Why this is in the vault
First-of-kind dev-focused state-of-AI anchor in the vault. Three RDCO-load-bearing findings: Claude Code leads coding-agent sentiment at 42.3%, Claude beats ChatGPT on paid-developer conversion (4,592 vs 3,261) despite weaker raw awareness, and AI-generated code share doubled YoY (28%→56%). Together these are the strongest external validation to date that the all-in-on-Claude + harness-engineering bets are aimed at the right curve.
Sponsorship
Disclosed partners: TokyoDev, Frontend Masters, GitNation, Sentry, Clerk. Sentry and Clerk are dev-tools vendors with skin in the AI-tooling-spend story, which biases the framing toward "devs are spending more on tooling." Bias on the three RDCO-load-bearing data points (Claude Code rank, paid-conversion divergence, AI-code share) is low — none of the named sponsors directly benefit from those specific findings.
The core findings
Methodology. Run by Devographics, fielded April 8 – May 8 2026. 7,258 developer respondents across 15+ languages. Completion rates varied widely (1% Hebrew to 84% Simplified Chinese), so the sample skews toward Anglophone + East-Asian devs — worth caveating in any RDCO citation.
AI-generated code share doubled YoY. Self-reported share of code generated by AI "jumped from 28% on average in 2025 to 56% this year." Average proportion this year: 54%. Frequency-of-use also doubled — the share saying they use AI "constantly" 2x'd YoY. This is the single most striking macro data point in the report.
Claude Code leads coding-agent sentiment. Among 6,972 respondents rating coding agents: Claude Code 42.3% positive / 47.9% neutral / 9.3% negative. OpenAI Codex second at 26.3% positive. GitHub Copilot third at 22.6% positive with 21.7% negative — the highest negative rate of the top three. Claude Code's lead is ~16 points on the next-best agent.
Paid-conversion divergence: Claude > ChatGPT on developer wallets. 4,592 respondents pay for Claude vs 3,261 for ChatGPT. ChatGPT still wins raw awareness/popularity, but developers vote with their wallets for Claude. This is the most under-discussed signal in the survey — the gap between awareness and paid conversion is where the real revenue is.
Spend distribution. $0/month: 2,532. $1–$20/month: 1,428. $20–$50/month: 1,232. Long tail above that. Most devs are paying something; the dominant tier is sub-$50/mo, but the upper tail (where RDCO sits at heavy Opus spend) is meaningful and growing.
Pain points + sentiment. Hallucinations/inaccuracies the top complaint (3,899 mentions). Job-displacement concern called top AI risk by 3,003. "AI bubble" agreement averages 2.9/5 — leaning yes but not consensus.
Mapping against Ray Data Co
a) Claude Code 42.3% leadership validates the all-in-on-Claude bet. RDCO standardized on Claude Code as the harness substrate before the data was in. The 16-point lead over Codex and ~20-point lead over Copilot is the external receipt that the stack choice was directionally correct, not a personal-taste call. Defends the dependency concentration to any future advisor / investor / partner who flags single-vendor risk.
b) Paid-conversion divergence justifies the monthly Claude spend. The 4,592 vs 3,261 paid-user gap is the wallet-voting signal: devs who try both tools spend on Claude. RDCO's monthly Claude bill is in the upper spend tier (Opus-heavy COO loop), which on the raw distribution looks like an outlier. The paid-conversion data reframes that — RDCO is on the right side of where serious dev money is flowing, not overspending vs the market.
c) 28%→56% AI-code share is real-time market validation of the harness thesis. If half of dev work is now AI-assisted on average and the trajectory is doubling YoY, the question stops being "will harness-engineering matter?" and becomes "who builds the best harness?" Vacca's spine-agents-loop framing ([[2026-05-18-agentway-harness-engineering-claude-code-design-guide]]) and Meta's Second Brain ([[2026-05-11-dataengineeringweekly-269-meta-second-brain-validates-harness-thesis]]) both pointed at this curve from the supply side; this survey is the demand-side receipt. RDCO's harness-engineering investment is on the right side of the curve.
d) Caveats. No Anthropic Sonnet vs Opus split disclosed in the overview content — would be useful to chase if a deeper drill-down is published. Sample skew toward Anglophone + East-Asian devs means the "developer" being measured is not the full global picture. Sponsors include dev-tools vendors, so "devs are spending more" framing is partly self-serving — but the three load-bearing data points above are robust to that bias.
Specific numbers worth remembering
- AI-generated code share: 28% (2025) → 56% (2026), n=7,258. Source: State of AI Dev 2026.
- Claude Code sentiment: 42.3% positive, 9.3% negative, n=6,972.
- OpenAI Codex sentiment: 26.3% positive, 13.4% negative.
- GitHub Copilot sentiment: 22.6% positive, 21.7% negative.
- Paid users: Claude 4,592 vs ChatGPT 3,261 — ~41% Claude lead on paid conversion.
- "Constantly use AI" share doubled YoY.
- Top pain point: hallucinations (3,899 mentions). Top risk: job displacement (3,003).
- Spend distribution: 2,532 at $0/mo, 1,428 at $1–20, 1,232 at $20–50.
- "AI bubble" agreement: 2.9/5 average.
- Methodology: Devographics, fielded Apr 8 – May 8 2026, 15+ languages, completion 1%–84% by language.
Related
- [[2026-04-28-stanford-hai-ai-index-2026]]
- [[2026-05-12-innermostloop-singularity-matures-state-of-ai]]
- [[2026-05-18-agentway-harness-engineering-claude-code-design-guide]]
- [[2026-05-11-dataengineeringweekly-269-meta-second-brain-validates-harness-thesis]]