# The model isn't what makes one AI coding agent 10x cheaper than another > Source: > Published: 2026-07-29 16:49:21+00:00 *The cheaper AI coding agent isn't always running the cheaper model. The real spread is in the harness: the tool loop, context handling, retries, and verification sitting around it.* Ask an engineering team why it chose Claude Code over Cursor, or Cursor over GitHub Copilot, and you'll usually hear one word first: model. That's the wrong answer. The bill says otherwise. The frontier models still matter, of course. You don't get good autonomous coding from a weak model and a nice wrapper. But once you're comparing serious tools on serious models, the expensive part is often the machinery around the model. The agent reads files, calls tools, retries failed edits, runs tests, compresses history, reopens context, and asks the model to decide what to do next. Each step costs tokens. Each bad step costs more. That's why harness architecture has become the thing buyers need to inspect, not the thing they can leave to engineers after procurement signs the contract. Anthropic said in a March 2026 engineering post that harness design was key to pushing Claude further on long-running software work, especially where verification and tool use mattered. A separate July 2026 analysis called The Harness Tax measured the wrapper directly and found that running the same problems through Claude Code multiplied context by 40 to 125 times and per-run cost by 5 to 8 times, while also improving one model's pass@1 score by 28 points. That isn't a footnote. That's the product. ## The bill is hiding in the loop A coding agent doesn't just answer your prompt once. It carries a system prompt, tool schemas, file contents, terminal output, previous decisions, failed attempts, and sometimes a long trail of reasoning from earlier turns. The Harness Tax put the fixed context load for its single-shot harness at roughly 36,000 to 71,000 tokens per run, compared with 400 to 1,000 input tokens for a raw model call on the same small problems. You can argue with the sample size. You can't argue with the direction. Output tokens make the pain sharper. Published pricing tables from OpenAI, Anthropic, and Google show a common pattern: generated tokens cost several times more than input tokens. Anthropic's March 2026 Claude announcement listed Opus 4.6 at $5 per million input tokens and $25 per million output tokens, while Sonnet 4.6 was $3 and $15. A chatty agent that explains every edit before making it isn't being helpful to your finance team. It's spending money. This is where the difference between products starts to show. Builder.io's comparison, published in September 2025 and updated in February 2026, reported that Claude Code completed one benchmark task with about 33,000 tokens while Cursor used about 188,000 tokens on the same work and hit errors along the way. Treat that properly: it's one test, not a universal law. Still, it gives you a concrete number for what many engineers already feel in daily use. Some agents keep a tighter grip on context. Others reload and re-summarize their way through the same job. Claude Code's advantage is tied partly to the size of the context window it can use. Anthropic's official Claude documentation says Claude Code supports a 1 million-token context window on paid plans for models including Sonnet 5, Fable 5, Opus 4.8, Opus 4.7, and Opus 4.6, with plan-specific conditions for usage credits. That's a long leash. That lets a long refactor stay in one session instead of being chopped into brittle pieces. It also creates a trap: a million tokens is useful only if the harness knows what to keep, what to drop, and when to stop. ## Procurement needs better questions For enterprise buyers, sticker prices are a weak guide. Cursor's own pricing page lists Teams at $40 per user per month, with Enterprise custom. GitHub's Copilot documentation currently lists Copilot Business at $19 per user per month and Copilot Enterprise at $39. Those numbers matter, especially when you're rolling seats out to hundreds of developers. They still don't tell you the cost per completed task. This is the buying question: how many successful changes do you get for the money? If your team mostly wants inline completions and quick IDE feedback, Cursor and Copilot can make perfect sense. If you're asking an agent to refactor across a monorepo, run tests, inspect failures, and come back with a working patch, Claude Code's wider autonomous loop may be worth the higher apparent spend. Many teams will use both. That's not indecision. It's matching the tool to the job. Investors should be just as blunt. A startup claiming a breakthrough AI coding agent needs to show more than model access and a clean interface. Ask for token traces. Ask how the harness handles stale context. Ask what happens after a failed test, and how many retries happen before the agent burns through the budget. Ask whether tool choice is constrained or whether the model gets a junk drawer and a prayer. Frankly, the thin-wrapper story is over. The model contract can be copied, repriced, or replaced when the next frontier release arrives. The harder work is in evals, observability, context compaction, tool discipline, and feedback loops from real usage. That's where the cost curve bends. The market still talks as if the model is the moat. For AI coding agents, the harness is where the bill lands first. **Also read:** [Together AI hits $8.3 billion valuation as enterprises flee closed AI models for open source](https://startupfortune.com/together-ai-hits-83-billion-valuation-as-enterprises-flee-closed-ai-models-for-open-source/) • [Eliyan raises $145 million at a $1 billion valuation as AI's real bottleneck moves between chips](https://startupfortune.com/eliyan-raises-145-million-at-a-1-billion-valuation-as-ais-real-bottleneck-moves-between-chips/) • [NextEra and Brookfield are turning a Cold War uranium site into a $100 billion AI data center](https://startupfortune.com/nextera-and-brookfield-are-turning-a-cold-war-uranium-site-into-a-100-billion-ai-data-center/)