Long-Context Costs: What 128K Tokens Actually Cost Per Call

A Single 128K-Token Call

Take the maximum context length on each model and assume a 2,000-token completion:

Model	Context	Input cost	Output cost	Total
GPT-4o	128K	$0.32	$0.02	$0.34
GPT-4.1	1.05M	$2.10	$0.016	$2.12
Claude Opus 4.7	200K	$3.00	$0.15	$3.15
Claude Sonnet 4.6	200K	$0.60	$0.03	$0.63
Gemini 2.5 Pro	2M	$2.50	$0.02	$2.52

A single Claude Opus call at full context is $3.15 — more than the entire daily LLM bill of many small applications. Make 1,000 of those a day and you're at $94K/month before any output growth.

Why people do it anyway

Long context replaces retrieval engineering. Instead of:

1. Chunk your codebase
2. Embed 50,000 chunks
3. Run a top-k similarity search per query
4. Pack the top 10 chunks into the prompt

You can simply:

1. Stuff the entire 80K-token codebase into the prompt
2. Ask a question

The trade-off is cost vs. complexity. Long context is right when:

• The corpus is small (a single repo, single document, single contract).
• Recall matters more than precision (you need every relevant fact).
• The query is a one-off (no monthly bill multiplier).

The cache-it-or-die rule

If you're going to put 80K+ tokens in a prompt, you must cache it. Without caching, repeated long-context calls scale linearly in cost; with Anthropic caching they scale at 0.1x. The example in Claude Prompt Caching shows the exact numbers.

Context-window pricing tiers

Both Gemini 2.5 Pro and Claude have higher pricing above a threshold (Gemini: 200K tokens triggers a 2x rate; Claude historically had similar tiers). Always check the provider's pricing page when planning very long calls.