cloudflareinsights.com |
| Tag managers and CDPs | Google Tag Manager | global: window.google_tag_manager |
| Fraud, bot and ad-verification | Cloudflare Bot Management / Turnstile | request: challenges.cloudflare.com request: /cdn-cgi/challenge-platform |
| Commerce platform, hosting and CDN | Next.js | global: window.next |
| Commerce platform, hosting and CDN | Cloudflare | request: cloudflare |
| Identity resolution and data enrichment | Clerk | request: clerk.com code: /clerk\.[a-z0-9.-]+/npm/@clerk/ |
Account and pixel IDs
| ID type | Value(s) |
|---|---|
| GA4 measurement ID | G-R8YZRJS2XN |
Tracking parameters their code reads
ref utm_source utm_medium utm_campaign utm_content utm_term ttclid fbclid gclid gbraid wbraid msclkid twclid epik rdt_cid li_fat_id. These show which click IDs and affiliate links the site is built to capture.
Other third-party hosts
| Host | Requests |
|---|---|
t0.gstatic.com | 40 |
analytics.google.com | 3 |
stats.g.doubleclick.net | 1 |
Cookies set
or_statsig_stable_id (openrouter.ai, 365d) · or_nav_arm (openrouter.ai, 365d) · or_consent_region (openrouter.ai, 365d) · _ga (.openrouter.ai, 400d) · cf_clearance (.openrouter.ai, 365d) · generaltranslation.locale (openrouter.ai, session) · generaltranslation.enable-i18n (openrouter.ai, session) · generaltranslation.referrer-locale (openrouter.ai, session) · __client (.clerk.openrouter.ai, 400d) · __client_uat (.openrouter.ai, 400d) · __client_uat_NO6jtgZM (.openrouter.ai, 400d) · _cfuvid (.clerk.openrouter.ai, session) · __cf_bm (.clerk.openrouter.ai, 0d) · _ga_R8YZRJS2XN (.openrouter.ai, 400d) · ph_phc_7ToS2jDeWBlMu4n2JoNzoA1FnArdKwFMFoHVnAqQ6O1_posthog (.openrouter.ai, 365d) · __cf_bm (.openrouter.ai, 0d)
10. Method and sources
Fresh logged-out headless Chrome scan of openrouter.ai (home, models, rankings, pricing, enterprise); Meta Ad Library scan; OpenRouter's public API (models, providers); GitHub, npm, PyPI and Discord public APIs. Blocked or unavailable: Google Trends, X follower count, SpyFu, Google Ads Transparency (no data rendered).
- Stripe newsroom: Stripe agrees to acquire OpenRouter
- Forrester on the deal · Channel Insider ($7.5B report) · Outlook Business (Bloomberg $7B+)
- OpenRouter $40M raise (June 2025) · Series B coverage · Sacra
- OpenRouter pricing · FAQ · Platform fee post · Rankings · App attribution docs · Models API
- Feb 2026 outage postmortem · Quasar Alpha stealth launch · a16z State of AI token study
- Similarweb · Semrush · Roo Code provider docs · Kilo Code partner page
- Gateway fee comparison (vendor blog) · Anthropic pricing · Gemini pricing
Questions
How does OpenRouter make money?
Mainly from a 5.5% fee on credit purchases (8% on the Business plan; 5% on crypto). It passes model prices through without markup and charges 5% on bring-your-own-key usage above $25K a month.
Who bought OpenRouter?
Stripe agreed to acquire OpenRouter on August 19, 2026. The price was not disclosed; Bloomberg reported over $7B and the New York Times about $7.5B.
What are OpenRouter's main competitors?
Vercel AI Gateway, Cloudflare AI Gateway and open-source LiteLLM (all 0% markup), Portkey and Helicone (subscription gateways), Requesty (about 5% markup), and the clouds' own multi-model services (Bedrock, Vertex, Azure).