{"slug": "claude-code-pricing-what-100-rounds-of-ai-website-edits-cost", "title": "Claude Code Pricing: What 100 Rounds of AI Website Edits Cost", "summary": "A documented 100-round Claude Code test found that repeatedly prompting the tool to \"make it better\" on a coffee shop website cost $29.93 in API billing on Claude Sonnet 5.5 at medium reasoning effort and $137 on Claude Opus 5.5 at the same effort, a roughly 4.5x gap for identical iteration counts. Both runs started from the same deliberately ugly site in clean sessions with no CLAUDE.md, skills, plugins or memory, and Sonnet's last real milestone landed around round 32, with no further milestones across the remaining 68 rounds even as lines of code kept climbing. The test concludes that forcing iteration past a point buys more features rather than better output, since the model must invent a \"bold change\" every round whether or not one is warranted.", "body_md": "# Claude Code Pricing: What 100 Rounds of AI Website Edits Cost\n\nA 100-round Claude Code test compares Sonnet and Opus API costs on medium effort, from under $30 to over $130, with real output shown.\n\n## What does it actually cost to run Claude Code for an extended, open-ended coding task?\n\nIn one documented test, telling Claude Code to “make it better” on a website 100 times in a row cost $29.93 in API billing on Sonnet (medium effort) and $137 on Opus (medium effort). Both runs started from the same deliberately ugly coffee shop site, used the same forced-iteration prompt, and ran in a clean session with no memory, skills, or plugins carried over between rounds. The gap between the two models wasn’t just about code quality. It was almost a 4.5x difference in dollar cost for the same number of iterations.\n\n## TL;DR\n\n- A **100-round Sonnet run** on medium effort cost $29.93 in API billing and generated 11 AI images and two videos along the way.\n- The equivalent **Opus run** cost $137, nearly four and a half times more, and produced 54 AI images and three videos.\n- Costs didn’t scale with quality. Lines of code climbed steadily round after round, but the **value added per round** bounced around rather than trending up.\n- Both models hit a **diminishing returns wall** well before round 100. Sonnet’s last real milestone landed around round 32, and Opus showed clear overengineering and bugs by round 100.\n- **Forcing iteration** with a strict prompt (rather than letting the model stop when it ran out of ideas) is what drove the cost up, since the model has to invent a “bold change” every single round whether or not one is warranted.\n- The experiment suggests that past a certain point, more spend on a bigger model doesn’t buy better output. It buys more features, which isn’t the same thing.\n\n## Other agents ship a demo. Remy ships an app.\n\nReal backend. Real database. Real auth. Real plumbing. Remy has it all.\n\n## How was the pricing test set up?\n\nThe creator ran Claude Code against a single intentionally bad coffee shop website, using the exact same prompt 100 times in a row: look at the site, think outside the box, ship one bold change that improves conversion or feel, no small tweaks, every round. This prompt had to be aggressive because earlier attempts saw the model stall out around round 20 to 25, declaring that no further improvements were needed.\n\nTwo models were tested under identical conditions: [Claude Sonnet 5.5 and Claude Opus 5.5](https://goremy.ai/blog/articles/gpt-6-sol-vs-claude-opus-5-5-for-coding-the-real-cost-math), both on medium reasoning effort. Each ran in a clean room, meaning no CLAUDE.md file, no custom skills, no plugins, and no memory of prior sessions. Every round started fresh, with the model seeing only the current state of the site and whatever built-in tools it had access to. This matters for the pricing conversation because it isolates the cost of raw iteration, not the cost of a model leaning on saved context or reusable instructions.\n\n## What did $30 of Sonnet usage actually buy?\n\nSonnet’s $29.93 run produced a site that moved from a bare-bones layout to something with a hero background image, scroll-based animations, a sticky order bar, AI-generated images tied to menu items, a background video, and eventually a working order form with pickup scheduling. The milestone log showed real structural jumps at round 1 (first image, first JavaScript, mobile-ready), round 2 (first animation), round 7 (first working image on the page), round 24 (first AI-generated video), and round 32 (first functional form).\n\nAfter round 32, nothing in the next 68 rounds qualified as a milestone. Lines of code kept growing in a straight line, but the “API value per round” metric moved up and down rather than tracking that growth. In practice, this means a meaningful chunk of that $30 budget was spent on rounds that added code without adding noticeable value. At one point the model even removed a feature (a navbar element present at round 65 was gone by round 100), and it swapped the site’s font late in the run seemingly just to have something to change.\n\n## What did $137 of Opus usage buy, and was it worth 4.5x more?\n\nOpus hit its milestones faster and with more polish per round. Round 1 already included a generated image, animation, and mobile readiness, more progress on a single first pass than Sonnet’s. By round 7 it had an AI-generated video background. Round 15 added a working form plus a QR code feature that could text an order summary based on in-site selections, a genuinely functional piece of UX. Round 32 introduced canvas and WebGL elements, specifically a flavor-matching tool that recommends drinks based on mood input.\n\n- ✕a coding agent\n- ✕no-code\n- ✕vibe coding\n- ✕a faster Cursor\n\nThe one that tells the coding agents what to build.\n\nThat’s where the extra spend was clearly earning its keep. But the pattern from Sonnet repeated, only louder. By round 100, the Opus version had accumulated so many forced additions (a pumpkin-carving mini-interaction, a caffeine-and-sleep calculator, a non-functional language toggle, background music options, a cluttered review board) that large parts of the page stopped working correctly or became unusable. One previously functional section broke outright. The creator’s own assessment was blunt: past a certain point, the site became AI slop, with more features but a worse user experience than versions from 30 or 40 rounds earlier.\n\n## Why didn’t more spending produce proportionally better results?\n\nThe core issue isn’t model capability, it’s the prompt structure. Telling an agent it “must ship a change every time” guarantees that every single API call produces output, regardless of whether a genuine improvement exists to make. Once the obvious, high-value changes are exhausted (which both models reached within the first third of their runs), the remaining budget goes toward manufactured changes: cosmetic tweaks, speculative features, and eventually changes that actively hurt the product.\n\nThis has a direct pricing implication for anyone running agentic coding loops: cost scales with round count regardless of whether value does. Lines of code is not a proxy for quality, and in this test it kept climbing linearly in both runs while perceived quality and the “value per round” metric plateaued or declined. The practical lesson is that letting an agent run indefinitely on vague, open-ended instructions is a way to burn API budget without a matching return, particularly on a more expensive model where each round costs more but doesn’t necessarily produce proportionally more useful work.\n\n## Is Opus worth the extra cost over Sonnet for this kind of task?\n\nBased on this single test, Opus produced richer features earlier (the QR-code ordering flow and the WebGL-based flavor matcher stood out as genuinely useful), and it used far more generated media overall (54 images and three videos versus Sonnet’s 11 images and two videos). For teams that want maximum creative output per round and are willing to pay for it, that’s a real difference. But the 4.5x price gap wasn’t matched by a 4.5x quality gap by round 100. Both models ended up overengineering the site once forced past their natural stopping point, and Opus’s final version had more visible bugs and usability problems than its own round-32 state. The takeaway isn’t that one model is better across the board, it’s that the value of either model drops sharply once you force iteration past the point where real improvements run out, and that drop happens regardless of price.\n\n## Frequently Asked Questions\n\n### How much did the 100-round Sonnet run cost in API billing?\n\nThe Sonnet 5.5 run, on medium reasoning effort, cost $29.93 in API billing for all 100 rounds combined, including the generation of 11 AI images and two AI videos.\n\n### How much did the 100-round Opus run cost?\n\nThe Opus 5.5 run, also on medium effort, cost $137 for the full 100 rounds, which included 54 AI-generated images and three AI-generated videos.\n\n### Why did the model keep making changes even after the site looked finished?\n\nThe prompt explicitly required a shipped change every round with no option to stop, which is why earlier unconstrained attempts had stalled around round 20 to 25. Forcing continuous output past the point of genuine improvement is what led to overengineering and, eventually, broken features in both runs.\n\n### Did spending more on Opus produce a proportionally better website?\n\nNot consistently. Opus produced more polished early milestones and more generated media, but by round 100 its site had usability problems and bugs that weren’t present in earlier rounds, suggesting the extra cost didn’t translate into a proportionally better end result once forced iteration ran past useful changes.\n\n### What’s the practical lesson for people running their own Claude Code sessions?\n\nRound count and lines of code are not reliable proxies for site quality, and letting an agent iterate indefinitely on vague instructions tends to produce diminishing or negative returns well before 100 rounds, regardless of which model is used.", "url": "https://wpnews.pro/news/claude-code-pricing-what-100-rounds-of-ai-website-edits-cost", "canonical_source": "https://www.mindstudio.ai/blog/claude-code-api-pricing-experiment-cost/", "published_at": "2026-10-30 00:00:00+00:00", "updated_at": "2026-10-30 10:49:52.076819+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "large-language-models", "ai-products"], "entities": ["Claude Code", "Claude Sonnet 5.5", "Claude Opus 5.5", "Anthropic", "Remy"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/claude-code-pricing-what-100-rounds-of-ai-website-edits-cost", "markdown": "https://wpnews.pro/news/claude-code-pricing-what-100-rounds-of-ai-website-edits-cost.md", "text": "https://wpnews.pro/news/claude-code-pricing-what-100-rounds-of-ai-website-edits-cost.txt", "jsonld": "https://wpnews.pro/news/claude-code-pricing-what-100-rounds-of-ai-website-edits-cost.jsonld"}}