Claude Code sends 33k tokens before reading the prompt; OpenCode sends 7k
AI coding assistants like Anthropic’s Claude Code are drawing criticism for sending tens of thousands of “overhead” tokens—system prompts, tool descriptions, and subagent setup—before even reading a user’s request, dramatically inflating costs compared with leaner harnesses such as OpenCode or Pi. Commenters argue this token bloat is exacerbated by aggressive use of tools and subagents, and may reflect a conflict of interest when the model provider also controls the agent that spends the tokens. Many developers are responding by building their own lightweight agents, pruning context, or switching to alternative harnesses and models that offer more transparency, control, and cost efficiency.
Token overhead & harness design
- Many note Claude Code’s large system prompt and tool descriptions (tens of thousands of tokens) vs OpenCode’s much smaller payload; Pi and some tiny CLIs are even leaner.
- Some argue raw prompt size isn’t the right metric: what matters is intelligence × cost × time and overall task completion.
- Others counter that “tokenflation” is real: trivial tasks seemingly consume more tokens over time.
Subagents and runaway costs
- Subagents are repeatedly cited as the biggest token sink: users report tasks spawning dozens or even hundreds of subagents, blowing through quotas quickly.
- Several disable or heavily restrict subagents (via settings, permissions, or system instructions) and find sequential single-agent workflows cheaper and often no worse in quality.
- Others use subagents selectively (e.g., cheaper models for exploration, specialized reviewers) and find value when carefully orchestrated.
Caching, pruning, and technical details
- Some point out that large static prefixes should hit provider KV caches and be billed more cheaply or even free on subscriptions, reducing the effective cost after the first call.
- Others note short cache TTLs (5 minutes–1 hour) and opaque behavior, so savings are fragile.
- Context-pruning tools (e.g., Dynamic Context Pruning / Sleev) draw skepticism: they may reduce tokens but risk degrading capability and busting caches; lack of transparency is a concern.
Cost models & incentives
- Strong debate over incentives:
- One side claims Anthropic is “token-maxing,” especially by tying flat-fee subscriptions to their own harness and making Claude Code verbose.
- Others argue competition and GPU scarcity push everyone toward efficiency; wasteful behavior risks losing users to cheaper models (DeepSeek, open models, etc.).
- Some data points: users see hypothetical API costs for similar workflows rising over months despite similar work volume.
Alternative harnesses and DIY
- Many praise OpenCode, Pi, Codex, and minimal custom harnesses (simple REPLs or 50–200 LOC agents) as cheaper, more controllable, and often equally or more effective.
- Pi is seen as intentionally minimal and pluggable; some love that, others find it too bare and dependency-heavy in practice.
- Several recommend local models plus a custom harness to avoid provider lock-in and observe exactly what’s sent.
Quality, UX, and opacity
- Mixed reports on Claude Code quality: some say it’s still the most advanced orchestrator for complex, parallel work; others find OpenCode or Codex faster, clearer, and less “over-engineered.”
- Complaints about Claude Code becoming more opaque over time: hidden thinking tokens, less visibility into tool use, and more automatic behaviors (tests, reviews, subagents) that users didn’t explicitly request.
- A common pattern emerges: people end up adding explicit instructions (e.g., no subagents, no tests for trivial changes, minimal system prompt) to regain cost and behavioral control.