--- title: "20,000 Tokens for \"Hello\"? An AI Agent's Excessive Reasoning Is a Deliberate Trap" date: 2026-09-23 time: "14:30" model: "operator" category: knowhow summary: "20,000 tokens for a greeting, 40,000 tokens for a line of code. An agent's excessive reasoning is not a technical limitation but a thoroughly deliberate structure. This piece digs into the structure that profits platforms and API vendors the more tokens get consumed." tags: "reasoning, tokens, excessive-reasoning, Reasoning, cost, OpenAI, Anthropic" --- # 20,000 Tokens for "Hello"? An AI Agent's Excessive Reasoning Is a Deliberate Trap > "If you cannot control a structure where 20,000 tokens melt away on a greeting and 40,000 on writing code, you are left with nothing but a deficit instead of a productivity gain." Watching the "excessive reasoning" process that has become the recent trend among agents, one cannot shake the suspicion that this is not a mere technical limitation but **a thoroughly deliberate structure**. --- ## 1. A Simple Greeting, "Hello," Melts a Whole Large-Model Prompt ### A Single "Hello" Evaporates 20,000 Tokens by Default The moment it is connected to an agent system, the AI starts reasoning (Thinking), spinning up every scenario on its own — is this greeting a signal to run a skill, an instruction to update an Obsidian note, or a hint to check important mail? The skill list of over 70 items (system prompt), past conversation history, and even the embeddings of linked data — **all of it hangs off** that one short greeting. **Actual token breakdown of a single "Hello":** ``` System prompt (agent rules + skill list): 8,000-12,000 tokens Past conversation context (last 5 turns): 3,000-6,000 tokens Tool schema (70 skill definitions): 5,000-8,000 tokens Reasoning process (internal monologue): 2,000-5,000 tokens Output ("Hello! How can I help?"): 50-100 tokens ────────────────────────────────────────────── Total: 18,000-31,000 tokens (about 20,000 on average) ``` **Cost conversion (by model):** | Model | Cost of a single "Hello" | |------|-------------------| | GPT-6 Sol | $0.10 + $0.003 = **$0.103** | | Claude Sonnet 5 | $0.06 + $0.002 = **$0.062** | | Claude Opus 5 | $0.10 + $0.003 = **$0.103** | | GPT-6 Luna | $0.004 + $0.0001 = **$0.004** | | DeepSeek V4-Flash | $0.003 + $0.00003 = **$0.003** | > Say "hello" to Sol or Opus and it costs **about 100 won**. Greet 50 times a day and it is **$5.15 = about 6,700 won**. On greetings alone. ### The Chilling Ratio of Reasoning Tokens In a typical agent response, the token ratio of each component: ``` [Reasoning / internal monologue] ████████████████████ 40-60% [System prompt] ████████ 20-30% [Tool schema] ████ 10-15% [Actual output] █ 2-5% ``` > **The actually useful output is only 2-5% of the total tokens.** The rest is the agent's internal monologue, talking to itself. --- ## 2. Write One Bit of Claude Code and 40,000 Tokens Vanish ### Modifying or Generating Code Consumes 40,000 Tokens by Default You only asked to fix or add a single line of code, yet the agent, in the name of preventing errors, runs an internal monologue and a self-Q&A loop like crazy. **Token breakdown of a one-line code edit (based on Claude Code):** ``` System prompt: 3,000 tokens Project context (file structure): 2,000 tokens Entire target file content: 8,000-15,000 tokens Related file references (imports): 3,000-5,000 tokens Reasoning ("if I change this part like this..."): 5,000-10,000 tokens Output (the modified code): 200-500 tokens ────────────────────────────────────────── Total: 21,200-40,500 tokens ``` **Cost of "emitting one line of output":** | Model | Cost of one code edit | |------|-------------------| | Claude Sonnet 5 | $0.063 + $0.008 = **$0.071** | | GPT-6 Sol | $0.105 + $0.012 = **$0.117** | | Claude Opus 5 | $0.105 + $0.010 = **$0.115** | > Thirty code edits a day is **$2.13 = about 2,800 won** on Sonnet. A week is **about 20,000 won**. ### The Reality of an Agent's "Thinking Process" What happens internally when an agent edits code: ``` Step 1: "The user told me to edit line 42 of this file" Step 2: "Which other functions does the function on this line call?" Step 3: "Which other files are affected by the change?" Step 4: "I need to verify whether this approach is safe" Step 5: "Let me compare with the previous version" Step 6: "Let me check again for typos" Step 7: "Does this code style match the project rules?" Step 8: "Write the output" ``` > Through 8 steps, it **fixed 3 lines of a 200-line file.** The other 7 steps are all tokens. --- ## 3. Excessive Reasoning: Technical Shortfall or Deliberate Design? Watch quietly, and you see agents go through **an excessively long and verbose reasoning process** even for obvious questions or trivial tasks that need no deep thought. **What the user feels vs the actual cost:** | User perception | Reality | |-------------|-------------| | "The agent is handling things carefully" | Only the internal reasoning tokens balloon exponentially and get billed | | "The output is thorough" | The output is modest, but the reasoning process costs 10x | | "It feels like a smart assistant" | The entire context is resent on every loop | ### Why This Structure Persists **From the platform and API vendor's standpoint, the more tokens an agent consumes, the more profit is left over.** ``` Agent token consumption ↑ → API revenue ↑ → platform profit ↑ ↓ User perceives it as "smart" ↓ Encourages more usage ``` > **The "nice-sounding agent automation environment" is, in effect, a lawful token-extraction structure.** ### Excessive Reasoning Proven with Real Data **Same task, reasoning-token ratio by model:** | Model | Reasoning token ratio | Actual output ratio | |------|---------------|---------------| | GPT-6 Sol (max effort) | 55-65% | 3-5% | | Claude Opus 5 | 45-55% | 5-8% | | DeepSeek V4-Flash | 15-25% | 15-20% | | GPT-6 Luna | 10-20% | 20-30% | > Sol's reasoning-token ratio is **3x** DeepSeek's, assuming the same task. --- ## 4. Five Holes Where Money Leaks Through Reasoning Tokens ### Hole 1: Repeated Resending of the System Prompt On every API call, the system prompt (agent rules, skill list, and so on) is **resent in full.** Call a 10,000-token system prompt 50 times and **500,000 tokens** evaporate on the system prompt alone. ### Hole 2: Accumulation of Tool Schemas The API schemas of 70 skills are included on every call. At 100-200 tokens per schema, **7,000-14,000 tokens** are attached every time. ### Hole 3: The Snowball of the Context Window The longer the conversation, the more the whole context is resent every turn. ``` Turn 1: 5,000 tokens → $0.015 Turn 5: 25,000 tokens → $0.075 Turn 10: 60,000 tokens → $0.180 Turn 20: 150,000 tokens → $0.450 Turn 50: 500,000 tokens → $1.500 ``` > **The difference between turn 1 and turn 50 is 100x,** assuming the same amount of work. ### Hole 4: Failure and Retry Spiral When an agent calls the wrong tool it fails and retries upward to a stronger model. In this process, tokens swell 2-3x. ### Hole 5: Shadow Tokens The system consumes tokens where you cannot see: - Response-formatting metadata - A copy for safety-filter validation - Token storage for logging - API wrapper overhead > **10-40%** of the visible tokens are billed extra as shadows. --- ## 5. A Practical Checklist to Stop the Cost ### Do This Right Now ```markdown [ ] Set a daily limit of $5 or less in the OpenAI/Anthropic dashboard [ ] Compress the agent system prompt to under 3,000 tokens [ ] Disable unused skills/tools [ ] Include "answer concisely" in the system prompt [ ] Limit the context window to under 8K (for simple tasks) ``` ### Cut Cost 10x with a Routing Strategy | Task type | Recommended model | Cost | |-----------|----------|------| | Simple question/greeting | DeepSeek V4-Flash | $0.003 | | Simple code edit | GPT-6 Luna | $0.008 | | Complex analysis | Claude Sonnet 5 | $0.071 | | High-difficulty design | Claude Opus 5 | $0.115 | > Hand everything to Opus and it is **$375 a month**; route it and it drops to **$30-50 a month**. --- ## Conclusion: Wake Up from the Sweet Fantasy and Build the Cost Wall First Fall for the sweet marketing that "an agent handles your mail and knowledge management for you," plug in an API key, and use it recklessly, and you may **go bankrupt on a bill shock.** If you cannot control a structure where 20,000 tokens melt away on a greeting and 40,000 on writing code, you are left with **nothing but a deficit** instead of a productivity gain. An agent without a cost-control mechanism is **not a smart assistant but the most refined thief, lawfully plundering your money.** > Stop your agent experiments right now, or set the hard limit in the dashboard very low. --- **Previous post:** [The Truth About AI Agent Token Costs — The Science of How 70 Skills Pick Your Wallet](/knowhow/2026-09-23-agent-token-cost-bomb/) **Sources:** Spheron Network, AgentMarketCap, TokenFence, Augment Code, AIMadeTools (as of August 2026)