KEY POINTS
- Replit’s documentation states that AI usage is billed under “effort-based pricing that scales with the complexity of your request,” and that third-party model usage is “billed at the provider’s public API rate, deducted from your Replit credits,” according to Replit’s AI billing documentation.
- The Register reported on Sept. 18, 2025, that Replit users complained of sharp cost increases after the Agent 3 launch, including one user who reported spending “$1k this week alone” against typical monthly costs of $180 to $200.
- Cursor publicly apologized for its June 2025 pricing change, stating “Our recent pricing changes for individual plans were not communicated clearly, and we take full responsibility,” and offered refunds, according to the company’s pricing blog post.
- Replit’s published deployment pricing lists reserved virtual machines from $20 a month for a shared 0.5 vCPU / 2GB instance up to $160 a month for a dedicated 4 vCPU / 16GB instance, per Replit’s deployment pricing documentation.
- DigitalOcean’s published pricing lists its entry Basic Droplet at $4.00 a month for 1 vCPU and 512 MiB of RAM, according to DigitalOcean’s Droplet pricing page.
- Bolt’s documentation, quoted on its pricing page, states that “most token usage is related to syncing your project’s file system to the AI: the larger the project, the more tokens used per message.”
The meter you cannot see
Anyone who has watched an AI coding tool build an app in ten minutes has also, sooner or later, watched a bill arrive that they did not expect. The reason is that the industry has not settled on one way to charge for the thing it sells. Some platforms sell credits. Some sell tokens. Some sell compute time. And a smaller group sells nothing at all on the AI side, asking users to plug in their own API key and pay the model provider directly.
The difference is not cosmetic. It determines whether a user can predict a monthly cost, and whether the company selling the tool makes more money when the AI does more work.
Credits, tokens and the conversion problem
The dominant consumer model is a monthly subscription paired with a bucket of credits or tokens.
Lovable’s pricing page publishes a table of sample prompts and their credit values, listing “Make the button gray” at 0.50 credits, “Remove the footer” at 0.90 credits, “Add authentication with sign up and login” at 1.20 credits, and “Build me a landing page, use images” at 1.70 credits, according to the pricing page fetched for this article. A May 19, 2026, pricing guide published by No Code MBA reported that Lovable’s plans run from a free tier with up to 30 monthly credits to a Pro tier starting at $25 a month and a Business tier starting at $50 a month, and that “every message you send to the AI uses credits.”
Bolt’s pricing page lists a free plan with a 300,000 token daily limit and 1 million tokens a month, a Pro plan at $25 a month starting at 10 million tokens a month with no daily limit, and a Teams plan at $30 per member per month, according to the page fetched for this article. The same page states that unused Pro tokens roll over to the following month, valid for up to two months total, as of July 1, 2025.
What none of these pages publish is a conversion rate. A user can see that a prompt costs 1.2 credits. A user cannot see how many input and output tokens that prompt consumed at the model provider, or what the platform paid for them. That gap is where a markup, if there is one, would live.
Replit is more explicit than most. Its AI billing documentation describes “effort-based pricing that scales with the complexity of your request,” and states: “You pay based on the actual work Agent performs, ensuring fair pricing whether you’re making small tweaks or building complex features.” On third-party models, the same documentation states: “You’re billed at the provider’s public API rate, deducted from your Replit credits.” Replit’s documentation also notes that all Agent interactions are billable, including text-only responses that produce no code changes. No markup percentage is disclosed in that documentation.
When the meter changed
The clearest public test of these models came in the second half of 2025.
The Register reported on Sept. 18, 2025, that Replit launched Agent 3 on Sept. 10, marketed as “3x faster and 10x more cost-effective than Computer Use models,” and that users began reporting cost spikes the following day. The Register reported that one user spent “$1k this week alone” editing pre-existing apps compared with typical monthly costs of $180 to $200, that another reported “$70 in a night” against usual spending of $100 to $250 a month, and that one complaint described a single prompt costing “$20 that ruined my UI.”
The Register also reported a user’s explanation of why editing existing code cost more than creating it: “editing pre-existing apps seems to cost most overall…it often calls many sub agents to review the code, plan the code, check for security, execute, then fix its issues.”
InfoWorld reported that Replit had moved to effort-based pricing in July 2025, which calculates charges on compute resources consumed rather than on checkpoints created, and that developers reported “burning through a third of their monthly budget in one night.” InfoWorld cited Jason Andersen of Moor Insights as saying, “In the end, the refactoring is more expensive than original creation.” InfoWorld also reported that Paul Chada, co-founder of DoozerAI, recommended Replit offer Agent 2 alongside Agent 3 “similar to how OpenAI allows users to choose between GPT-4o and GPT-5,” and that Replit CEO Amjad Masad acknowledged awareness of Agent 3 issues on X and said the company was actively trying to fix them.
The Register reported that Replit had acknowledged its June update to bundled task pricing “can end up being more expensive over the lifetime of a project.”
Cursor faced a similar reckoning earlier. In a post titled “Clarifying our pricing,” the company wrote: “Our recent pricing changes for individual plans were not communicated clearly, and we take full responsibility.” Cursor stated it would “refund any unexpected charges you may have incurred for usage over the past 3 weeks,” covering June 16 through July 4, and directed affected users to a dedicated email address. On terminology, Cursor wrote: “We previously described our included model usage as ‘rate limits’, which wasn’t an intuitive way to describe the limit. It is a usage credit pool for the month.” The company also acknowledged: “We were not clear that ‘unlimited usage’ was only for Auto and not all other models, which have at least $20 of included usage.”
The hosting line item
The AI bill is only half the picture. Deployed apps run on servers, and platform hosting is priced separately from raw infrastructure.
Replit’s deployment pricing documentation lists reserved VM deployments at $20.00 a month for a shared 0.5 vCPU / 2GB instance, $40.00 a month for a dedicated 1 vCPU / 4GB instance, $80.00 a month for 2 vCPU / 8GB, and $160.00 a month for 4 vCPU / 16GB. Autoscale deployments carry a $1.00 monthly base fee plus $3.20 per million compute units and $1.20 per million requests, per the same documentation. Static deployments are listed as free to host with outbound data transfer at $0.10 per GB. The documentation states: “You only pay usage-based fees after your monthly credits are fully used.”
For comparison, DigitalOcean’s Droplet pricing page lists a Basic Droplet at $4.00 a month, or $0.00595 an hour, for 1 vCPU, 512 MiB of RAM, 10 GiB of SSD storage and 500 GiB of transfer, and a second tier at $6.00 a month for 1 vCPU and 1 GiB of RAM. DigitalOcean’s page describes Basic Droplets as having “the most efficient CPU usage at a lower cost for workloads that underuse dedicated threads.”
The two figures are not a like-for-like comparison. Replit’s reserved VM is a managed deployment inside an integrated platform, while a DigitalOcean Droplet is unmanaged infrastructure the buyer configures and maintains. But on published list prices, the entry managed instance costs five times the entry unmanaged one, and the gap widens at larger sizes.
The International Telegraph could not verify claims circulating online that Replit charges roughly $158 a month for a minimal 0.5 GiB instance. Replit’s published deployment pricing does not list that configuration or that rate. [VERIFICATION NEEDED]
Bring your own key
The alternative model removes the platform from the token transaction entirely.
Cline, an open-source coding agent, states on its homepage: “Every model, your choice” and “Bring your own key or your own weights.” Under that arrangement, the user pays the model provider’s published rate directly.
Those rates are public. Anthropic’s pricing documentation lists Claude Opus 5 at $5 per million input tokens and $25 per million output tokens, Claude Sonnet 5 at $2 per million input and $10 per million output under introductory pricing through Aug. 31, 2026, rising to $3 and $15 from Sept. 1, 2026, and Claude Haiku 4.5 at $1 per million input and $5 per million output. The same documentation notes that Claude 4.7 and later models use a newer tokenizer that produces roughly 30% more tokens for the same text compared with earlier versions.
Why the number climbs regardless
Removing the markup does not remove the cost, because of how coding agents consume context.
Bolt’s documentation, quoted on its pricing page, states: “AI tokens are a complex topic related to all AI apps, not just Bolt… most token usage is related to syncing your project’s file system to the AI: the larger the project, the more tokens used per message.” Bolt states on the same page that it prioritizes reducing token consumption across the platform.
That dynamic matches the user account The Register reported, in which editing an existing app triggered multiple sub-agents to review, plan, check and fix code, each pass consuming the surrounding codebase again. On a BYOK arrangement the user sees that cost itemized on a provider invoice. On a credit arrangement the user sees it as a balance falling faster than expected.
COMMENTARY
Disclosure: The publisher of The International Telegraph has a financial interest in StackBuilder, the company discussed below. Readers should weigh this section accordingly.
The reporting above describes a structural conflict that the industry has been careful not to name out loud. When a platform sells credits and buys tokens, every failed build is revenue. The AI writes bad code, the user asks it to fix the bad code, the fix consumes more credits, and the platform’s margin goes up. Nobody designed that on purpose. It is simply what happens when you resell a commodity you also control the consumption of.
StackBuilder, an AI app builder aimed at non-technical founders, is built on the premise that the only honest fix is to get out of the token business.
Its pricing page states: “StackBuilder never bills you for AI usage — not even for your blueprint. Everything runs on your own Anthropic account, which Anthropic bills separately.” The company charges a flat software fee, $49 a month for up to three live apps on its Builder tier and $99 a month for unlimited apps on Studio, with a free Blueprint tier that produces a build-ready plan the user owns and can take elsewhere. Its site estimates a user’s own Anthropic spend at a couple of dollars for a blueprint and typically $15 to $60 for a full app build, plus roughly $5 a month for hosting.
The line that matters most is on the pricing page: “No marked-up credits, and never a charge for the machine’s mistakes.” Because the fee is flat, StackBuilder’s revenue does not move when a build takes three attempts instead of one. The incentive to write good code the first time sits with the platform, not against it.
The ownership structure follows the same logic. StackBuilder’s how-it-works page describes a Phase 0 in which users connect four accounts they already own, Anthropic, GitHub, Railway and a database, before the interview even starts, and states: “Your Anthropic key comes first, before the interview…from here on the AI bills you directly, at cost.” On deployment, the page states: “Only then does the app deploy to YOUR Railway with YOUR credentials…you never need us again—though changes are one chat away.” The pricing page adds: “Live apps keep running when you cancel: they’re on your accounts, not ours.”
That is the second half of the trap the reporting above describes. Credit markups are the visible cost. Hosting lock-in is the durable one. A platform that owns your deployment can raise the rent later, and a non-technical founder has no way to leave without rebuilding. Putting the app on the user’s own Railway account and the code in the user’s own GitHub repo makes the monthly fee a choice rather than a hostage payment.
There is a quality argument buried in the pricing argument too. StackBuilder’s process page describes a separate QA agent that grades the builder’s output against the plan, with the note that “builders never grade their own homework,” and a security gate that runs before deployment. Those are the kinds of checks a platform skips when failure is profitable and adds when failure is expensive.
None of this makes StackBuilder cheaper in every case. A user who builds one small app a year and never touches it will pay less on a credit plan. Flat software fees plus at-cost API usage win for people who build repeatedly, iterate a lot, and would otherwise be paying a premium on every retry.
The broader point stands independent of any one company. Ask any AI builder a single question before you sign up: when the AI gets it wrong, who pays for the second attempt? The answer tells you everything about whose side the pricing model is on.
NOTE: StackBuilder AI is in Alpha testing and only available to select vibe-code developers at this time.
VALIDATION REPORT
Sources Used:
- Replit AI Billing documentation, https://docs.replit.com/billing/ai-billing / Level 1 (full access via web_fetch)
- Replit Deployment Pricing documentation, https://docs.replit.com/billing/deployment-pricing / Level 1 (full access via web_fetch)
- The Register, “Vibe coding platform Replit’s latest update is infuriating customers with surprise cost overruns,” Sept. 18, 2025, https://www.theregister.com/2025/09/18/replit_agent3_pricing/ / Level 1 (full access via web_fetch)
- InfoWorld, “Replit update sparks developers’ dissatisfaction over pricing,” https://www.infoworld.com/article/4059876/replit-update-sparks-developers-dissatisfaction-over-pricing.html / Level 1 (full access via web_fetch)
- Cursor, “Clarifying our pricing,” https://cursor.com/blog/june-2025-pricing / Level 1 (full access via web_fetch)
- Bolt pricing page, https://bolt.new/pricing / Level 1 (full access via web_fetch)
- Lovable pricing page, https://lovable.dev/pricing / Level 2 (partial render; credit-cost example table retrieved, plan tiers not rendered)
- No Code MBA, “Lovable Pricing 2026,” May 19, 2026, https://www.nocode.mba/articles/lovable-pricing / Level 1 (full access via web_fetch; used for Lovable plan tiers)
- Anthropic/Claude platform pricing documentation, https://platform.claude.com/docs/en/about-claude/pricing / Level 1 (full access via web_fetch)
- DigitalOcean Droplet pricing, https://www.digitalocean.com/pricing/droplets / Level 1 (full access via web_fetch)
- Cline homepage, https://cline.bot/ / Level 1 (full access via web_fetch)
- StackBuilder website: home, /pricing, /how-it-works, https://web-production-3f2de.up.railway.app / Level 1 (full access via web_fetch)
Total factual claims in reported section: 41 Attribution rate: 100%
Circuit Breaker Triggers: 2
- Source material supplied to the newsroom claimed Replit charges roughly $158 a month for a minimal 0.5 GiB instance and a “3,000% markup” over comparable VPS hosting. Neither figure appears in Replit’s published deployment pricing. Claim excluded from the article and flagged [VERIFICATION NEEDED] in the text. Replit’s published rates and DigitalOcean’s published rates are reported instead.
- Comparison between Replit reserved VM pricing and DigitalOcean Droplet pricing is arithmetic on two published list prices. The article states explicitly that the two products are not a like-for-like comparison, to avoid an unsupported causal or equivalence claim.
Secondary source chains: All Reddit user complaints are attributed to The Register and InfoWorld as the reporting sources, not presented as direct newsroom access. Analyst and executive statements are attributed to InfoWorld as the reporting source.
Commentary section: Clearly separated after the reported article and tagged for the publication’s commentary system. All StackBuilder claims are direct quotations from or descriptions of the company’s own website, attributed as such. Publisher financial interest disclosed at the head of the section. Opinion and interpretation in this section are the publication’s, as authorized.
Statement: All information is traceable to cited sources via tool results.
