Long-Context Pricing Cliffs: >100K, >200K, and >272K Traps

Full-Request Repricing vs. Marginal Tiers

In LLM billing, long-context cliffs are step functions applied to the whole request, not progressive marginal tax brackets. If a model triggers a long-context threshold at 100,001 tokens, every single input token, cached read, cache write, and output token in that API call jumps to the higher tier.

Only models that explicitly advertise flat 1M pricing—such as Claude Sonnet 5.5 ($2/$10), Claude Fable 5.2 ($8/$40), Gemini 4 Carbon, Gemini 3.8 Flash, DeepSeek V4.1 Flash, and Mistral Large 4—avoid mid-window rate jumps.

Verified Threshold Cliffs (November 2026)

Claude Haiku 5.5 (>100K total input = input + cache_read + cache_write): Jumps 5x across the board—from $0.10/$0.50 ($0.01 cached read, $0.125 5m write) to $0.50/$2.50 ($0.05 read, $0.625 5m write, $1.00 1h write). Once a prompt crosses 100K tokens, Haiku 5.5 flips from 3x cheaper than DeepSeek-V4.1-Flash Peak ($0.10 vs. $0.30 input) to 1.67x more expensive on input ($0.50 vs. $0.30) and 2.08x more expensive on output ($2.50 vs. $1.20).

Gemini 3.1 Pro Preview (>200K prompt tokens): Doubles standard input from $2.00 to $4.00/1M, doubles cached read from $0.20 to $0.40/1M, and raises output by 1.5x from $12.00 to $18.00/1M (storage stays $4.50/1M/hr).

OpenAI GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna (>272K input tokens): Doubles input, cache read, and cache write rates (2.0x) and multiplies output by 1.5x for the full request. For instance, GPT-6.1 Sol jumps from $2.00/$10.00 ($0.10 read) to $4.00/$15.00 ($0.20 read), within a 922K max input limit.