Published 2026-07-23 · reference numbers, verify before budgeting
Every AI cost conversation starts with tokens. We build calculators for a living and the token calculators are the ones people use — the price comparison is the busiest page on this site by a wide margin. So this week we did the opposite exercise: we wrote guides for four pieces of agent infrastructure — Cerebras for inference, E2B for code sandboxes, Browserbase for headless browsers and LiveKit for realtime transport — and then priced one realistic task across all four at once.
The task: a ten-minute voice call where an agent talks to a user, browses four minutes of web pages, and runs five short scripts. The result was not close. The LLM tokens came to 0.89% of the bill. Proxy bandwidth came to three quarters of it.
The thing that makes agent infrastructure hard to budget is that none of these services bill on the unit you think about the task in. You think in "calls" or "tasks"; they bill in sandbox-seconds, browser-hours, gigabytes and participant-minutes.
| Service | Meter | Rate | The trap |
|---|---|---|---|
| Cerebras | tokens in / out | $0.35 / $0.75 per 1M | none — this is the honest, well-understood one |
| E2B | sandbox lifetime, per second | ≈$0.10 / hour | idle time bills the same as running time |
| Browserbase | browser hours + proxy GB | ≈$0.12 / hr · ≈$12 / GB | the GB meter is 100× the hour meter |
| LiveKit | participant min + agent min | ≈$0.0005 · ≈$0.01 | the agent meter is 20× the transport meter |
Reference rates, July 2026. Report outdated price →
Look at the trap column. Three of the four services have a secondary meter that costs an order of magnitude or more than the one on the headline of their pricing page. That is not deception — every one of these numbers is published openly. It is just that the mental model you bring ("a browser costs twelve cents an hour, fine") attaches to the wrong number.
Here is the ten-minute call, costed twice: once by a team that closes its sandboxes and blocks images in the browser, and once by a team that does neither. Same feature, same user, same models.
| Line item | Disciplined | Sloppy | What changed |
|---|---|---|---|
| LiveKit — 10 agent min + 20 participant min | $0.1100 | $0.1100 | nothing; 91% of it is the agent meter |
| Cerebras — 8k in / 2k out tokens | $0.0043 | $0.0043 | nothing |
| E2B — 5 sandboxes | $0.0028 | $0.5040 | closed after 20s vs left to a 1-hour timeout |
| Browserbase — 4 browser minutes | $0.0080 | $0.0080 | nothing |
| Browserbase — proxy bandwidth | $0.3600 | $2.4000 | 30 MB with images blocked vs 200 MB without |
| Total per call | $0.485 | $3.026 | 6.2× |
| Per 1,000 calls | $485 | $3,026 | the same product, twice |
Two things stand out. The first is that the gap between the columns is not a pricing decision — nobody switched provider, nobody negotiated a tier. It is two configuration habits: calling kill() on a sandbox instead of letting it time out, and blocking images and fonts in a headless browser that is only there to read text. Between them they account for a 6.2× difference in unit economics.
The second is that the token line, the one every cost article on the internet is about, is $0.0043. Less than one percent of the disciplined bill, and less than two tenths of a percent of the sloppy one. You could halve your token spend — switch models, compress prompts, cache aggressively — and move the total by 0.4%. You could block images in one line of Playwright config and move it by 42%.
E2B bills per second of sandbox lifetime, not per second of execution. A script that runs for twenty seconds inside a sandbox costs about $0.00056. The identical script inside a sandbox left to run out its one-hour session cap costs $0.10 — 180 times more for the same computation, with the extra 3,580 seconds spent doing absolutely nothing. In an agent loop that spawns a sandbox per step, this is not a rounding error; it is the difference between agent infrastructure being cheap and being the second-largest line on the invoice.
The fix is one line of teardown code, and the reason teams miss it is that it works fine in development. A prototype with three test runs a day never notices; the meter only becomes visible at production volume, by which point the habit is baked into the code path. If you are building an agent loop, model it with the agent loop budget calculator before you scale it, not after.
Browserbase charges roughly $0.12 for a browser hour and roughly $12 for a gigabyte of proxy bandwidth. That ratio — 100:1 — means a single browser hour needs to move only about ten megabytes before bandwidth and compute cost the same, and modern web pages move ten megabytes without trying. Our four-minute browsing session at 200 MB cost $2.40 in bandwidth against $0.008 in browser time: bandwidth was 300× the compute.
Blocking images, fonts, video and stylesheets in a scraping browser typically cuts payload by 80-90%, which is where our 30 MB disciplined figure comes from. If the agent is reading text — and for most agent tasks it is — none of that payload was ever going to be used. This is the same lesson as the hidden costs that double your AI and SMS bills: the shape of the usage decides the invoice far more than the sticker rate does.
LiveKit's participant-minute rate is the number in every comparison article: about $0.0005 a minute, which sounds like nothing because it is nothing. The agent-session meter is about $0.01 a minute — twenty times higher — and it is the one that actually runs during an agent call. In our ten-minute example the participant minutes came to a cent and the agent minutes to ten. An estimate built from the headline rate lands 10× low.
None of which includes the speech models: transcription, the LLM turn and text-to-speech are all billed by whoever provides them, on top of everything above. Price that whole stack with the AI voice agent cost calculator.
Agent infrastructure is not expensive. Agent infrastructure configured with development-time habits at production-time volume is expensive, and the two are separated by about six times on our numbers. The uncomfortable part is that all the money is in meters nobody writes blog posts about — sandbox seconds, proxy gigabytes, agent-session minutes — while the meter everybody writes about turned out to be the smallest line on the invoice.
Read the guides: Cerebras · E2B · Browserbase · LiveKit · all AI & agent APIs.
Methodology: rates are the published reference figures as of July 2026, cross-checked against each provider's public pricing page while writing the four guides linked above. The worked example is a constructed scenario, not telemetry from a production system — the point is the ratio between meters, which holds across a wide range of assumptions, not the absolute total. Confirm current pricing before budgeting.