E2B, which gives AI agents their own cloud computers, passes $120 million in annual recurring revenue with 100% inbound. The San Francisco company runs secure, isolated computers where AI agents write code, browse the web, and finish tasks for businesses. Ninety-four of the Fortune 100 have signed up Manus found E2B organically and scaled with no sales calls or demos AI agents started 2 billion computers on E2B by October, double the June... November 5, 2026 at 10:09 a.m. Media error: Format(s) not supported or source(s) not found * The San Francisco company runs secure, isolated computers where AI agents write code, browse the web, and finish tasks for businesses. * Ninety-four of the Fortune 100 have signed up * Manus found E2B organically and scaled with no sales calls or demos * AI agents started 2 billion computers on E2B by October, double the June total E2B, which builds secure cloud computers for AI agents, today announced it has passed $100 million in annual recurring revenue (ARR). This year alone, they're projected to add over $150M of net new revenue. The company was at $8m just a year ago. The milestone marks E2B's first public revenue disclosure and comes almost 100% from ALG: Agentic-Led Growth. Companies building AI products use E2B to give their agents programs that carry out tasks on their own, a computer to work on, so they can open files, run software and browse the web the way a person would. Each computer runs a single task, walled off from other tasks and from the customer's own systems. Developers have discovered E2B through its open-source tools, word of mouth and, increasingly, recommendations from AI assistants. Close to 200,000 teams use E2B. Manus found E2B organically, signed up through E2B's self-serve platform, and scaled from there, with no sales calls or demos. Usage is accelerating. E2B crossed 1 billion agent computers started in June 2026 and passed 2 billion in October and will hit 3 billion in the next 30 days. "The future of agentic infrastructure is increasingly more automated," said Vasek Mlejnsky, co-founder and CEO of E2B. "We see agents managing more complex systems that go beyond just running a single computer in a cloud. Eventually, agents will be managing full distributed systems. We're building a cloud of clouds. A future where each agent is managing and building its own private cloud that is able to adapt on the fly to the task the agent is working on at any given moment." E2B gives every agent a full operating system: * A sealed-off computer. Each task runs on its own isolated machine that shares nothing with any other, not even memory, so a problem in one task never reaches other users or the customer's own systems. * Robust networking and firewall: Each computer has its own dedicated networking setup, firewall, and filtering allowing for tight control of the traffic going into and outside the agent's computer. * Instant copies. Companies set up an agent's workspace once, and E2B copies it in milliseconds, as many times as they need. * The full set of tools. A web browser, a password vault for logging in on a user's behalf, and separate identities for different jobs. * Enterprise security. E2B is SOC 2 Type II compliant and HIPAA-ready, so companies can plug it into their existing systems Agents have moved from answering questions to doing real work. Gartner predicts 40% of enterprise apps will include task-specific AI agents by the end of 2026, up from less than 5% in 2025. Each of those agents needs a safe place to run. E2B started in an apartment in Prague. Mlejnsky and co-founder and CTO Tomas Valenta have known each other since middle school, about 18 years, and built more than 150 projects together. Before raising outside funding, they entered hackathons and used the prize money to pay for rent, insurance, and food. In 2023, OpenAI President Greg Brockman came across Mlejnsky's E2B demo and shared it on X with no introduction. The post reached 524,000, and the founders moved to San Francisco soon after. "Vasek and I have spent nearly all of our time working on our own things, and we never really stopped," said Valenta. "We've always believed AI agents will need a computer to do work. Nobody would hire someone without giving them one and the software they need. An agent needs the same, only it can use a thousand of them at once." E2B has raised $38 million to date, including a $21 million Series A led by Insight Partners in July 2025. Other investors include Decibel, Kaya VC and Sunflower Capital. Developers can start building at e2b.dev. E2B provides open-source cloud sandboxes for AI agents, equipped with all tools necessary for them to complete complex tasks like generating analyst-grade reports or building websites from a single prompt. E2B is trusted by enterprises and hyper-growth startups like Meta, Hugging Face, and Groq to securely deploy, scale, and manage complex AI agentic workflows. Media Contact VSC for E2B [email protected]

Cloudflare Sandbox pricing, compared: what drives the bill. Cloudflare Sandbox pricing explained with a worked cost model, compared against E2B and Daytona - and why idle time and egress decide the bill. CreateOS · August 5, 2026 How much does a Cloudflare Sandbox cost? Cloudflare Sandboxes run on the Containers billing model: a $5/month Workers Paid plan, then CPU at $0.000020 per vCPU-second, memory at $0.0000025 per GiB-second, and disk at $0.00000007 per GB-second. The plan includes 375 vCPU-minutes, 25 GiB-hours of memory and 200 GB-hours of disk before overage. Those rates are published in the Cloudflare Workers pricing docs and were checked on 5 August 2026. Sandbox pricing changes; re-check before you budget on it. The per-unit rate is the easy part. What decides your bill is which units you are charged for while the agent is doing nothing. Why per-second rates do not tell you the cost. An AI agent spends most of its wall-clock life waiting - on a model response, a tool call, a human approval. A code-execution benchmark runs flat out. An agent does not. That gap matters because vendors bill idle time differently: * Cloudflare bills CPU on active usage only - a change made in its November 2025 CPU pricing update - but bills memory on provisioned resources for as long as the sandbox exists. * E2B bills vCPU and RAM per second of sandbox lifetime, whether or not the CPU is busy. * Daytona publishes a Windows rate of $0.0858/vCPU/h and $200 in free compute, but does not publish a general per-second rate on its pricing page. So a comparison table of rates can rank three vendors in an order that reverses the moment your agent starts waiting. What does a realistic agent workload actually cost? Take 1,000 agent sessions a month, five minutes each, on 2 vCPU and 4 GiB of RAM, with the CPU genuinely busy about 20% of the time. That is 83.3 hours of wall-clock sandbox time and 300,000 sandbox-seconds. Working it through on published rates: Cloudflare | Component | Quantity | Included | Billable | Cost | | CPU (active only, ~20%) | 120,000 vCPU-s | 22,500 | 97,500 | $1.95 | | Memory (provisioned, full session) | 1,200,000 GiB-s | 90,000 | 1,110,000 | $2.78 | | Workers Paid plan | - | - | - | $5.00 | | Total | / | / | / | ≈ $9.73 | E2B, billed on sandbox lifetime at $0.000028/s for 2 vCPU and $0.0000045/GiB/s: | Component | Quantity | Cost | | vCPU | 300,000 s | $8.40 | | RAM | 1,200,000 GiB-s | $5.40 | | Usage total | / | ≈ $13.80 | E2B's Hobby tier is free with a one-time $100 credit, so that usage is covered for a while. Its Pro tier adds $150/month on top. Both figures exclude egress and disk. That omission is the point of the next section. Which costs do sandbox pricing pages hide? Egress and idle memory are where an agent bill grows faster than the compute line. Neither appears in a per-second rate comparison. Three things to check before you commit: * Egress. Cloudflare charges $0.025-$0.05 per GB beyond an included 500 GB-1 TB, varying by region. An agent that pulls large documents or writes artifacts to external storage moves real volume. * Idle memory. If memory is billed on provisioned resources, a sandbox left alive between turns bills the whole time - which is exactly the pattern a conversational agent produces. * Monthly floors. A $150/month plan is irrelevant at scale and dominant at pilot scale. Model your actual volume, not the vendor's headline tier. Cloudflare's own Sandbox SDK docs are the primary source for what a sandbox does; the pricing page is the primary source for what it costs. Read both. How does CreateOS price sandboxes? CreateOS publishes a pricing model rather than a rate card: per-second billing of vCPU and memory, no egress fees, no minimums and no monthly commitment. The AI Gateway is a flat 2.5% platform fee on the lowest available rate across nine providers. CreateOS is not going to print a per-second number here that CreateOS would then have to keep in sync. CreateOS pricing models usage with you directly, because the honest answer for an agent workload depends on your idle ratio and your egress volume, and a headline rate hides both. What is worth comparing is the model, not the number: | / | Idle CPU billed | Egress charged | Monthly floor | | Cloudflare | No (active only) | Yes, beyond allowance | $5 | | E2B | Yes (lifetime) | Not published as separate | $0 Hobby / $150 Pro | | CreateOS | Per-second vCPU + memory | No egress fees | None published | If your agents are chatty and move data, the egress line is the one to model first. Should you pick a sandbox on price? Usually not - pick on isolation model and state handling, then check the price is not disqualifying. The cost difference between these providers on a pilot workload is single-digit dollars a month. The cost of choosing an isolation boundary your security team later rejects is a re-platform. Start with what an AI agent sandbox is if you are still scoping, and with Firecracker sandboxes if the question is what the boundary actually guarantees. Common questions. Cloudflare Sandboxes bill on the Containers model: a $5/month Workers Paid plan, then $0.000020 per vCPU-second of active CPU, $0.0000025 per GiB-second of memory, and $0.00000007 per GB-second of disk. The plan includes 375 vCPU-minutes, 25 GiB-hours and 200 GB-hours before overage. Figures checked 5 August 2026. Is Cloudflare Sandbox cheaper than E2B? On a modelled workload of 1,000 five-minute sessions at 2 vCPU and 4 GiB, Cloudflare came to roughly $9.73 and E2B to roughly $13.80 in usage, before egress. The gap narrows or reverses depending on your idle ratio, because Cloudflare bills CPU only when active while E2B bills for the sandbox lifetime. Does Cloudflare charge for sandbox egress? Yes. Network egress is $0.025-$0.05 per GB depending on region, with an included allowance of 500 GB to 1 TB per month. For agents that move documents or write artifacts externally, this line can exceed the compute line. What is Daytona sandbox pricing? Daytona's pricing page publishes $200 in free compute, the first 5 GiB of storage free, and a Windows rate of $0.0858 per vCPU-hour. It does not publish a general per-second Linux rate, so a like-for-like comparison needs figures from their sales team. Why is my AI agent sandbox bill higher than the rate card suggests? Almost always idle time and egress. Agents wait on model responses and tool calls, so wall-clock sandbox lifetime is several times actual CPU work - and if memory is billed on provisioned resources, you pay for all of it. Check whether your provider bills idle CPU, and model egress separately. CreateOS bills per second of vCPU and memory with no egress fees, no minimums and no monthly commitment, and charges a flat 2.5% platform fee on the AI Gateway's lowest available provider rate. Exact rates are modelled against your usage rather than published as a rate card. About CreateOS. CreateOS is the unified AI execution layer for the enterprise - route, govern, validate, observe. It runs agent workloads in Firecracker/KVM micro-VM sandboxes with fork, pause-resume, VPC and S3-backed disks, with kernel-level egress control via eBPF, and can run entirely inside a customer's own boundary. Built by the team at NodeOps. Next step. Model your own numbers before you choose. Take your session count, average session length, vCPU and memory, and an honest idle ratio, then run them through the arithmetic above against each provider's published rates.

StackAI partners with E2B to give every agent a computer. Jun 18, 2026 StackAI and E2B have partnered to bring secure, isolated code execution to enterprise AI agents in regulated industries. The partnership now powers three core capabilities on the StackAI platform. Custom Code lets users execute Python or JavaScript inside agent workflows, with every run automatically sandboxed. Computer Use and RPA replays recorded desktop sessions inside E2B sandboxes so agents can operate enterprise software securely and reliably. Browser Use keeps web navigation and scraping sessions fully isolated. "With E2B, it's really easy to provision sandboxes. They're secure by default, highly capable, and very performant," says JJ Miller, Senior Product Manager. "Sandboxes are a key primitive for agents going forward. For us to keep growing as a company, we need a partner that can grow with us. We're happy to have found one in E2B." The collaboration scaled quickly, going from a few sandboxes during evaluation to hundreds of thousands in production. And it's set to deepen, according to JJ: "Sandboxes are a key primitive for agents going forward. For us to keep growing as a company, we need a partner that can grow with us - we're happy to have found one in E2B." Read the full story here. Pratik paudel. Engineering at StackAI Make your organization smarter with AI. Deploy custom AI Assistants, Chatbots, and Workflow Automations to make your company 10x more efficient.