# What a million-token context window actually costs to use > Anthropic bills a 900k-token request at the same rate as a 9k one. Same rate, a hundred times the tokens, and $4.50 a call on Opus 5. URL: https://getculpa.com/long-context-cost Last reviewed: 2026-08-05 Rates effective: 2026-07-02 ## Answer Anthropic charges no premium for long context: a 900,000-token request bills at the same per-token rate as a 9,000-token one. The rate holds and the token count doesn't, so the same request costs a hundred times more. Culpa, a local-first LLM cost, margin, and forecast ledger, prices each call from its own token count. ## Why this happens Long context is sold as a capability and priced as a volume, and the gap between those two framings is where budgets break. Anthropic is explicit that Claude 4.6 and later include the full million-token window at standard pricing, and states plainly that a 900,000-token request bills at the same per-token rate as a 9,000-token one. Caching and batch discounts apply across the whole window too. All of this is genuinely good news and it's routinely misread as long context being cheap. The rate is flat. The bill isn't, because you're sending a hundred times the tokens. A single 900,000-token request on a frontier model costs more than most teams' entire daily spend, and the failure mode is an agent that keeps appending to its context until every subsequent call carries the whole history. No rate change announces that. It just gets slowly more expensive per call while the per-token price on the pricing page stays exactly where it was. ## What this usually looks like - Cost per call climbs through a long conversation while volume is flat. - An agent appends to context and nothing ever trims it. - Someone concluded long context was cheap because there's no premium tier. - A single request costs more than you expected an entire session to. - Nobody knows your average context length, only your call count. ## Common mistakes - Reading no premium as cheap. Why it hurts: A flat rate on a hundred times the tokens is a hundred times the cost. Do instead: Price the request, not the rate, and watch context length as its own metric. - Letting an agent grow its context unbounded. Why it hurts: Every later call re-sends the whole history, so cost per call climbs with conversation age. Do instead: Trim, summarise or checkpoint, and measure cost per call against turn number. - Not caching a long stable prefix. Why it hurts: Caching applies across the full window, so a large reused prefix is the clearest saving available. Do instead: Cache the stable part and confirm the reads are actually landing. - Tracking calls instead of tokens. Why it hurts: Call count is flat while context grows, so the metric that moves is the one nobody plotted. Do instead: Track tokens per call over time, which is where this shows up first. ## Self-check - Plot your average input tokens per call over the last month. - Find your largest single request and price it on its own. - Check whether cost per call rises with conversation turn number. - Work out what share of your input tokens is a stable prefix you could cache. ## The same rate, a hundred times over Anthropic states that Claude 4.6 and later models include the full 1M-token context window at standard pricing, and that a 900k-token request is billed at the same per-token rate as a 9k-token request. Priced on the price book's real rates: Claude Sonnet 5 at $2.00 per million input under its introductory pricing, and Claude Opus 5 at $5.00. Input only, since the output is unchanged between the two. Everything below is published rather than assumed. 9,000 input tokens on Sonnet 5: 9,000 x $2.00/M = $0.0180 900,000 input tokens on Sonnet 5: 900,000 x $2.00/M = $1.8000 the rate is identical, the request costs 100 times as much the same 900,000 tokens on Opus 5: 900,000 x $5.00/M = $4.50 per request Four dollars fifty for one request, on a page that correctly says there's no long-context premium. Both facts are true at once, and only one of them ends up in the summary somebody repeats. The number to watch isn't the rate, it's your average tokens per call, and that one moves without anybody publishing a change. ## Cost figures Every figure carries its confidence and its source. No figure on this page is provider-reported. - $1.80 against $0.0180, input cost of a 900k-token request and a 9k-token request on Claude Sonnet 5, at the same rate [calculated] Source: Anthropic's pricing page, read 2026-08-05, states that Claude 4.6 and later include the full 1M-token context window at standard pricing and that a 900k-token request is billed at the same per-token rate as a 9k-token request. At Claude Sonnet 5's introductory $2.00 per million from the price book effective 2026-07-02, 900,000 tokens is $1.80 and 9,000 is $0.0180, a factor of exactly 100. The same 900,000 tokens on Claude Opus 5 at $5.00 per million is $4.50. Rates and the no-premium statement are published, the request sizes come from Anthropic's own example. ## FAQ Q: Does a longer context cost more per token? A: Not on Claude 4.6 and later. Anthropic includes the full million-token window at standard pricing and says a 900k-token request bills at the same per-token rate as a 9k one. Prompt caching and batch discounts apply across the full window as well. Q: So is long context cheap? A: The rate is unchanged and the bill isn't. A 900,000-token request on Sonnet 5 costs $1.80 of input against $0.0180 for a 9,000-token one, exactly 100 times, and $4.50 on Opus 5. No premium and expensive are both true at the same time. Q: Why is my cost per call rising when volume is flat? A: Usually context growth. An agent that appends to its history re-sends everything on every subsequent call, so cost per call climbs with conversation age while your call count stays level. Plot tokens per call against turn number and it shows up immediately. Q: What's the cheapest way to use a long context? A: Cache the stable part. Caching applies across the full window, and a cache read costs a tenth of the base input rate, so a large prefix reused across calls is the clearest saving on offer. After that, trim or summarise the part that isn't stable. ## Sources - Anthropic pricing: https://platform.claude.com/docs/en/docs/about-claude/pricing Run the free Cost Leak Scan: https://app.getculpa.com/scan?source=pseo&slug=long-context-cost&cluster=problem Machine-readable index of every guide: https://getculpa.com/api/pages Human-readable index of every guide: https://getculpa.com/guides Site overview: https://app.getculpa.com/llms.txt Privacy: Culpa runs on your infrastructure. Your prompts and responses never leave it. Culpa counts calls to run your plan, and it fails open, so if it ever breaks your app keeps running.