July 12, 2026
Token gluttony goes public
Claude Code sends 33k tokens before reading the prompt; OpenCode sends 7k
People are roasting Claude Code for eating the bill before the chat even starts
TLDR: A test found Claude Code can use far more paid AI input than OpenCode before your real request even begins, which matters because it can raise cost and slow things down. Commenters split between “that’s the price of a powerful assistant” and “these agents are wildly overengineered.”
The big number that set people off? Claude Code reportedly burns about 33,000 words' worth of AI budget before it even reads your actual request, while OpenCode comes in around 7,000 on the same setup. In plain English: some users feel like they’re being charged for the appetizer, the tablecloth, and the waiter’s life story before they even say “hi.” The article says that in more realistic work setups, the hidden pre-chat baggage can balloon to 75,000 to 85,000 tokens per request, and using sub-agents can send costs flying even higher.
And wow, the comments did not stay calm. One camp basically shrugged and said, of course this happens: if Anthropic wants the “best coding agent,” then expensive overkill is the whole strategy. Another crowd sounded far less forgiving, accusing modern AI helpers of trying too hard to look smart by rummaging around everywhere and chewing through paid usage in the process. The funniest jab came from users pointing out that even a tiny greeting can trigger a ridiculous flurry of tool use, turning “Hey” into a full-blown office fire drill.
There was also some nerdy side-eye over the test itself, with one commenter sarcastically asking whether the article and the benchmark were both done by AI because of the model choice. Meanwhile, others jumped in with alternatives, flexing that smaller agents like pi keep things leaner. The vibe: half outrage, half vindication, and a lot of “we KNEW these agents were doing too much.”
Key Points
- •In the minimal one-line reply test, Claude Code reportedly sent about 33,000 tokens before the user prompt, compared with about 7,000 for OpenCode on claude-sonnet-4-5.
- •A rerun on claude-fable-5 reduced the Claude Code versus OpenCode overhead gap to about 3.3x, indicating the multiple depends on the model.
- •The article says OpenCode reused a byte-identical cached prefix, while Claude Code repeatedly rewrote large prompt-cache segments and used up to 54x more cache tokens on the same task.
- •Repository instruction files and MCP servers materially increased request size, with a 72KB instruction file adding about 20,000 tokens and five MCP servers adding 5,000 to 7,000 more.
- •On a multi-step task, Claude Code's total token use came out lower than OpenCode's because it batched tool calls into fewer requests despite its higher initial overhead.