For the past few years, AI code assistants operated on a simple model: flat monthly fee, unlimited use.
That model is dead.
In spring 2026, every major AI development tool shifted to usage-based pricing. The reasons are obvious in hindsight—someone running an AI agent 24/7 wasn’t supposed to cost the same as someone using it interactively—but the impacts on development budgets are significant.
Here’s what changed, and what it means for your organization.
The GitHub Copilot Shift
GitHub Copilot moved from a flat $10 per month per user model to usage-based billing effective June 1, 2026. Developers now pay for “GitHub AI Credits” based on token consumption—think of it like API billing where every completion, every suggestion, every agentic action consumes tokens from your account. The trigger was practical: a handful of organizations were running Copilot in agentic mode, using code generation in a loop for autonomous coding tasks, and were already costing 5–10x the monthly subscription price in a single week. GitHub realized the flat-fee model was economically unsustainable.
Interactive coding assistance—a developer typing a function and asking Copilot to complete it—will likely remain cheap at maybe $10–30 per month. But agentic use cases like running code generation on a large codebase, continuous refactoring, and automated test generation will be billed separately and will be significantly more expensive. For legacy modernization projects where you’re using Copilot heavily (converting 50+ SYNON programs, refactoring a large codebase), your AI tool costs will rise substantially. Budget for 2–3x the old flat-fee cost if you’re doing aggressive automation.
The Anthropic (Claude) Restructuring
Anthropic’s Claude Enterprise shifted from a per-seat flat fee to a hybrid model: $20 per month base per seat plus actual API usage on top. A user who runs one chat per day might pay $20 per month, whereas a user running 100 API calls per day might pay $50–100+ per month. The complication is that Claude Code briefly disappeared from the $20 Pro plan in late April 2026 before being reversed within 24 hours—a sign that pricing restructuring is still in flux and will likely continue. Analysts estimate that heavy Claude users could see costs triple compared to the old flat-rate model.
If you’re using Claude for code review, architecture analysis, or continuous development support, you need to monitor your usage carefully. The economics have fundamentally changed, and what was a fixed cost is now variable.
IBM’s Different Approach
IBM watsonx Code Assistant, the AI tool specifically for RPG, COBOL, and mainframe, adopted a tiered, enterprise-negotiated model rather than usage-based pricing. Pro tier costs $20 per seat per month (individual developers), Pro+ costs $60 per seat per month (team collaboration with role-based access), and Enterprise is $200 per seat per month (centralized dashboard, team management, audit logs). IBM didn’t go usage-based—instead, IBM negotiated task-based or volume-based enterprise contracts with predictable pricing. The tool is purpose-built for IBM i modernization.
For IBM i shops using watsonx for SYNON conversion or RPG modernization, you get predictable costs without worrying about usage spikes. The trade-off is that the per-seat cost is higher than other tools, but the tool is fine-tuned for exactly your use case.
Why This Matters
The shift from flat-fee to usage-based pricing has several critical implications. Agentic use cases get expensive—if you’re using AI in autonomous mode running code generation in a loop, doing continuous refactoring, or performing automated testing, your costs rise significantly. What looked cheap at $20 per month per developer now looks expensive when you multiply it by thousands of API calls per day. This matters particularly for IBM i modernization because converting large legacy applications often uses agentic approaches: “convert all SYNON programs in this portfolio,” “refactor all RPG to Free Form,” “generate tests for all procedures.” Those operations, if they were cheap under flat-fee pricing, now have real costs you need to account for.
Development teams now need visibility into usage. GitHub and Claude both provide usage dashboards, but you need to actually monitor them. Budget owners need to track these costs the way they monitor cloud infrastructure. The tool ecosystem is stratifying—GitHub and Claude are going after general-purpose developers, while IBM is going after enterprises with specific needs. That stratification creates opportunity: if you’re doing IBM i modernization, watsonx might be cheaper and more purpose-built than generic tools. If you’re doing general development, you need to think carefully about usage patterns to predict costs accurately.
The Cost Equation for Modernization
Let’s work through a concrete example: converting 50 SYNON programs to RPG Free Form.
Old model (flat-fee per developer):
- 1 developer + 1 AI tool: $20/month × 6 months = $120 in tool costs
- Developer labor: 10 weeks × $50/hour = $25,000
- Total: ~$25,120
New model (usage-based):
- Same developer labor: $25,000
- GitHub Copilot (heavy agentic use): $1,000–$2,000 (estimated based on 2026 reports)
- Claude Code review and optimization: $500–$1,000
- Total: ~$26,500–$27,000
The AI costs tripled, but the total project cost only increased ~5–7%. The real win is still developer productivity (6 weeks instead of 10).
Alternative (using IBM watsonx):
- Same developer labor: $25,000
- IBM watsonx Pro+ for 6 months: $60 × 6 months = $360
- Total: ~$25,360
If you’re doing IBM i modernization, using the IBM-specific tool is more cost-effective and the tool is actually built for your use case.
The Honest Assessment
The shift to usage-based pricing is economically rational—unlimited consumption at a flat price was always unsustainable—but it means you can’t ignore AI tool costs anymore.
For organizations modernizing legacy systems (especially IBM i), the question has changed from “should we use AI?” to “which AI tool makes economic sense for our use case, and what will we actually spend?”
Usage-based tools (GitHub, Claude) make sense for exploratory work or periodic assistance. Tiered enterprise pricing (IBM watsonx) makes sense if you’re doing heavy, repetitive modernization work.
The good news: even with the cost shift, AI-assisted development is still dramatically cheaper and faster than manual work.
If you’re planning a major modernization initiative and want to understand how to budget for AI tooling, let’s talk. We can help you model the costs and pick the right tools for your specific use case.