{
  "id": 7177300,
  "title": "Your LLM cost estimate is wrong above 200,000 tokens",
  "url": "https://urgent.news/2026/09/13/your-llm-cost-estimate-is-wrong-above-200-000-tokens",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-13T21:12:32.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/llmwaves/your-llm-cost-estimate-is-wrong-above-200000-tokens-26ck"
  },
  "original_language": "en",
  "account": "The cost estimate for large language models (LLMs) is often incorrect when dealing with prompts that exceed 200,000 tokens. The widely cited rate of $2 per million input tokens and $12 per million output tokens only holds true up to a context length of 200,000 tokens; beyond that threshold, the price doubles. This discrepancy is not isolated to a single model, as Gemini 3.1 Pro and Grok 4.6 both apply the same pricing structure above the 200,000 token mark. OpenAI's pricing page subtly mentions that their published rates apply to contexts under 270,000 tokens, leaving no explicit rate beyond that limit. Similarly, Claude's pricing model changes once it surpasses 4.6, eliminating context tiers and maintaining consistent rates across all token quantities within the 1 million token window.\n\nThe significance of this issue extends beyond theoretical comparisons; it has tangible implications for users employing large-scale applications such as coding agents or Retrieval-Augmented Generation (RAG) pipelines. For instance, a coding agent handling a substantial repository in context or a RAG pipeline that triggers the model's full context window could face unexpectedly higher costs when the prompt exceeds 200,000 tokens. In short prompts, Gemini 3.1 Pro's input rate may appear 2.5 times cheaper than Claude Opus 5's; however, once the token count surpasses 200,000, this advantage diminishes significantly due to the doubled pricing of the latter.\n\nDeepSeek stands apart from other models because it does not adhere to a context tier system. Instead, it adjusts its pricing based on the time of day, with rates doubling during peak hours from 01:00-04:00 and 06:00-10:00 UTC. Consequently, the actual cost of using DeepSeek is not a fixed rate but rather a variable that fluctuates depending on when the traffic occurs. Additionally, the assumption that caching reduces costs by 10% is a common misconception. While this approximation holds for some models like Claude Fable 5.1, which reads cached data at 0.025x the base rate, it does not apply uniformly. For example, Gemini 3.5 Flash-Lite lacks any form of context caching, meaning that any comparison based on assumed cache discounts would be inaccurate. These observations underscore the importance of understanding the specific pricing structures and limitations of each LLM provider when conducting cost analyses.",
  "summary": "I went looking for a straight answer to \"what will this cost per month\" and found that almost every comparison table gets it wrong the moment prompts get long. Not slightly wrong. Double. The threshold nobody shows you Gemini 3.1 Pro is $2 per million input tokens, $12 per million output. That is the number in every roundup. It holds up to 200,000 prompt tokens, and above that line Google doubles…",
  "key_points": [
    "Cost estimate for LLMs is inaccurate beyond 200,000 tokens",
    "Pricing doubles for prompts exceeding 200,000 tokens",
    "DeepSeek's pricing varies by time of day"
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}