AI coding agents have transformed developer productivity. They have also transformed IT budget lines. Claude Code, OpenAI’s Codex, and GitHub Copilot now run continuously, in parallel, across entire codebases. Every action consumes tokens. For a ten-person engineering team, the annual bill can exceed $100,000. Many companies found out only when the monthly invoice arrived.
Key Takeaways
- Claude Code costs between $150 and $250 per developer per month in intensive use, with peaks reaching $500–800 depending on the models used
- OpenAI’s Codex switched to token-based billing on April 2, 2026, replacing the previous per-message system and significantly increasing costs for heavy users
- The industry expects stabilization between $100 and $200 per month for professional developers by end of 2026, up from around $20 in 2025
Why Tokens Accumulate So Fast
Understanding the billing mechanics is the first challenge. AI coding tools do not charge per request or per message. They charge per token: fragments of text roughly three to four characters long in English. A simple sentence represents about a dozen tokens. Harmless in isolation.
The trap is agentic behavior. Claude Code does not answer a question: it reads the codebase context, executes tool calls, and runs multi-step verification loops. Every file modification, every search, every terminal command consumes tokens. A developer using Claude Code six hours a day with Opus 4.7 without any optimization can easily spend $600–900 per month.
One European SaaS startup illustrates what this looks like at scale. The company had integrated an AI assistant into its product. Its monthly bill went from €800 to over €12,000 in three months, simply because the user base had grown and no one had capped conversation history length.
Unlike traditional cloud infrastructure, where cost alerting is mature and well configured, token spend monitoring remains rudimentary in most organizations. Many companies discover budget overruns after the monthly invoice, with no ability to react in real time.
A Sector-Wide Pricing Shift Since April 2026
April 2026 marks a turning point for the entire sector. On April 2, OpenAI updated Codex pricing to align with the API’s token-based billing model, replacing the previous per-message system. The change hit Business and Enterprise customers first, then rolled out to Plus and Pro subscribers in the following weeks.
For the GPT-5.3-Codex model, the reference price sits at $1.75 per million input tokens and $14 per million output tokens, with a context window reaching up to 400,000 tokens. OpenAI estimates Codex costs between $100 and $200 per developer per month on average, with significant variance depending on the model used.
GitHub Copilot is following a similar trajectory. Microsoft is considering a shift to token-based billing for future offerings, after removing access to Opus models from its $10 Pro plan, an implicit acknowledgment that the previous pricing was not sustainable with agentic workflows.
Optimizing token consumption has become a skill in its own right for engineering teams. What was once a marginal consideration is now a concrete budget issue for IT leadership.
Also on Horizon:
- Sam Altman on the Stand: The Question of Trust
- Ramp AI Index: Anthropic Edges Out OpenAI in Enterprise
- Claude Code and AI Agents Are Blowing Up Dev Budgets
Open Source Alternatives Waiting in the Wings
Faced with rising costs, part of the ecosystem is looking for alternatives. Three open source tools dominate the scene in 2026. Cline, with 58,200 GitHub stars and 5 million VS Code installs. Aider, with 39,000 stars and 15 billion tokens processed per week. Continue.dev, aimed at enterprise needs.
These tools let teams bring their own API key and take full control of costs, with no provider markup. A developer switching to this setup can choose exactly which model to use, set context limits, and make real-time trade-offs between performance and cost.
On the native optimization side, Anthropic documents that the average cost in enterprise deployment is around $13 per developer per active day. 90% of users stay below $30 per active day. This means extreme cases are real, but far from representative. Anthropic’s native prompt caching can significantly reduce costs on repetitive contexts.
The industry expects stabilization between $30 and $50 per month for average users and between $100 and $200 for heavy users by end of 2026. Agentic workflows are not going back. Open source players should gain market share as IT leaders look to regain control over spending.
For IT leadership, the question is no longer whether AI tools are productive. It is how to measure and control what they actually cost, before the next invoice arrives.
Follow the story on Horizon.
