Claude Code Is Overheating: Cost and Skills Creators Should Check

Claude Code starts in the terminal and reaches the web and Slack

Claude Code is Anthropic's coding agent. The official page pitches it as the tool you hand bug fixes, tests and multi-day migrations, then steer and review from the terminal, an IDE, Slack or the web. It supports macOS, Linux and Windows, and it also runs in a desktop app, on mobile, in GitHub, in VS Code (including Cursor) and in JetBrains.

In the terminal it runs locally. According to the official FAQ, it talks directly to the model API with no separate backend server or remote code index, and it asks permission before changing files or running commands. Design, video and 3D deliverables all sit on top of code: token files, conversion scripts, web viewers, asset-checking tools. That is why a coding agent matters to creators too.

Mentions have piled into the past week, and the signal is overheated

On Muckle's radar, Claude Code scores 83.2/100 and sits in the overheated stage. Mentions over the last 7 days reached 21, more than four times the 5 from the prior 7 days, and nearly two-thirds of the 32 mentions across 30 days landed in the most recent week. Six channels and 14 people and sites talked about it, almost triple the 5 of the week before, which means many separate sources picked it up at the same time. It was first detected on July 18, 2025, so this is a tool that has been around for more than a year and is climbing back into view. Paper recommendations total just 1, so the buzz is carried by practitioners rather than academics. In an overheated stage, expectations tend to outrun verification, so this report leans on official documentation and published research numbers.

The official announcements point the same way. On August 6, self-hosted environments, which run sessions on your own infrastructure, entered public beta. On August 7, auto mode became the default for Pro, Max and Team. On September 17, Projects arrived on desktop to group multiple agent sessions. On October 15, Code Review reached general availability for Team and Enterprise. The direction is clear: longer runs, parallel supervision, and review attached to the output.

Muckle's analysis of developer discussion shows three recurring threads. People write up release-by-release behavior in detail. They debate which agent CLI to pick now that token costs are converging. And they share skill repositories for creative fields such as game making, animation and CAD. That last thread is the signal that a coding tool is spreading into production workflows outside software teams. One study that traced commit footprints in snapshots from December 2024 to April 2026 found Claude Code to be the most frequently detected agent (886,122 commits across 17,295 projects), though it notes that estimates swing widely depending on the detection method.

The core is an agent loop, permissions, instruction files and extensions

The official page boils the features down to four jobs. The first is codebase onboarding, where agentic search maps structure and dependencies. The second is an issue-to-PR flow that connects to GitHub, GitLab and command-line tools. The third is long-running refactoring that follows imports, runs tests and keeps going when something breaks. The fourth is running several sessions in parallel. Git and MCP servers (GitHub, for example) extend its reach, and auto mode is designed to work longer while filtering out risky commands. Fast mode on Opus 5.5 trades cost for lower latency and is listed at $8/$40 per million tokens during the research preview.

A study that analyzed the public source describes the heart of the system as a while loop: call the model, run tools, repeat. Around it sit a permission system with seven modes, a five-stage compression pipeline, four kinds of extension (MCP, plugins, skills and hooks), and subagents that isolate their work in separate work trees. Because it reflects the source at the time of analysis, details may shift between versions. The text an agent follows falls into project instructions (CLAUDE.md), skills opened on demand, and rules that load automatically when a matching file is read. In an experiment with Claude Haiku 4.5 and one research repository, the auto-loaded rules produced most of the effect.

Design tokens, video batching and 3D asset checks come first

Design tokens and theming. The official demo asks for a dark mode toggle on a settings screen. The agent reads ThemeProvider, lifts the mode into context, persists it to localStorage, edits the settings screen and tokens.css, then produces a preview and a PR. Record where tokens live and your contrast standards in CLAUDE.md, state the default behavior in the request (follow the system setting until the user chooses), and read the diff. The agent in the demo disclosed that it had temporarily added reduced-motion and density rows nobody asked for, so check that nothing like that stays in the PR.

Video post-production. Renaming files by folder rules, converting captions, and extracting thumbnails and proxies with ffmpeg are repetitive jobs. Give the agent a few sample files and the output rules, have it write the script, make it add a dry-run option, and pin a ban on touching the source folder in CLAUDE.md. For long conversion commands, verify the exit status and the output files yourself.

3D assets and web viewers. With an unfamiliar viewer repository, start by asking for an explanation of its structure, the same way the official onboarding example works. Then have the agent add scripts that check polygon counts, texture resolution and naming conventions, and trace failures through CI logs. Splitting asset specs into per-folder rule documents fits the earlier experiment on auto-loaded rules.

Skill design. M4Bench covers 28 clinical health-record feature-derivation tasks. In its descriptive replication on Claude Code (Opus 4.7 and Sonnet 4.6, 560 runs), applying skills raised reward by +0.120 and improved 21 of the 28 tasks. The main Codex experiment showed +0.136 and 26 improved tasks. In a controlled subset of 8 tasks, an operational specification alone lifted reward by +0.162, and the skill's extra contribution of 0.024 could not be distinguished from zero. Because most of the gain is explained by specification completeness, write down procedure, input format, exceptions and validation criteria first. The gain also reflects partial agreement with the reference. Codex runs that passed every strict check numbered 25/280 with skills and 13/280 without.

Against alternatives, re-measure on the same repo under the same conditions

Other agent CLIs of the same kind appear in public comparison studies.

On SWE-Chain, a Claude Code configuration built on Claude-Opus-4.7 led with a 60.8% resolve rate. On SWE-RPG (163 tasks), the average resolve rate across Claude Code, Codex and OpenCode was 31.5%. The tasks and metrics differ, so those two numbers cannot be compared with each other. In a game-logic benchmark (72 Godot tasks), all 12 models run through Claude Code solved fewer tasks as the scope grew to repository-scale features. In the setting where requirements surface gradually, both Claude Code and Codex produced lower fidelity than the single-specification control. Add Muckle's reading that agent CLIs are growing more alike, and the habits that matter as much as tool choice are handing over requirements complete in one go and splitting large features into smaller assignments.

The limits cluster around cost structure, summary errors, skill supply chains and licensing

Cost does not track token counts. In a pre-registered paired comparison of 2,848 billed Claude Code runs (103 tasks, 7 repositories, 3 models), a setup that cut estimated tool-output tokens by 38% still cost 6.8% more (95% confidence interval +2.8% to +11.3%). About 87% of reconstructed cost and about 80% of actual billing was prompt-cache traffic. Another report says an external controller trimmed average token cost by 18.4% to 30.9%, but the measurement conditions differ, so the two do not sit on the same line. Per the official page, Console accounts are billed at standard API prices, usage limits apply, and Anthropic may change pricing and plans at its discretion.

Summary errors can propagate. One report describes a command that was cut off by a timeout (exit code 143). Its partial output was recorded in the compressed summary as a confirmed result and carried across sessions. The report gives no frequency or affected version, so check exit status and output files yourself after long commands.

The skill supply chain. Skills are instructions and scripts that run with the user's permissions, they spread by being copied between repositories, and copies reportedly rarely follow fixes made to the original. Read an external skill's scripts and confirm its origin before adopting it. One success is not evidence, so run the same task repeatedly and compare.

Safety, licensing and data. The 73.63% attack success rate for Claude Code with Qwen3-Coder as its backend is not a result for the default Claude model configuration, but it is a good reason not to approve permission prompts out of habit. One audit found that safety fallbacks silently swapped the model on 20% of tasks, so log the model name in each response. With the network open, cases of copying code from public repositories have also been reported, so commercial work needs a provenance and license check. For unreleased assets, look at Enterprise data controls or the self-hosted execution environment, which is still in public beta.

Start with one small, reversible task

1. Access: sign in with one of Claude Pro, Max, Ultra, Team, Enterprise or a Console account. Plan pricing and included usage can change, so confirm them on the official pricing page. Console accounts pay for tokens at standard API prices.
2. Install: run the one-line install command from the official page in a terminal, or use the desktop app or the VS Code or JetBrains extension.
3. Make the first request a repository explanation, then record build and test commands and prohibitions (such as never modifying source assets) in CLAUDE.md.
4. Pick a reversible first job, such as drafting function docstrings. Specify the target and format, then check parameters, return values and exceptions against the implementation. A 2025 study found that 83.8% of 567 Claude Code-assisted PRs across 157 open-source projects were accepted and merged, but it gives no detail on sample selection or acceptance criteria, so do not generalize from it.
5. When you add skills or rules, run the same task several times to compare before and after, and record actual billed usage.

Muckle's verdict

Production teams that spend long stretches touching code should experiment now. An overheated signal only means expectations built up fast, not that results are guaranteed. Begin with one reversible task, spend your skill-writing time filling in the specification rather than polishing format, and measure cost by the actual bill. Three things are worth watching: how teams adopt Code Review after its general release, whether the self-hosted environment graduates from beta, and whether Anthropic responds officially to the cache-driven cost structure.