How Token Roulette works
Token Roulette is a game, not a meter. Nothing here calls a real model. Every run is a simulation that follows fixed rules, seeded so the same settings and the same seed always produce the same trace.
The simulation
Each line of the trace is one cycle of an agent working the ticket: read code, edit, run tests, read the failures. A cycle stands for about 10 model requests. Every cycle sends context to the model and gets output back, and the cost of that cycle is those tokens multiplied by the list prices below.
- Model tier changes the price per token and how likely each cycle is to land the fix.
- Context strategy changes how much is re-sent every cycle, how much is cached, and how often the agent loses the plan.
- Max iterations stops a run that hasn't shipped. Unlimited means it keeps going until it ships or the money runs out.
- Sub-agents work in parallel: more chances to land the fix, more tokens, and sometimes they overwrite each other.
- Guardrails stop the run at $15, optionally with a nudge from you at $10.
Luck matters. Every run draws a random multiplier for how verbose and loopy the agent is, which is why the same settings can cost very different amounts. The numbers under each run come from 20,000 simulated runs of every combination of settings. Your budget is $50 per ticket. “Your month” assumes 6 medium tickets a day for 21 working days.
Prices used
List prices in US dollars per million tokens, from each vendor's official pricing page.
| Tier | Priced as | Input | Cached | Output |
|---|---|---|---|---|
| Frontier | Claude Opus 5.5 | $4 | $0.20 | $20 |
| Mid | Claude Sonnet 5.5 | $2 | $0.10 | $10 |
| Small | Claude Haiku 5.5 | $0.10 | $0.01 | $0.50 |
Claude Haiku 5.5 charges more when a single request's prompt exceeds 100,000 tokens ($0.50 input, $2.50 output); the game applies that.
Checked 2026-10-31. Pricing source
The same simulation drives the monthly estimates for Claude Code, Cursor and GitHub Copilot.
Sources
Gartner analyst Nitish Tyagi told The Register that AI coding bills are leaping from $20–100 to $2,000–5,000 per developer per month, with extreme cases reaching $20,000. The Register, 24 Jun 2026
Gartner predicts that by 2028, AI coding costs will overtake the average developer's salary. Gartner press release, 24 Jun 2026
Gartner: agentic models require 5–30x more tokens per task than a standard GenAI chatbot. Gartner press release, 25 Mar 2026
Bai, Brynjolfsson, Pentland et al. ('How Do AI Agents Spend Your Money?', arXiv:2604.22750, Apr 2026): agentic coding tasks consume ~1000x more tokens than code reasoning and code chat. arXiv:2604.22750 (v3, 2 Oct 2026)
Bai, Brynjolfsson, Pentland et al. (arXiv:2604.22750, 2026): runs of the same task can differ by up to 30x in total tokens. arXiv:2604.22750 (v3, 2 Oct 2026)
On 1 June 2026 GitHub moved all Copilot plans to usage-based billing, charging for input, output and cached tokens. GitHub Blog, 27 Apr 2026; changelog 1 Jun 2026
Search interest in 'claude code pricing' peaked at ~36k US searches/month in Jan–Apr 2026. (Ahrefs data, Oct 2026)
Anthropic's Claude Code docs: across enterprise deployments, the average cost is around $13 per developer per active day and $150–250 per developer per month, with 90% of users below $30 per active day. Claude Code docs, Manage costs
Limits
- The token counts per cycle are calibrated so that outcomes look like real agent runs; they are not measured from your repository.
- Cache-write surcharges, batch discounts, fast modes and regional pricing are left out.
- Real agents differ by tool, prompt and codebase. Treat the numbers as orders of magnitude, not a quote.
Names
Product and model names are used only to say whose list price a tier uses or which tools people run. Not affiliated with any model or tool vendor.