If your prompts run long, stop comparing per-token rates and check whether the model has a length cliff at all. Above its threshold a tiered model bills every token in the request at the higher rate, so crossing by one token can quintuple a turn. Claude Haiku 5.5’s cliff is the steepest (5x everything) and the earliest (100,000 tokens, about 47,000 words). The models with no cliff are every Claude 4.6+ except Haiku 5.5, every Gemini Flash, DeepSeek Flash and V4 Pro, and Mistral Large 4. GPT-4.1 also had no cliff and is the reason this used to have an OpenAI answer; it retires from the API on 14 January 2027. On xAI, every current Grok is tiered. The cheapest long-prompt option on list price is GPT-6 Luna, which stays on standard rates to 272,000 tokens — four times further than Haiku.
The four cliffs are not where the rate cards say they are
Each vendor denominates its threshold in its own tokens, and those tokens are not the same size. Normalising to characters of English text, using each vendor’s own published estimate, moves the picture a long way:
| Vendor | Threshold (vendor tokens) | ≈ characters | ≈ English words | vs Anthropic |
|---|---|---|---|---|
| Anthropic (Haiku 5.5) | 100,000 | 269,200 | 47,000 | — |
| Google (Gemini 3.1 Pro, 2.5 Pro) | 200,000 | 800,000 | 140,000 | 3.0x later |
| xAI (Grok 4.7, 4.3, Build) | 200,000 | 800,000 | 140,000 | 3.0x later |
| OpenAI (GPT-6, GPT-5.6 families) | 272,000 | 1,088,000 | 191,000 | 4.0x later |
Character estimates from each vendor’s own documentation: Anthropic’s glossary and token-counting guide (≈2.69 chars/token on the Claude 4.7+ tokenizer, which Anthropic says produces about 30% more tokens for the same text), OpenAI and Google (≈4 chars/token). xAI publishes no figure; OpenAI’s is assumed.
The raw numbers suggest Anthropic’s threshold is a little under half of OpenAI’s. On the same document it is a quarter of it. A 47,000-word input is not an exotic workload — it is a mid-sized repository, a quarter’s worth of support tickets, or one long technical manual.
The models with no cliff — which is the actual decision
Anthropic states the exemption in writing on its pricing page: Claude 4.6-and-later models except Haiku 5.5 “include the full 1M token context window at standard pricing,” and “a 900k-token request is billed at the same per-token rate as a 9k-token request.” Prompt caching and batch discounts apply at standard rates across the whole window.
| Model | Input / cached / output | Window | Length cliff |
|---|---|---|---|
| DeepSeek Flash | $0.15 / $0.003 / $0.60 | 1M | None (2x at peak hours) |
| GPT-6 Luna | $0.10 / $0.01 / $0.50 | 1.05M | 272,000 → $0.20 / $0.02 / $0.75 |
| Claude Haiku 5.5 | $0.10 / $0.01 / $0.50 | 1M | 100,000 → $0.50 / $0.05 / $2.50 |
| Gemini 3.5 Flash-Lite | $0.30 / $0.03 / $2.50 | 1M | None |
| DeepSeek V4 Pro | $0.66 / $0.022 / $1.98 | 1M | None (2x at peak hours) |
| Gemini 3.8 Flash | $0.75 / $0.075 / $3.75 | 1M | None (promotional to 31 Dec 2026) |
| Mistral Large 4 | $1.36 / $0.14 / $4.18 | 1M | None |
| Grok 4.3 | $1.25 / $0.20 / $2.50 | 1M | 200,000 → $2.50 / $0.40 / $5.00 |
| Claude Sonnet 5.5 | $2 / $0.10 / $10 | 1M | None |
| GPT-6.1 Sol | $2 / $0.10 / $10 | 1.05M | 272,000 → $4 / $0.20 / $15 |
| GPT-4.1 (retires 14 Jan 2027) | $2 / $0.50 / $8 | 1.05M | None |
| Gemini 3.1 Pro | $2 / $0.20 / $12 | 1M | 200,000 → $4 / $0.40 / $18 |
| Grok 4.7 | $2 / $0.50 / $6 | 500K | 200,000 → $4 / $1.00 / $12 |
| Claude Opus 5.5 | $4 / $0.20 / $20 | 1M | None |
| GPT-6 Astra | $10 / $1 / $50 | 1.05M | 272,000 → $20 / $2 / $75 |
USD per million tokens, standard (non-batch) rates. Prices verified 11 October 2026 against Anthropic, OpenAI, Google and xAI. The same figures drive our API cost calculator and pricing tracker.
Two rows deserve attention. GPT-4.1 was the cheapest insurance policy on the list and is now a dead end — the only OpenAI model that could not surprise you at 300,000 tokens, deprecated on 14 October 2026 for removal on 14 January 2027 with GPT-6 Luna named as the replacement (what the 92-day notice means). OpenAI now has no uncapped long-context option at any price: every GPT-6 model carries the 272,000-token cliff. And Gemini 3.8 Flash’s no-cliff rate is promotional: on 1 January 2027 it reverts to $1.50 / $0.15 / $7.50, doubling. If you are picking it today for long prompts, price the January number.
The same document, priced on every vendor
Take a 600,000-character input — roughly 105,000 words — sent as a cached turn with 500 output tokens. The document is identical; only the tokenizer and the threshold differ.
| Model | Tokens | Tier | Cost per turn |
|---|---|---|---|
| DeepSeek Flash | 180,018 | standard | $0.00084 |
| GPT-6 Luna | 150,000 | standard | $0.00175 |
| DeepSeek V4 Pro | 180,018 | standard | $0.00495 |
| Claude Haiku 5.5 | 222,883 | over cliff | $0.01239 |
| Gemini 3.8 Flash | 150,000 | standard | $0.01313 |
| GPT-6.1 Sol | 150,000 | standard | $0.02000 |
| Claude Sonnet 5.5 | 222,883 | standard | $0.02729 |
| Grok 4.3 | 150,000 | standard | $0.03125 |
| Gemini 3.1 Pro | 150,000 | standard | $0.03600 |
| Claude Opus 5.5 | 222,883 | standard | $0.05458 |
One document, and it puts Haiku 5.5 over its cliff while leaving every other tiered model comfortably under. GPT-6 Luna and Haiku 5.5 have identical rate cards — $0.10 / $0.01 / $0.50 — and on this prompt Luna costs about one-seventh as much. Nothing about the models’ advertised prices predicts that; only the threshold does. Haiku-over-the-cliff barely still beats Gemini 3.8 Flash, and will lose to it in reverse once Google’s promotion ends and Haiku’s 5x stays put.
Push the same document to 1,000,000 characters (~175,000 words) and the ordering scrambles again: Grok 4.3 and Gemini 3.1 Pro cross their 200,000-token thresholds, taking Grok to $0.1025 and Gemini Pro to $0.1090 per turn, while GPT-6 Luna is the only tiered model still on standard rates at $0.00275. OpenAI’s 272,000-token threshold corresponds to about 1,088,000 characters, so the GPT-6 family stays cheap right up to the edge of a million-character prompt.
Above 100,000 tokens, the Claude decision inverts
This is the consequence buyers miss. Haiku 5.5’s surcharge applies to every rate; Sonnet 5.5 has no surcharge at all. So crossing the threshold does not just make Haiku more expensive — it closes the gap that was the entire reason to use Haiku.
| Haiku 5.5 vs Sonnet 5.5 | Cached input | Output |
|---|---|---|
| Prompt ≤ 100,000 tokens | Sonnet costs 10x | Sonnet costs 20x |
| Prompt > 100,000 tokens | Sonnet costs 2x | Sonnet costs 4x |
A tenfold saving justifies tolerating a smaller model. A twofold saving on cached input usually does not. If your prompts sit reliably above 100,000 Anthropic tokens, the honest recommendation is Sonnet 5.5 — and that is the opposite of what the headline rate cards imply. We covered the launch arithmetic when Haiku 5.5 shipped, and the same pattern appeared at OpenAI’s 272,000-token cliff on GPT-6 Astra and in xAI’s flat-looking Grok 4.7 rate card.
Three things the rate cards do not tell you
Cache hits count toward the threshold. Anthropic’s pricing page: “A request’s prompt length counts all of its input tokens, including cache reads and cache writes,” and “a request over the threshold pays the higher prices even when part of its prompt is a cache hit.” A 150,000-token prompt that is almost entirely a cache hit is still over the cliff, and the cached tokens bill at $0.05 rather than $0.01. Caching lowers the rate; it never moves the line.
Each request is priced on its own. Earlier requests “keep the prices they were billed at,” so one long turn in a conversation does not retroactively reprice the session — but it also means a single oversized turn is billed at 5x in full, with no averaging in your favour.
EU and US residency premiums stack on top. Anthropic’s US-only inference_geo parameter adds a 1.1x multiplier to all token categories, and the pricing page states that on Haiku 5.5 “the multiplier also applies to the higher prices for prompts over 100,000 tokens” — so a long prompt on US-pinned Haiku 5.5 is $0.55 / $0.055 / $2.75. Buying Claude through a cloud marketplace for residency reasons adds its own 10% on top; we mapped those routes in Bedrock vs Vertex and across vendors in the EU-resident API guide.
The trim lever, and why it beats every other optimisation
Because the surcharge reprices the entire request, the return on removing context is discontinuous near the threshold. On Haiku 5.5’s cached input:
| Trim | Context removed | Cached-input cost | Saving |
|---|---|---|---|
| 101,000 → 100,000 | 1% | $0.00505 → $0.00100 | 80% |
| 110,000 → 100,000 | 9% | $0.00550 → $0.00100 | 82% |
| 127,000 → 100,000 | 21% | $0.00635 → $0.00100 | 84% |
| 200,000 → 100,000 | 50% | $0.01000 → $0.00100 | 90% |
Dropping a single percent of context saves four-fifths of the bill, provided that percent is the one that crosses the line. Another way to read it: a 100,001-token prompt is billed exactly as a 500,005-token prompt would be at standard rates. You are paying for 400,000 tokens of phantom context.
The practical implication is a hard budget, not a soft one. Cap the assembled prompt at the threshold minus a margin for the tokenizer’s variance, measured with the vendor’s own token-counting endpoint rather than a character estimate, and drop the lowest-ranked retrieved chunks to fit. Above the threshold the lever stops working and ordinary proportionality returns.
Which to buy
| If your prompts are… | Buy | Why |
|---|---|---|
| Under ~40,000 words, cost-sensitive | Claude Haiku 5.5 or GPT-6 Luna | Identical $0.10 / $0.01 / $0.50; both safely under their cliffs |
| 40,000–190,000 words, cost-sensitive | GPT-6 Luna | Same rate card as Haiku, cliff four times further out |
| Any length, cheapest possible | DeepSeek Flash | No length cliff at all; manage the peak-hour 2x instead |
| Any length, frontier quality | Claude Sonnet 5.5 or Opus 5.5 | Full 1M window at standard rates, no cliff to model |
| Over 190,000 words on OpenAI | No good option | GPT-4.1 was the only uncapped OpenAI model and dies 14 Jan 2027; GPT-6 Luna is cheapest but bills 2x input above 272,000 tokens |
| Unpredictable length, needs a hard ceiling | A no-cliff model | A cliff turns a length bug into a 5x billing event |
If you are choosing between ecosystems rather than models, our Claude vs ChatGPT and DeepSeek V4 Pro vs Claude Opus 4.8 comparisons cover the non-price differences, and the coding-tool guide covers the subscription products built on top of these APIs. Model lifecycle dates are tracked on model retirements.
What changed
- 14 Oct 2026: GPT-4.1, GPT-4o and GPT-4o mini were deprecated for removal on 14 January 2027, taking OpenAI’s only uncapped long-context model with them — see analysis.
- 11 Oct 2026: All four thresholds and multipliers re-verified against vendor pricing pages; Grok 4.3 confirmed tiered at 200,000 tokens despite its single-rate presentation elsewhere in xAI’s docs.
- 8 Oct 2026: Claude Haiku 5.5 launched with the first Anthropic prompt-length tier — 100,000 tokens, 5x on every rate — see analysis.
- 30 Sep 2026: GPT-6.1 Sol arrived at $2 / $0.10 / $10 with the same 272,000-token cliff as the rest of the GPT-6 family — see analysis.
Frequently asked questions
Which models have no long-prompt surcharge at all?
On Anthropic, every Claude 4.6-and-later model except Haiku 5.5 — the pricing page states that they '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.' On Google, every Flash and Flash-Lite model; only Gemini 3.1 Pro and 2.5 Pro are tiered. On OpenAI, the pre-GPT-5.4 models, including GPT-4.1 at a flat $2 / $0.50 cached / $8 across its full 1,047,576-token window - but GPT-4.1 was deprecated on 14 October 2026 and is removed from the API on 14 January 2027, so it is no longer somewhere to build. DeepSeek charges no length surcharge on any model, and neither does Mistral Large 4. On xAI there is no escape: Grok 4.7, Grok 4.3 and grok-build-0.1 are all tiered at 200,000 tokens. Prices verified 11 October 2026.
Why is Anthropic's 100,000-token threshold worse than it looks?
Because it is counted in Anthropic's newer tokenizer, which Anthropic says 'produces approximately 30% more tokens for the same text' than the one used by Claude 4.6 and earlier. At the glossary figure of about 2.69 characters per token, 100,000 tokens is roughly 269,200 characters — about 47,000 English words. OpenAI's 272,000-token threshold, at its own documented four-characters-per-token estimate, is about 1,088,000 characters, or roughly 191,000 words. So on the same document OpenAI's cliff arrives four times later than Anthropic's, not 2.7 times later as the raw token counts suggest. In Haiku-4.5-equivalent tokens, Anthropic's cliff lands at about 77,000.
Does prompt caching keep me under the threshold?
No, and this is the most expensive misreading of the Haiku 5.5 rate card. Anthropic's pricing page is explicit: 'A request's prompt length counts all of its input tokens, including cache reads and cache writes,' and 'a request over the threshold pays the higher prices even when part of its prompt is a cache hit.' A 150,000-token prompt that is 95% cache hit is still a 150,000-token prompt for threshold purposes, and every token in it — cached or not — bills at the higher rate. Caching reduces the rate you pay; it does not move the cliff.
Above 100,000 tokens, is Claude Haiku 5.5 still the cheap option?
It is still cheaper per token, but by so little that the reason to choose it mostly disappears. Below the threshold, Sonnet 5.5 costs 10x Haiku 5.5 on cached input ($0.10 against $0.01) and 20x on output ($10 against $0.50). Above it, Haiku's rates quintuple while Sonnet's do not move at all, so the same two gaps collapse to 2x and 4x. If your workload is reliably over 100,000 tokens, the question stops being 'can I afford Sonnet' and becomes 'why am I accepting Haiku quality for a 2x cached-input saving.'
What is the cheapest way to cut a long-prompt bill?
Get back under the threshold. The surcharge applies to the whole request, so the saving from trimming is wildly non-linear: on Haiku 5.5, cutting a 127,000-token cached prompt to 100,000 tokens — 21% less context — cuts that turn's cached-input cost by 84%. Trimming from 101,000 to 100,000 tokens, a 1% reduction, still saves 80%. No other cost lever on any of these platforms pays that well. Below the threshold the arithmetic is ordinary again: 10% less context is 10% less money.
Does DeepSeek really have no long-prompt surcharge?
Correct, but it charges on a different axis instead. DeepSeek Flash and V4 Pro both carry a 1,000,000-token window at one flat rate regardless of prompt length. What DeepSeek varies is the clock: peak hours — 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday — cost twice the off-peak rate. So a long-prompt workload on DeepSeek is priced by when you run it, not by how much you send, which makes it the only major vendor where batching overnight is a bigger lever than trimming context.