The deploy platforms grew up selling one moment: a developer pushes to git and a URL comes back. Between August and October 2026 every one of them filed here rewrote that pitch around a different reader, the agent. Until now this thread read their free tiers as one movement, wider every month. That was half the picture. The free door is being re-cut rather than widened: more room for an agent’s many disposable projects, less for the human app serving traffic, and a new cost in the keys those projects hold.

The free door, re-cut

Render made the trade in one announcement. On 23 April its $0 Hobby plan, “our free plan for personal projects”, gained unlimited projects and lost bandwidth: the allowance was “reduced” to 5 GB a month, with overage at $0.15 a GB. Render’s editor’s note says the new plans reached every existing workspace from 1 August. Render does not state the old figure. A plan sized for one app serving visitors became a plan sized for many small projects that serve almost nobody - and Render’s pricing page now lists Sandboxes in beta, to “give agents an isolated computer”.

Fly.io prices the two doors on its own pages. Its pricing page reads “Every new user gets a $1 free trial, no credit card required”; its home page, which now opens on “deploy apps and agents”, says “New Sprites organizations get $30 in trial credit”. Sprites is Fly’s sandbox for agents. When the $1 wording arrived we could not date, so this is the gap as it stands on 3 November, not a measured cut: thirty times more trial for the agent’s computer than for the app.

The widening this thread filed is real and still stands. Vercel gave away the trust layer: the Sandbox egress firewall on every plan including free (August 11), and full deployment protection for production URLs on every plan from September 9, retiring the $150/month add-on Pro needed for that scope. Railway gave away the front door: deploy with no account and 60 minutes to claim (August), then a free 2 vCPU / 2 GB VM from ssh railway.new with 60 minutes to build and 24 hours to claim (September 25). Neon doubled free storage to 1 GB a project across 100 projects on October 2, for agents’ “lots of projects, lots of experiments, most of them idle most of the time”. DigitalOcean starts its Agent Droplets at $5 of credit with no card. Every one of those widens the agent’s side, and the money moves to the plan above: DigitalOcean’s $50 and $200 tiers at 15% and 20% off usage, Vercel’s Token Pass at $30 a month for $80 of credit on open-weight models.

What the free door now costs: the keys

Three independent hosts disclosed that customer environment variables were exposed inside twelve days, and the variables are where an agent’s AI provider keys live.

  • Railway, 20 to 22 October: at most 1.85% of workspaces, about 100,000, had some variables potentially exposed. Railway says the attacker was “attempting to obtain AI provider API keys to perform distillation attacks” - its own reading, with the entry point still undescribed.
  • Northflank, 30 October: users were told to “assume all environment variables have been decrypted”, and that night Northflank wrote that “users have reported unauthorised use of AI API keys supplied to the Northflank platform”. By 2 November it said it believed the attack vector was contained within an hour of the first alert and that the evidence suggested the target was secrets in its control plane, and that it intends to share a root cause analysis once findings are confirmed.
  • Trigger.dev, 1 November: an attacker reached environment variables for “a subset of organizations”; no count and no cause given.

None of the three has published a root cause, so these are three vendor disclosures and not yet one proven campaign. What followed at Railway is the first front door to close since the thread opened. Railway’s forum told a would-be user on 25 October that “new account signups are currently paused as part of an ongoing incident”; its report’s 26 October update says it had “paused new signups as we prioritize uptime” for its “5 million existing users” and would “reopen signups this week”. We found no confirmation by 3 November that it had. That is not a price, and Railway gives uptime rather than abuse as the reason - but it is the first time a host filed here shut the door it had spent the autumn widening.

The unit was AWS’s first

AWS has billed AgentCore Runtime, its serverless microVMs for agents, “per second, using actual CPU and memory consumption, with a 1-second minimum”, with “CPU scales to zero during I/O wait” - the waits on a model response or a tool call. AgentCore went generally available on 13 October 2025; the AWS pages we read do not date when that billing rule began. It shipped a second generation on 18 September 2026 that reclaims idle memory mid-session and, by AWS’s own test, cuts the P75 cold start from 5.4-30 seconds to 1.9-2.0. New AWS accounts get up to $200 of Free Tier credit; there is no standing free allowance.

So when Railway, Fly.io and DigitalOcean each shipped compute built for agents between August and October - the 2026-09-26 signal counted three clouds in five weeks - they were converging on a shape AWS had already priced. DigitalOcean’s Managed Agents bill per second of active CPU, “the waits in between cost $0”, at $0.044 per vCPU-hour. AgentCore lists $0.0895 per vCPU-hour on the first generation and $0.1276 on the second; Fly’s Sprites list $0.0385 per CPU-hour. What the independents sell on the unit is not a new idea; it is a price at half AgentCore’s first-generation rate or below.

That gap has a caveat on AWS’s side. Its launch post says the new Runtime comes “with lower costs”, while its pricing page lists higher per-unit rates than v1 for both CPU and memory; both are true only if the memory reclaim saves more than the rate rise costs, and the page’s one worked example prices v2 alone. AWS also lists a committed-baseline option at $0.0997 per vCPU-hour, “launching by October 2026” - wording still on the page on 3 November. Cloudflare’s answer was never a rate: route file work to isolates and keep containers under 10% of agent workloads, which is Cloudflare’s stated target, not a measured result.

Two exits, both to platform builders

On 17 February 2026 Mistral AI bought Koyeb, an independent serverless platform with 13 employees, in its first acquisition. Koyeb said the platform would keep running and become a core component of Mistral Compute; it would focus on enterprise clients, and new users could no longer sign up for its Starter tier (TechCrunch’s report; Koyeb’s own announcement now sits behind a login). Koyeb’s pricing page still opens at Pro, $29 a month plus compute.

On 9 October the Deno team announced it was joining Cloudflare. Cloudflare’s post says the two will merge celld, Deno’s self-hostable implementation of the Workers and Durable Objects model, with workerd; Deno’s says the runtime gets a year of monthly bug and security releases and that Deno Deploy “will continue operating for six months before shutting down”, with migration support for paying customers moving to Workers. Neither post discloses terms, a customer count or a calendar date for the shutdown.

Both buyers are companies assembling an agent platform of their own - a model maker building a cloud, an edge network building an agent runtime - and neither is a hyperscaler. The one exit that had a low-priced entry closed it to newcomers; Koyeb’s Starter was pay-as-you-go with a free instance that its FAQ says still comes with every plan, so what rose was the entry price, not provably the free allowance. We found no third exit by 3 November.

The oldest PaaS left the race a different way. On 6 February Heroku moved to “a sustaining engineering model … rather than introducing new features” and stopped offering enterprise contracts to new customers, while promising card-paying customers no change. Not an exit; a host stepping out of the repositioning the others were about to start.

Where the customer is won now

With the unit a hyperscaler’s and the rate the only edge on it, the competition went to three other places.

The language. Cloudflare took Python Workers to GA on September 22; Vercel shipped Python GA at Ship 2026 (its recap is dated October 15). Two platforms whose customers wrote JavaScript both reached, inside a month, for the language the AI app is written in.

The code’s home. Netlify built git into the platform (Netlify Source, August 14) and added Cursor Origin as a deploy source beside GitHub, GitLab and Bitbucket (September 11). Cloudflare put Artifacts, a versioned filesystem for agent-driven development, into open beta with $25,000 in credit for the best “next Git platform” built on it (October 2).

The procurement path. Vercel listed AI Gateway on AWS Marketplace (August 5), billed through the buyer’s AWS account, and shipped Bring Your Own Cloud on AWS at Ship. Vercel’s Sandbox also runs Cognition’s Devin sessions (August 4), and by Ship it sold persistent storage for agents (Drives) and Daemons beside the deploy. The host that cannot win the enterprise’s cloud commitment sells into it.

The tension

The case for the independents is real and dated: Daytona, which sells a computer per agent, raised $60 million at a $1 billion valuation, reported by TechCrunch on 23 October (other outlets date the announcement the 20th; the terms are the company’s own). And Render’s bandwidth cut was announced in April, before the agent repositioning began, so reading it as a resize for agents could be fitting an ordinary margin move to this story - except that the same announcement is the one that made projects unlimited.

The evidence points one way from four directions. The unit the independents sell was priced first by AWS, so their edge on it is a rate. The two exits went to companies building agent platforms. The free door is being re-cut toward the agent, not simply opened. And holding an agent’s keys has become the thing attackers come for, which is a cost a free tier pays first. The thread’s bet: the independent hosts hold their customers only where the agent’s door stays cheaper to open than to copy, and to secure - and the loops above say what would prove it wrong.

What would settle it

3 open loops

  1. 01 By Deno Deploy's shutdown, around April 2027

    A third exit, or the first from the hosts filed here. Koyeb (to Mistral AI, 2026-02-17) and Deno (to Cloudflare, 2026-10-09) both went to companies assembling an agent platform. Does Railway, Fly.io, Render, Netlify, Northflank or another independent host fold into a larger rail, or announce a sunset, before Deno Deploy shuts? One that goes to a hyperscaler would make this plain consolidation rather than platform builders buying parts.

  2. 02 By end of Q1 2027

    Does anyone cut the agent's door? The cuts so far fell on the human app (Render's bandwidth) while the agent's side widened. A host that shortens a no-account claim clock, asks for a card before a sandbox or free VM, or reopens signups behind a new gate after a breach would be the first public price on what an agent's disposable projects cost to give away. Railway, which paused signups on or before October 25 and promised to reopen within the week, is the nearest test.

  3. 03 By end of 2026

    Does AWS price the gap away? AgentCore's v2 list rates ($0.1276 per vCPU-hour) sit well above DigitalOcean's ($0.044), and the committed-baseline rate AWS said would arrive "by October 2026" was still unlaunched on November 3. A baseline or a v2 rate under the PaaS list prices removes the one thing the independents sell on the unit; a worked AWS example showing v2 cheaper than v1 in practice would show how wide the gap really is.

Each is checkable by a date. If one goes quiet, that is itself an answer.