The Best Deal in AI Isn't for the People Who Build¶
Source: The best deal in AI isn't for the people who build · LinkedIn Pulse / Substack
Author: Andrew Kulakov
Date Published: June 2026
Tagline: Subscriptions are priced for the person who won't use them.
TL;DR¶
Premium AI subscriptions (ChatGPT Pro, Claude Max) are priced 40–70× below API-equivalent cost — a deliberate predatory pricing strategy to capture developer mindshare and build tooling dependency before pricing rationalises. At full utilisation, these plans operate at −900% to −1,650% gross margins. Enterprise case studies (Uber blew its entire 2026 AI budget in 4 months; Microsoft cancelled most Claude Code licences) show the subsidy is already breaking. Meanwhile, RCT evidence shows developers feel 20% faster even when they're measurably slower — a dangerous perception gap that compounds the risk.
The Great AI Subsidy¶
Premium subscriptions are strategically priced far below API-equivalent cost. This is a classic predatory pricing play — capture mindshare, build dependency, then rationalise.
"Subscriptions that don't make financial sense now make strategic sense later."
The Raw Economics¶
SemiAnalysis (June 2026) — converting token consumption to API-equivalent retail pricing at max usage:
| Plan | Price | API-Equivalent Value at Max Use |
|---|---|---|
| ChatGPT Pro 20× | $200/mo | ~$14,000 |
| Claude Max 20× | $200/mo | ~$8,000 |
| ChatGPT Plus | $20/mo | ~$700 |
| Claude Pro | $20/mo | ~$400 |
Margin Reality: - At 75% API gross margins, labs lose money above 11% utilisation. - At full utilisation: −1,650% margin (ChatGPT Pro) and −900% margin (Claude Max).
"The deal is real. It's designed for the person who won't use it."
The author's trajectory: started Feb 2026 at $16.67/mo (Claude Pro annual), hit limits by April ($30–$45 extra usage), upgraded to Max 5× then Max 20× on June 11. $325 in 4 months. Current: $200/mo + API usage.
The Strategy & Trajectory¶
- The Uber Playbook: Subsidise access with outside capital, build dependency, extract value once alternatives are gone.
- The Economics: OpenAI projects a $14 billion operating loss in 2026 ($1.69 burned per $1 earned).
- The Pivot is Visible: June 22, 2026 — Anthropic moved Fable 5 out of flat-rate plans into usage credits.
- The Endgame: SemiAnalysis projects Opus-class models could "eventually" be delivered profitably around $20/mo.
- The Trap: Real switching costs (CLAUDE.md, MCP setups, agent harnesses) accumulate while pricing is artificially low.
"A utility doesn't sell kWh at 40× below cost on a flat subscription. Not because they're mean. Because the physics doesn't allow it."
Enterprise Case Studies: Where the Subsidy Breaks¶
Uber (Q1 2026)¶
- Scale: Autonomous agents wrote >10% of committed production code. Claude Code adoption: 32% → 84% in 3 months (5,000 engineers).
- Spend: Burned entire 2026 AI budget in 4 months. Q1 R&D: $951 million (+17% YoY). Per-engineer cost: $500–$2,000.
- Reaction: Capped spend at $1,500/tool/employee.
- Measurement Gap:
COO Andrew Macdonald: "That link is not there yet... very hard to draw a line between one of those stats and 'Okay, now we're actually producing 25% more useful consumer features.'"
Microsoft¶
- Ran Claude Code internally for ~6 months, then canceled most licences.
- Shifted engineers to GitHub Copilot CLI.
- Mustafa Suleiman: "Anthropic is extremely expensive."
"Same conclusion, different spreadsheets: the bill arrives before the ROI case is clear."
The Dangerous Perception Gap¶
You feel faster even when you aren't. This is the critical psychological blind spot.
RCT Evidence¶
| Study | Prediction | Actual | Felt |
|---|---|---|---|
| METR (arXiv 2509.19708) | 24% faster | 19% slower | 20% faster |
| Anthropic RCT | — | No stat sig gain | — |
- Anthropic developers scored 17% lower on comprehension of code shipped with AI assistance.
- Result: "The feeling survived the evidence."
Token Economics (arXiv 2604.22750)¶
- Agentic tasks consume roughly 1,000× more tokens than simple code chat.
- Same task can cost 30× more on different runs.
- "You can't predict what a task costs before running it."
The Honest Calculator¶
"Run that against your headcount before your next planning cycle."
Uber's $1,500/tool/month cap: - An engineer on Claude Code + Cursor = $3,000/month - $36,000/year per person - ≈ 11% of a US senior engineer's total comp
The Subsidy Dependency Test:
"What would you pay today if your subscription disappeared and you moved entirely to API pricing? If the number is more than 3× your current bill, you're dependent on a subsidy."
Key Takeaways¶
- Premium AI subscriptions are not bargains for heavy users — they are loss-leaders designed to build dependency. The economics don't work at scale and the pivot to rational pricing has already begun.
- Enterprise adoption is outpacing ROI measurement. Uber blew its annual AI budget in 4 months; Microsoft cancelled most Claude Code licences. Both concluded the bill arrives before the business case is clear.
- The perception of productivity gain is unreliable. RCT evidence shows developers feel faster while being empirically slower or producing lower-quality code. Trust usage data, not intuition.
- Cap spend before you scale — don't copy the $200/mo subscription model without monitoring actual token consumption. Design for prompt caching (41–90% savings) and route tasks to appropriate model tiers from day one.