Short version: GPT-5.5's rate card says $5 / $30 per million tokens. That's true — until your input crosses 272K tokens. Above that line the request bills at $10 / $45 (2x input, 1.5x output), and the higher rate applies to the whole request, not just the tokens past the cliff. GPT-5.5's window is 1M, so a long-running agent can slide over 272K with no error and a silently doubled bill. Here's the mechanic and four ways to stay under it.

It's a cliff, not a slope#

The word "long-context pricing" makes it sound like a metered surcharge — you pay the base rate up to 272K, then a bit more on the overflow. It isn't. Once a single request's input passes 272,000 tokens, OpenAI reprices the entire call: every input token at 2x ($10/M) and every output token at 1.5x ($45/M).

That means the marginal cost of the tokens that push you over the line is enormous. A request at 271K input tokens bills at $5/$30. The same request with 3K more context — 274K — bills at $10/$45 across all 274K input tokens and every output token it generates. Two thousand tokens of extra context can add far more than two thousand tokens of cost. There is no partial rate; you are on one side of the cliff or the other.

The 271K-to-273K request isn't 1% more expensive. It falls off a cliff — the whole session reprices the instant you cross the line.

Why agents fall off it and chatbots don't#

A human chat almost never approaches 272K tokens. An agent does — routinely, and by accretion. Every loop, it appends: tool outputs, retrieved documents, prior reasoning, a growing transcript, persistent memory it re-injects each turn. None of that throws an error, because GPT-5.5's window is 1M tokens — nearly four times the cliff. The model happily accepts the 300K-token prompt. The invoice is where you find out.

This is the cruel geometry of it: the large window is precisely what lets you sail past the cheap tier without noticing. The context ceiling and the price cliff are 728K tokens apart, and everything in that gap is billed at double.

Four ways to stay under 272K#

1. Cap working context with compaction. Keep the live prompt small even as the agent's history grows — summarize old turns, drop stale tool output, and carry forward only what the next step needs. This is the same discipline that keeps latency down; see context editing vs compaction for long-running agents.

2. Cache the stable parts. Your system prompt, tool definitions, and durable history don't change turn to turn. Prompt caching bills re-sent context at the cheaper cached-input rate ($0.50/M standard), which both cuts the input cost and keeps the tokens you pay full freight on lower. Start with caching your tool definitions.

3. Retrieve less. The fastest way over the cliff is dumping a whole repo or corpus into the prompt "just in case." Pull the specific files and passages the task actually needs. Smaller retrieval keeps you under 272K and usually improves answer quality too.

4. Route long-context jobs to a flat-rate model. Claude Sonnet 5 has no long-context tier — one flat rate across its full 1M window ($2/$10 introductory, then $3/$15). If a task genuinely needs 400K tokens of context, a cliff-free model can be cheaper than GPT-5.5 at its long-context rate, even before you weigh capability.

"Flat" isn't the same as "cheap"#

Sonnet 5's lack of a cliff is real, but don't read it as a free lunch. Sonnet 5 ships a new tokenizer that can turn the same text into up to ~35% more billable tokens — prose-heavy workloads sit near the top of that range. So its effective cost rises even though the per-token rate card looks unchanged. The two models hide their long-context cost in different places: GPT-5.5 in a threshold, Sonnet 5 in the token count.

The takeaway generalizes past these two models: read the pricing mechanic, not the headline number. A rate card that shows one price per million tokens can still bill you two very different ways depending on how long your context runs and how its tokenizer counts. Before you commit an agent to a model, read its pricing page like an adversary — and measure cost per completed task, where the cliff actually shows up.