The cost
Blind is not cheap. It is billed somewhere else.
The industry spent two years measuring what AI gives back. The measurements that matter now are about what it costs when it does not know what your team already knows. Every figure below is sourced, including the ones published by companies with something to sell.
Uber spent its entire 2026 AI budget in four months of a twelve month year, after putting Claude Code in front of roughly 5,000 engineers.
Praveen Neppalli Naga, CTO, Uber · May 2026The Information, via Forbes
What Salesforce expects to spend on Anthropic this year. Asked what for, Benioff answered in one word: “Coding.”
Marc Benioff, CEO, Salesforce · All-In podcast · May 2026His projection, not a filed figureOne two hour coding session, run by Uber’s own CTO. Engineers there average $150 to $250 a month. Power users reach $2,000.
Praveen Neppalli Naga, CTO, Uber · May 2026The Information, via Forbes
“It’s very hard to draw a line between one of those stats and ‘okay, now we’re actually producing like 25% more useful consumer features.’ That link is not there yet.”Andrew Macdonald · President and COO, UberRapid Response podcast, via Fortune · May 2026
“GitHub has absorbed much of the escalating inference cost behind that usage, but the current premium request model is no longer sustainable.”Mario Rodriguez · Chief Product Officer, GitHubCopilot moved to usage-based billing, 1 June 2026
The buyer cannot justify it. The seller cannot absorb it.
“AI adoption’s positive benefits depend on organizations having AI-accessible internal data.”DORA · Google Cloud · State of AI-assisted Software Development 2025 · ~5,000 respondents
dora.dev/dora-report-2025
The single biggest frustration developers report with AI tools: answers that are almost right, but not quite. Debugging AI-generated code ranks second.
Stack Overflow Developer Survey 2025 · 49,000+ respondentssurvey.stackoverflow.co/2025/ai
Lost per developer per week to organizational friction, for half of all developers surveyed. The largest single cause is finding information. At 500 developers that is $7.9M a year.
Atlassian State of Developer Experience 2025 · 3,500 developersatlassian.com · developer-experience-report-2025Atlassian sells knowledge tools
Merge rate across 878 agent-authored pull requests in dotnet/runtime, against 87.1% for human ones. The agent PRs that did merge needed 16.5 review comments each, versus 12.4.
Microsoft · .NET Blog · ten months of Copilot coding agentdevblogs.microsoft.com/dotnetMicrosoft’s own agent
Tokens per prompt, pulling in as many as 50 related files, by the end of Airbnb’s test migration. That context is what moved it from 75% automated to 97%.
Airbnb Engineering · accelerating large-scale test migrationmedium.com/airbnb-engineering
Per developer, per month, at the top of Gartner’s observed range, up from $20. They expect AI coding costs to pass developer salaries by 2028, and find no direct relation between token spend and productivity.
Gartner, via The Register · June 2026theregister.com · ai-coding-agents-cost
Estimated drop in delivery stability for every 25% increase in AI adoption. Throughput barely moved. More code shipped, more of it broken.
DORA · Accelerate State of DevOps 2024 · ~39,000 respondentsdora.dev/research/2024
Every one of these is the same bill in a different envelope. The agent does not know what your team already knows, so somebody pays to tell it again, every session. An engineer pays it in review comments, or you pay it in tokens. Uber’s answer was to feed its agents source code, engineering docs, Slack and Jira, and call it memory.
This is the problem. The measured answer to it, against self-hosted competitors on the same model and the same budget, is on the benchmarks page.
See the benchmarks →