As of October 2026, Claude Sonnet 5.5, GPT-5.5, and Gemini 3.1 Pro Preview all advertise roughly 1 million-token context windows, but their per-token prices and the fine print around those windows differ enough to change which model is cheaper for a given workload by more than 2x. This post pulls the numbers straight from each provider’s own pricing documentation — Anthropic’s, OpenAI’s, and Google’s — and runs one real prompt through all three so the difference isn’t abstract.
Why This Comparison Is Harder Than It Looks
A context window is the maximum number of tokens — word pieces, roughly three-quarters of a word each in English — a model can hold in one request, counting both what you send and what it generates back. The three providers rename and re-version their flagship models often enough that a comparison written even two months ago can be pointing at a model that’s already been superseded: this one covers Claude Sonnet 5.5 and Claude Opus 5.5 (Anthropic’s current flagships), GPT-5.5 (OpenAI’s), and Gemini 3.1 Pro Preview (Google’s), as listed on each provider’s pricing page in October 2026. If you’re reading this later, treat the structure of the comparison — tiered pricing, caching discounts, context ceilings — as the durable part, and re-check the dollar figures against the links at the end.
Context Window Size, Compared
All three now sit close to the 1-million-token mark, but the number means something slightly different for each one.
| Model | Context window | Max output |
|---|---|---|
| Claude Sonnet 5.5 / Opus 5.5 | 1,000,000 tokens (standard pricing, per Anthropic’s 4.6+ rule) | Not capped separately in the pricing page |
| GPT-5.5 | 1,050,000 tokens | 128,000 tokens |
| Gemini 3.1 Pro Preview | ~1,000,000 tokens (per Google Cloud’s model page) | Not stated on the pricing page |
GPT-5.5’s window is the only one of the three specified down to an odd number (1,050,000) rather than a round one, and it’s also the only one of the three with a published hard output cap in the same breath as the context figure. That cap matters for anything that generates long documents or large code diffs in one pass — Claude and Gemini’s documentation don’t surface an equivalent limit in the same place, though in practice output is still bounded by the overall context window.
Input and Output Pricing, Compared
Prices are quoted per million tokens (MTok) in US dollars, at each provider’s standard, non-batch rate.
| Model | Input / MTok | Output / MTok |
|---|---|---|
| Claude Sonnet 5.5 | $2.00 | $10.00 |
| Claude Opus 5.5 | $4.00 | $20.00 |
| GPT-5.5 (≤272K input) | $5.00 | $30.00 |
| Gemini 3.1 Pro Preview (≤200K input) | $2.00 | $12.00 |
At standard rates and short prompts, Claude Sonnet 5.5 and Gemini 3.1 Pro Preview tie on input price, Gemini is slightly more expensive on output, and GPT-5.5 is the most expensive on both at this tier. None of this holds once a prompt crosses each provider’s size threshold, which is where most head-to-head comparisons stop looking.
The Tiers That Don’t Make the Headline Number
Both OpenAI and Google quietly raise the price once a prompt passes a size threshold, and the threshold is lower than most people expect.
- GPT-5.5: prompts over 272,000 input tokens are billed at 2x the input rate ($10.00/MTok) and 1.5x the output rate ($45.00/MTok) — and OpenAI’s documentation specifies this applies to the full session once the threshold is crossed, not just the overflow tokens.
- Gemini 3.1 Pro Preview: prompts over 200,000 input tokens jump to $4.00/MTok input and $18.00/MTok output, a flat 2x and 1.5x respectively over the sub-200K rate.
- Claude Sonnet 5.5 / Opus 5.5: Anthropic’s standard pricing page lists one flat rate across the full 1M-token window — no second tier appears in the published pricing for these models.
Practically, this means the model that looks cheapest on a short prompt can flip to the most expensive one on a long document, and the crossover point is exactly the kind of detail that’s easy to miss skimming a pricing page once.
Batch Processing: The Discount Most Comparisons Skip
Batch processing means submitting a set of requests to be completed within a longer window — Anthropic’s documentation gives up to 24 hours — in exchange for a flat discount, rather than paying full price for an immediate, synchronous response. For Claude, that discount is 50% off both input and output tokens, and Anthropic’s pricing page states explicitly that it stacks with prompt caching, so a cached, batched request can end up priced at a small fraction of the standard rate. This matters for anything that doesn’t need an answer in real time: bulk document classification, overnight re-summarization of a knowledge base, or evaluation runs against a prompt test suite. It’s worth checking whether GPT-5.5 and Gemini 3.1 Pro Preview’s own batch tiers (both providers have historically offered similar discounts on earlier model generations) still apply at the same rate on these specific models before assuming the discount carries over — pricing pages are usually explicit about this per model, and it’s exactly the kind of detail that goes stale fastest.
Which Pricing Shape Favors Which Kind of Work
The four pricing structures above favor different workloads, not just different budgets. A flat rate across the full context window, which is what Anthropic currently publishes for Claude Sonnet 5.5 and Opus 5.5, is easiest to forecast for a product that occasionally sends very long prompts — a 900,000-token prompt costs exactly proportionally more than a 90,000-token one, with no step change to budget around. A tiered rate, like GPT-5.5’s and Gemini 3.1 Pro Preview’s, rewards keeping prompts under the threshold deliberately — trimming a 280,000-token prompt down to 270,000 tokens on GPT-5.5 avoids the 2x input multiplier entirely, which is a bigger saving than the token trim itself would suggest. And a workload built around the same large context reused many times over a short span — a codebase a coding agent keeps re-reading, a style guide attached to every call in a session — is where Claude’s 95%-off cache-hit rate on Sonnet 5.5 and Opus 5.5 does the most work, since the discount applies to the expensive first-read tokens on every call after the first.
What Prompt Caching Changes
Prompt caching lets a provider charge less to reprocess input it has already seen recently, and Anthropic’s version of it is the most aggressively discounted of the three at this pricing tier: a cache hit on Claude Sonnet 5.5 or Opus 5.5 costs 0.05x the base input rate — $0.10/MTok for Sonnet 5.5, a 95% discount — against 0.1x (90% off) for Anthropic’s own older Sonnet 5 and most other current Claude models. Writing to the cache costs more than a normal request (1.25x the base rate for a 5-minute cache, 2x for a 1-hour one), so caching only pays off when the same large context — a system prompt, a codebase, a long document — gets reused across several calls within that window. If you’re sending a one-off 150K-token prompt, as in the worked example below, caching doesn’t apply, because there’s nothing cached yet to hit.
For a deeper look at how the three providers’ caching discounts, write costs, and time-to-live windows compare against each other directly, see our context caching comparison, which this post intentionally doesn’t duplicate.
A Worked Example: Pricing One 150K-Token Prompt
To make the tiers concrete, here’s the actual cost of sending a single 150,000-token input prompt and getting a 2,000-token response back, with no caching involved, at each provider’s standard rate.
| Model | Input cost (150K tokens) | Output cost (2K tokens) | Total |
|---|---|---|---|
| Claude Sonnet 5.5 | $0.300 | $0.020 | $0.320 |
| Gemini 3.1 Pro Preview | $0.300 | $0.024 | $0.324 |
| Claude Opus 5.5 | $0.600 | $0.040 | $0.640 |
| GPT-5.5 | $0.750 | $0.060 | $0.810 |
At 150,000 tokens, every model is still under its higher-tier threshold, so this comparison is a fair like-for-like at standard rates. Claude Sonnet 5.5 and Gemini 3.1 Pro Preview land within four-tenths of a cent of each other; GPT-5.5 costs roughly 2.5x what Sonnet 5.5 does for the identical input and a comparable output length. Push the input past 272,000 tokens and GPT-5.5’s cost roughly doubles again on the input side alone, while Claude’s stays on the same flat rate through the full million-token window. Scale the same math up to a 900,000-token prompt — close to the ceiling all three advertise — and the gap compounds: GPT-5.5’s prompt is now well past its own threshold and billed at the 2x/1.5x rate for its entire length, pushing its total meaningfully further ahead of Sonnet 5.5 and Gemini 3.1 Pro Preview than the 150K-token example shows.
Where to Verify This Yourself
Every figure above comes from a provider’s own documentation rather than a third-party aggregator, and all three pages get updated without much notice: Anthropic’s Claude pricing page, OpenAI’s GPT-5.5 model page, and Google’s Gemini API pricing page. If you’re choosing a model for a workload with long, reused context rather than one-off long prompts, read this alongside the reasoning-effort parameter comparison and, if you’re working with PDFs you send repeatedly, Claude’s Files API — both change the effective price per task well beyond what the headline per-token rate suggests.
Photo: CERN data centre by Unnerving duck, licensed under CC BY-SA 4.0, via Wikimedia Commons.
