The Real Cost of AI Coding Tools (And Why the Price Will Change)
I pay $200 a month for Claude Max. I run multiple agent sessions in parallel, most days. My bill for AI coding tools, if I add Cursor on top, clears $250. And honestly, I've never felt better about a software purchase.
That said, I think most engineers paying for these tools have no idea what they're actually buying. The pricing on AI coding tools right now doesn't reflect real costs. It reflects a race for market share, backed by billions of dollars in venture capital. And that's going to change.
Here's what the economics actually look like, the pricing table, the subsidy gap, the productivity data (which is messier than you think), and what I think you should do about it.
What AI Coding Tools Actually Cost Right Now
As of early 2026, this is the pricing landscape:
| Tool | Free | Standard | Premium |
|---|---|---|---|
| GitHub Copilot | $0 (limited) | $10/mo | $39/mo (Pro+) |
| Claude Code | - | $20/mo (Pro) | $100-200/mo (Max) |
| Cursor | $0 (limited) | $20/mo | $200/mo (Ultra) |
| OpenAI Codex | Included in ChatGPT | $20/mo (Plus) | $200/mo (Pro) |
On paper, these prices look trivially small. The average US software engineer costs a company $150,000–200,000 a year in salary, more with benefits. A $200/month Max subscription is $2,400 a year. That's roughly 1–2% of total compensation. If the tool gives you even a 10% productivity lift, the ROI math writes itself.
That math is real. But it's also incomplete.
The Subsidy Problem: Why These Prices Don't Add Up
The $200/month Max plan gives heavy users access to compute that would cost $3,000–4,000 at standard API rates. The gap between what you pay and what your usage actually costs is roughly 18x for power users.
That gap is a subsidy. Anthropic, OpenAI, and the others are eating that difference to acquire users and build market share. They can do it because they're sitting on massive funding rounds. But it's not a business model.
We've been here before. Uber rides were cheap in 2015 because VC money covered the gap between what riders paid and what rides actually cost. DoorDash was the same story. Both eventually had to charge prices that reflected real economics. Fares and delivery fees went up, sometimes a lot.
AI coding tools are in that phase right now. The signs are already showing. Cursor switched from flat-rate to usage-based pricing in June 2025. Some users saw their bills go from $20/month to $350 a week. GitHub added metered billing for overages at $0.04 per request. Microsoft raised Microsoft 365 prices partly to offset Copilot infrastructure costs.
Sam Altman said publicly that the $200/month ChatGPT Pro subscription is "currently unprofitable because users hammer it so hard." OpenAI generated $4.3 billion in revenue in the first half of 2025 and posted an operating loss of $7.8 billion.
That's not a business. That's growth-at-any-cost funding.
The Counterargument: Compute Is Getting Cheaper Fast
That said, there's a real case that prices stay manageable.
LLM inference costs dropped roughly 99% between GPT-4's launch in March 2023 and late 2025. Anthropic cut Opus pricing by 67% when they released Opus 4.5. Models keep getting more efficient. Claude Code uses prompt caching heavily, with 90%+ of tokens being cache reads, which dramatically reduces the actual compute cost per session. One analysis by Martin Alderson estimated that for the average Claude Code user consuming around $6/day in API-equivalent spend, the real compute cost to Anthropic is about $18/month. That's well within the $20 subscription price for most users.
So the all-you-can-eat tier is unsustainable for power users, but the base economics might work out for average usage patterns. The direction the industry is heading is clear: usage-based pricing, tiered by actual consumption. The question is whether falling compute costs outpace the pressure to become profitable before the VC runway runs out.
I think prices will stay roughly affordable. But the flat-rate unlimited plans won't last in their current form.
The Productivity Question (And Why the Data Is Messy)
Here's where I want to slow down, because the productivity numbers get repeated a lot and they don't all point the same way.
The optimistic case is real. Anthropic reports 70% productivity increases internally. Atlassian's CTO describes 2–5x output gains for certain tasks. In the Pragmatic Engineer's 2026 survey, 95% of engineers said they use AI tools at least weekly.
But the controlled studies tell a different story.
The METR study from July 2025 is the most rigorous I've seen. Sixteen experienced open-source developers were randomly assigned to work with or without AI tools on real issues in repos they maintained. Before the study, developers predicted AI would save them 24% of their time. The actual result: AI made them 19% slower. Even after seeing the data, they still believed AI had helped. The study's authors attributed this to suboptimal prompting, context-switching overhead, and generated code that didn't meet their quality bar.
The DORA 2025 report found that while 80%+ of developers report subjective productivity gains, actual software delivery metrics, lead time, deployment frequency, change failure rate, stayed flat. A 25% increase in AI usage correlated with a 7.2% decrease in delivery stability.
In the Pragmatic Engineer survey, only 16% of respondents said AI made them "significantly more productive." 41% said it had "little to no effect."
GitClear's analysis of 153 million changed lines found that code duplication increased from 8.3% to 12.3% since widespread AI adoption, while refactoring activity dropped from 25% to under 10% of changed lines.
I don't think the skeptical studies are wrong. But I don't think they tell the full story. The METR study tested expert developers on codebases they knew deeply, the exact scenario where AI adds the least value, because you already have the context and skill to do the work yourself. AI adds the most value in unfamiliar codebases, cross-stack work, boilerplate-heavy tasks, and parallel exploration. Those scenarios are harder to design a controlled study around.
My experience matches the messier picture. AI doesn't make me faster at everything. It makes me faster at the tedious parts and frees up time for the work that needs real attention. Whether that shows up in a study as "productivity" depends entirely on what you measure.
What Companies Are Actually Doing
Companies aren't waiting for the research to settle. They're placing bets now.
Shopify's CEO sent a company-wide memo in April 2025 making AI usage a baseline job expectation. Teams requesting headcount increases must first show why AI tools can't fill the need. Shopify tracks token consumption on an internal leaderboard. Salesforce reduced support staff from around 9,000 to roughly 5,000, explicitly crediting AI. Amazon laid off 16,000 employees in 2026.
That said, the pattern isn't always "same output, fewer people." Fast Company reported that several companies replacing entry-level engineers with AI discovered those tasks didn't disappear, they got pushed onto senior staff, who were supposed to be doing higher-value work instead. The honest picture is usually the same headcount shipping more, not smaller teams shipping the same. Whether that justifies tool costs depends on what you actually build with the extra output.
How I Think About It
I'm going to be direct: I think the current pricing is a gift and you should use it.
I pay for Max. I run agent sessions on multiple problems in parallel. The amount of work I can move through in a day is meaningfully higher than it was before. If the price doubles to $400/month, I'd still pay it without thinking twice. The math is obvious when you consider what an hour of an engineer's time costs.
At the same time, I'm not naive about what I've built on. When usage-based pricing arrives at scale, and it will, I'll need to adapt. That might mean routing simpler tasks to cheaper models. DeepSeek V3.2 is priced at $0.28 per million input tokens, which is 10–50x cheaper than Claude Sonnet, and for straightforward work it can be good enough. It might mean running open-source models locally for routine tasks and saving the expensive models for complex problems.
The practical position: use these tools hard right now, while the economics are in your favor. Build the skills. Build the workflows. But track what you're spending and stay flexible. The companies that thrive aren't the ones that locked into one tool at one price point. They're the ones that built genuine skill and adaptability.
If you want to go deeper on this, I cover it in my book How to Be a Great Software Engineer in the Age of AI, including how to build workflows that stay useful as pricing evolves.