State of AI coding spend: benchmarks from 800 developers and $2.3M of usage

nikshepsvn1 pts0 comments

State of AI Coding Spend 2026: Benchmarks From 800 Developers and $2.3M of Usage | Viberank Blog<br>Back to BlogState of AI Coding Spend 2026<br>June 10, 2026·9 min read<br>Everyone has an opinion about what AI coding agents cost. We have receipts. Developers on the Viberank leaderboard measure their own usage locally with ccusage — real Claude Code, Codex, and Gemini CLI logs, not survey answers — and submit it publicly. This report covers everything on the board as of June 2026.<br>792<br>developers

2.5T<br>tokens

$2.3M<br>usage value

29,230<br>coding days tracked

The median serious user has burned $1,285<br>Lifetime usage value per developer, at API-equivalent prices:<br>p25

$379

median

$1,285

p75

$3,128

p90

$6,494

p99

$30,720

The top user has consumed $56,694 of compute value across 81 billion tokens. The mean ($2,904) sits at more than twice the median, because AI coding spend is a textbook power law:<br>The top 1% of users account for 14% of all spend<br>The top 10% account for 51%<br>The top 25% account for 74%<br>Daily burn: $29 is the new normal, $200+ days are routine<br>Across all 29,000 tracked coding days:<br>median day

$29

p90 day

$215

p99 day

$708

11% of all tracked days exceeded $200 — an entire Claude Max monthly subscription's worth of compute, consumed in one day. The single biggest day we've recorded: $3,820 . And the median submitter is active 26 days per submission window — this is daily-driver usage, not weekend tinkering.<br>Power users have outgrown their subscription price 5–10×<br>Normalizing each developer's usage to a monthly rate: roughly half run at $1,000+/month in API-equivalent value, and ~73% run above $400/month. Most of them pay $100–200/month flat for Claude Max. The arbitrage is stark — the median heavy user extracts 5–10× the sticker price of their plan in raw compute. Flat-rate pricing made always-on agentic coding economically rational, and the usage curves show developers responded exactly as you'd expect.<br>AI coding is 95% rereading, 0.2% writing<br>The token mix across the entire dataset is the most lopsided stat in this report:<br>cache reads

94.8%

cache writes

4.2%

input

0.8%

output

0.2%

For every token an agent actually writes, it re-reads ~406 tokens of cached context. The 2.5 trillion token headline is really ~5.9B tokens of generated work product riding on a mountain of context re-reads. This is why prompt caching, not model price, is the real economic engine of agentic coding — and why long sessions in big repos get expensive even when output is small.<br>The always-on cohort exists<br>Seven developers have logged 200+ active days , and the longest consecutive-day streak on the board is 238 days straight — eight months without missing a single day. The top of the leaderboard isn't people who type fast; it's people who've turned agents into infrastructure that runs whether they're at the keyboard or not.<br>Weekends barely slow anyone down: Saturdays and Sundays are 24% of active days (a uniform week would be 28.6%), and a weekend day burns the same ~$82–84 as a Tuesday. When your agent does the typing, "logging off" is a softer concept.<br>Model mix: Opus is the workhorse, not the treat<br>Sonnet

96%

Opus

90%

Haiku

64%

Nine in ten developers run Opus — the most expensive tier is now the default for serious work, not a special occasion. Haiku's 64% tells the quieter story: cheap models doing subagent and utility passes inside bigger workflows.<br>Multi-agent developers are 9% of the board — and most of the top 10<br>A note on coverage: viberank started as a Claude Code leaderboard and only recently opened to every agent ccusage tracks — Codex, Gemini CLI, Copilot, OpenCode and friends. So this dataset is still overwhelmingly Claude Code, and the cross-tool numbers should be read as early signal, not market share.<br>That said: about 9% of the board already reports usage from more than one coding agent, and they're heavily over-represented at the very top — several top-10 spenders run three or more agents in parallel. If you want a preview of the median 2027 workflow, look at the current multi-tool tail — see our Codex vs Claude Code vs Gemini CLI comparison.<br>If you use Codex, Gemini CLI, or anything else alongside (or instead of) Claude Code, your submission makes the next edition of this report meaningfully better — one command covers every tool ccusage detects:<br>npx viberank-cliUsage is at an all-time high<br>Usage value logged per calendar month across all submissions:<br>$109k

Jun 25

$347k

Jul

$323k

Aug

$112k

Sep

$44k

Oct

$54k

Nov

$121k

Dec

$165k

Jan 26

$152k

Feb

$226k

Mar

$302k

Apr

$221k

May

$77k

Jun*

*June 2026 is 10 days in — tracking toward ~$230k, on pace with the launch-spike months. Earlier months only include usage from developers whose submissions cover them.

After the launch spike (Jul–Aug 2025) and an autumn lull, tracked usage has climbed for six straight months. Spring 2026 is running at launch-hype levels — except now it's steady-state behavior, not novelty.<br>Who these...

usage coding developers days claude median

Related Articles