Claude Code burns 33k tokens before it reads your prompt — OpenCode uses 7k
A new benchmark shows Claude Code front-loads 33,000 tokens of context before touching your actual prompt. OpenCode does the same job in 7,000. That gap costs real money.
A new benchmark shows Claude Code front-loads 33,000 tokens of context before touching your actual prompt. OpenCode does the same job in 7,000. That gap costs real money.
A post from Systima AI is making the rounds on Hacker News with 633 points: Claude Code sends roughly 33,000 tokens of system context before it even reads your prompt. OpenCode, an open-source alternative, does the same setup in about 7,000 tokens. If you are paying per token on a busy project, that overhead adds up fast.
The Systima AI breakdown compares the two coding agents side by side. Claude Code's 33k-token preamble includes tool definitions, system instructions, and context scaffolding that Anthropic bakes in. OpenCode trims that to around 7k. On a single prompt the difference is a few cents. Across a full workday of agentic coding sessions, it can mean a meaningfully higher bill — and slower first-token latency.
OpenCode is open-source and runs locally. If you are already comfortable with llama.cpp (which just hit release b9986 on GitHub), you can point OpenCode at a local model and pay nothing per token. That combination is now a real alternative to Claude Code for builders watching costs.
A study covered by IEEE Spectrum finds that AI tools boost individual researcher productivity and career outcomes — but the ideas being explored are getting narrower. Researchers using AI tend to converge on similar directions, which means the overall diversity of scientific inquiry shrinks. This is not a fringe concern: the paper landed 150 points on Hacker News.
For builders, the practical read is this: if you use AI to generate your product ideas, your feature list, or your research directions, you may be converging on the same answers as everyone else. The tool is useful; the output is less original than it feels.
Yael Writes published a piece titled 'Stop Telling Me to Ask an LLM' that hit 217 points. The argument: reflexively pointing people at ChatGPT or Claude as a first answer is often unhelpful, especially for questions that require lived experience, community knowledge, or nuanced judgment. Worth reading if you are building anything that involves directing users toward AI answers.
llama.cpp tagged release b9986 this week. No dramatic changelog item stands out, but the project continues its steady cadence. If you run local models, staying current here matters — each release tends to include quantization improvements and new model format support.
empero-ai's Qwythos-9B-v2 and its GGUF variant both jumped into the top 30-31 on Hugging Face models this week. A 9B model in GGUF format runs on a machine with 8-16 GB of RAM. No independent benchmark from CBW yet — but the ranking movement suggests community interest.
1. Run a token audit on your Claude Code sessions. Open the Systima AI post, check your own API logs, and calculate your real monthly overhead. Then download OpenCode from GitHub and run the same task to compare output quality and token cost side by side.
2. Try a local coding setup. Install llama.cpp b9986, pull down Qwythos-9B-v2-GGUF from Hugging Face, and point OpenCode at it. You get a zero-per-token coding assistant that runs offline. Good weekend project if you have a machine with 16 GB RAM.
3. Before your next product brainstorm, write down three directions without AI first. Then use Claude or GPT to generate ideas and compare. The IEEE Spectrum research suggests AI narrows the idea space — testing this on your own work costs nothing and might surprise you.
Confirmed: The Systima AI post comparing Claude Code (33k tokens) and OpenCode (7k tokens) reached 633 points on Hacker News as of the time of writing, indicating broad community interest.
Not independently verified by CBW: We have not run our own token-count test on Claude Code or OpenCode. The 33k/7k figures come from Systima AI's methodology, which we have not audited.
Confirmed: llama.cpp release b9986 is live on GitHub at the linked URL.
Not independently verified by CBW: Qwythos-9B-v2's quality or benchmark scores. Hugging Face rank movement reflects downloads and likes, not independent evaluation.
Worth noting: The IEEE Spectrum piece on AI narrowing scientific discovery is based on a study — CBW has not read the underlying paper, only the IEEE summary. Treat the framing as a prompt to read further, not a settled finding.
Source: Systima AI — Claude Code vs OpenCode token overhead — https://systima.ai/blog/claude-code-vs-opencode-token-overhead
Source: IEEE Spectrum — AI boosts research careers but narrows discovery — https://spectrum.ieee.org/ai-science-research-flattens-discovery
Source: Yael Writes — Stop Telling Me to Ask an LLM — https://blog.yaelwrites.com/stop-telling-me-to-ask-an-llm/
Source: GitHub — llama.cpp release b9986 — https://github.com/ggml-org/llama.cpp
Source: Hugging Face — Qwythos-9B-v2-GGUF — https://huggingface.co/empero-ai/Qwythos-9B-v2-GGUF
Get tomorrow's best AI project in your email. With a guide that works. Free. No spam.