---

Abstract #

Providers now advertise context windows of one million tokens as a headline feature, but their price lists treat that capability in very different ways. We ask whether long context is priced as a distinct feature or bundled into ordinary per-token rates. We built a dataset of 83 offerings from 14 model developers, sold first-party or through three hosts, read directly from public pricing pages on 20 October 2026, and combined it with a 365-model market snapshot that carries a public capability index. We then computed the cost of single requests from 32K to 990K prompt tokens, including tier surcharges, caching and batch discounts. Three results stand out. First, once capability is controlled for, a larger window does not raise the base price per token: the elasticity is -0.231 (95% CI -0.546 to 0.083) across 213 market listings and -0.244 within providers. The premium is instead charged through request-level tiers, used by 32.7% of first-party models with windows of one million tokens or more. Second, crossing a tier threshold raises the cost of a request by a median of 98.2% for a single extra token, and the surcharge on a 990K request is a median 99.5%. Third, caching at 90% cuts the surcharge in dollars by a median 80.5% but leaves it at 97.5% in relative terms. Measured against effective rather than advertised context, a full window costs a median 314 times more per reliably used token than its list price suggests.

1. Introduction #

Our previous article, Inference Economics: The Hidden Cost Crisis Behind Falling Token Prices[1]^[2], argued that cheaper tokens do not mean cheaper AI, because volume and token depth grow faster than unit prices fall. Context length is the most direct form of token depth. Retrieval, document review and long sessions push prompts towards the window limit, which in our October 2026 sample is one million tokens or more for 49 of 81 priced offerings.

The headline per-token rate has become a poor guide to what a long request costs. A systematic study of OpenRouter and Epoch AI price records documents a roughly 600-fold fall in token prices since 2020 and a market that now sorts itself into economy, mid and flagship tiers[2]^[3]. That work treats the token as the unit of value. A recent mechanism-design paper notes that pricing theory usually treats tokens as a homogeneous commodity and relaxes that assumption, deriving the cost of inference from a compute-bound prefill phase and a bandwidth-bound decode phase[3]^[4]. Long prompts load the prefill phase and the memory that holds the key-value (KV) cache, which is exactly where serving cost rises with length.

Two questions follow for anyone buying inference. Is the capability to accept a long prompt reflected in the list price of a model, so that one pays for the window whether or not it is used? Or is it charged only when it is used, through surcharges on long requests? No public source answers this with current data: price pages present tiers as footnotes, and market studies average them away.

RQ1: After controlling for model capability, does a larger maximum context window raise the list price per input token, and how common are explicit long-context surcharges? RQ2: What does a single long request cost at 32K, 128K, 200K, 500K and 990K prompt tokens, and where do step functions appear? RQ3: How much of the long-context premium survives prompt caching and batch discounts, and what does a full window cost per token that the model reliably uses?

The answers matter for the AI Economics series because they move the unit of analysis from the price of a token to the price of a request, which is the quantity that budgets, routers and caches actually act on.

2. Existing Approaches (2025–2026 state of the art) #

2.1 Hedonic pricing of model features #

The classical way to ask whether a product attribute carries a price is a hedonic regression: regress price on measured characteristics and read the implicit price of each one[4]^[5]. Applied to language models, the characteristics are capability, context window, modality and speed. Du’s analysis of 318 OpenRouter models fits price trends by tier and reports a reasoning premium averaging 31.5 times non-reasoning prices, but it does not isolate the window as a separate attribute[2]^[3]. Where hedonic analysis fails is collinearity: models with larger windows are also newer and stronger, so a raw price-on-window slope mixes the two. Our RQ1 addresses that directly with a public capability control.

2.2 Request-level tiering #

Five providers in our sample charge by prompt length at the level of the request. OpenAI bills prompts above 272K input tokens at twice the input rate and 1.5 times the output rate for the whole request[5]^[6]. Google bills Gemini 3.1 Pro Preview and Gemini 2.5 Pro at a higher rate when the prompt exceeds 200K tokens, and the cached-token price doubles with it[6]. xAI applies its higher rate when the prompt reaches 200K tokens, so a request of exactly 200,000 tokens is already a long one[7]. Alibaba Cloud’s Model Studio uses up to three tiers, billing all tokens in a request at the unit price of its tier[8]. Anthropic offers the full one-million-token window at standard rates on its 4.6 and later models, with one exception: Claude Haiku 5.5 charges five times more for prompts over 100,000 tokens, including the cached part of the prompt[9]^[7]. Tiering prices the window only when it is used, but it creates cliffs, which RQ2 measures.

2.3 Flat per-token pricing and host arbitrage #

The alternative is a flat rate over the whole window. Anthropic’s 4.6-and-later models other than Haiku 5.5, Google’s newer Flash and Argon models, DeepSeek, Mistral and Moonshot all publish a single rate per token regardless of prompt length in our snapshot. Third-party hosts add a twist. Together AI lists Qwen3.7 Plus at a flat $0.32 per million input tokens over a one-million-token window[10], while Alibaba’s own page tiers the same model above 256K tokens[8]. A measurement study of hosted open-weight APIs argues that the operational unit is a provider-specific service, not the model, and that listed prices are more anchored than context capacity or latency[11]^[8].

2.4 Caching and batching as the main discount levers #

Prompt caching bills a repeated prefix at a fraction of the input rate. An audit of real APIs detected caching from response timing and found global cache sharing across users at seven providers, a reminder that caching is an engineering mechanism with privacy consequences, not only a discount[12]^[9]. Across OpenAI, Anthropic and Google, prompt caching reduced API cost by 41–80% in long multi-turn sessions with 10,000-token system prompts[13]^[10]. A study of Anthropic’s cache found a two-tier behaviour with hit rates around 0.83 below a 3,500-token prefix and showed that query-specific prompt compression breaks prefix caching[14]^[11].

2.5 Why long context costs more to serve #

The engineering literature explains why providers want a length signal in the price. Moonshot’s Mooncake serving platform separates prefill and decode clusters and builds a disaggregated KV cache, and it reports its largest gains in long-context scenarios[15]^[12]. Strata reports that the KV cache of long contexts quickly exceeds GPU memory, which forces hierarchical caching and leaves systems bound by data loading rather than compute[16]^[13]. A 2026 benchmark of KV-cache compression finds that the compression ratio alone is a poor predictor of end-to-end throughput for long-context serving[17]^[14]. Architecture choices matter too: cost-optimal grouped-query attention can cut memory use and FLOPs for long contexts by more than 50% relative to Llama-3’s configuration[18]^[15]. Together they imply that marginal serving cost per token rises with prompt length, the economic case for tiers.

flowchart TD
    A[Long-context pricing 2026] --> B[Flat rate over full window]
    A --> C[Request-level tier]
    A --> D[Multi-tier ladder]
    A --> E[Third-party host rate]
    B --> B1[Anthropic 4.6 and later, DeepSeek, Mistral, Moonshot]
    C --> C1[OpenAI above 272K]
    C --> C2[Google Pro models above 200K]
    C --> C3[xAI at 200K and above]
    C --> C4[Claude Haiku 5.5 above 100K]
    D --> D1[Qwen3.7 Flash: 32K, 256K]
    E --> E1[Same model, flat price, different host]

3. Method and Data #

Provider-direct dataset. We read the public pricing pages and model documentation of Anthropic, OpenAI, Google (Gemini API), xAI, Alibaba Cloud Model Studio, DeepSeek, Mistral, Moonshot, Amazon Bedrock (Amazon Nova and hosted Meta Llama models), Cohere, Together AI and Groq on 20 October 2026. For each offering we recorded the maximum context window, input, output and cached-input prices per million tokens, every long-context tier with its threshold, and the batch discount. Every row carries its source URL and access date. Where a page did not publish a value, the field is empty: Cohere’s current Command A model has no public per-token price, Groq lists Llama 3.3 70B as “contact sales”, and 15 priced rows have no published cache-hit price. The result is 83 rows, 81 with prices. Bedrock prices were resolved from the price map that the AWS page loads, published 16 October 2026, for US East.

Market cross-section. To test RQ1 with more statistical power, we used a snapshot of the OpenRouter models API taken the same day[19]. After removing free and batch duplicates and non-text models, 365 listings remain. The API carries the Artificial Analysis Intelligence Index for 213 of them; version 4.3.2 of that index combines ten evaluations across reasoning, knowledge, maths and programming, one of which (AA-LCR) tests long-context reasoning[20]. We use it as the capability control.

Effective context. For RQ3 we took effective context lengths from NoLiMa, an ICML 2025 benchmark that removes literal overlap between question and evidence, defining the effective length as the longest context at which a model keeps at least 85% of its short-context score[21]^[16]. We matched ten NoLiMa models to current list prices. We also used RULER’s effective length for Llama 3.1 70B as a contrast[22]^[17].

Computations. The cost of a request is the prompt length times the input rate of the tier the request falls in, plus 2,000 output tokens at the matching output rate. A request must fit the window including its output, so the one-million-class point is 990K prompt tokens. The surcharge is the difference between the tiered cost and the cost of the same tokens at short-tier rates. The cached scenario bills 90% of the prompt at the cache-hit rate in steady state and ignores cache writes and storage fees; cache and batch discounts are combined only where both pages state that they stack (Anthropic and OpenAI). The regressions are ordinary least squares of log10 input price on log10 window size with HC1 robust standard errors, with and without the capability index and provider fixed effects.

All code, data and outputs are in stabilarity/hub/research/the-context-window-premium-why-long-context-models-command-higher-prices-than-token-count-alone-explains.

flowchart LR
    P[Pricing pages, 83 rows] --> R[Tier-aware cost model]
    O[OpenRouter snapshot, 365 listings] --> G[Price regressions]
    I[Capability index, 213 listings] --> G
    N[NoLiMa effective lengths] --> U[Cost per usable token]
    R --> Q2[RQ2 request costs and cliffs]
    R --> Q3[RQ3 cache and batch scenarios]
    G --> Q1[RQ1 context elasticity]
    P --> U

4. Results #

4.1 RQ1: The window is not priced into the base rate, the tier is #

Across all 365 listings the elasticity of input price with respect to window size is 0.226 (95% CI 0.09 to 0.362), and in the 213 listings with a capability score it is 0.286, which means a ten-fold larger window goes with 1.93 times the price. The fit is weak (R² 0.031 for all listings), and Chart 1 shows why: at every common window size, prices span two to four orders of magnitude.

Adding the capability index reverses the sign. With capability held constant, the window elasticity is -0.231 (95% CI -0.546 to 0.083), a price multiple of 0.59 per ten-fold window, and not distinguishable from zero. Capability carries the price instead: each index point is worth a factor of 1.057 in input price, and R² rises to 0.158. Output prices tell the same story, with a window elasticity of -0.125 (95% CI -0.46 to 0.21). Within our provider-direct data the result holds. Among 67 first-party rows, the raw elasticity is 0.343 with a confidence interval spanning zero; with provider fixed effects it rises to 0.585 (95% CI 0.027 to 1.143), consistent with each provider’s newest and strongest models also having the largest windows. Once the capability index is added for the 49 rows where it is available, the elasticity falls to -0.244 (95% CI -0.898 to 0.409).

Direct comparisons inside one product line confirm the regression. Claude Sonnet 4.5 with a 200K window and Claude Sonnet 4.6 with a 1M window carry identical prices of $3 input and $15 output per million tokens[9]^[7]. OpenAI’s GPT-4.1 offers a 1,047,576-token window for $2 per million input tokens, the same rate as o3 with 200,000 tokens[5]^[6].

The premium exists, but it sits in the tiers. In the provider-direct data, 19 of 81 priced offerings (23.5%) have a long-context surcharge, rising to 16 of 49 (32.7%) among models with windows of at least one million tokens. The market snapshot agrees: 14.2% of all listings and 31.8% of one-million-token listings carry a tier. The median tier multiplies the input rate by 2.0 (range 2.0 to 5.0) and the output rate by 1.5. The thresholds cluster at five values: 32K, 100K, 200K, 256K and 272K tokens.

4.2 RQ2: Step functions turn one extra token into a doubled bill #

Table 1 shows what a single request costs at five prompt lengths, each with 2,000 output tokens.

Model (provider)	Window (tokens)	32K (USD)	128K (USD)	200K (USD)	500K (USD)	990K (USD)
Claude Sonnet 5.5 (Anthropic)	1,000,000	0.084	0.276	0.42	1.02	2.0
Claude Haiku 5.5 (Anthropic)	1,000,000	0.0042	0.069	0.105	0.255	0.5
gpt-6.1-sol (OpenAI)	1,050,000	0.084	0.276	0.42	2.03	3.99
gpt-5.4 (OpenAI)	1,050,000	0.11	0.35	0.53	2.545	4.995
gemini-4-argon (Google)	1,048,576	0.084	0.276	0.42	1.02	2.0
gemini-3.1-pro-preview (Google)	1,048,576	0.088	0.28	0.424	2.036	3.996
grok-4.7 (xAI)	500,000	0.076	0.268	0.824	n/a	n/a
qwen3.7-plus (Alibaba)	1,000,000	0.016	0.0544	0.0832	0.6096	1.1976
Qwen3.7 Plus (Together)	1,000,000	0.0128	0.0435	0.0666	0.1626	0.3194
deepseek-v4-pro (DeepSeek, peak)	1,000,000	0.0502	0.1769	0.2719	0.6679	1.3147
Mistral Large 4 (Mistral, sale price)	1,000,000	0.0259	0.0912	0.1402	0.3442	0.6774
kimi-k3 (Moonshot)	1,048,576	0.126	0.414	0.63	1.53	3.0
Nova 2 Lite (Amazon)	1,000,000	0.0146	0.0434	0.065	0.155	0.302

Source: provider pricing pages accessed 2026-10-20[9]^[7][5]^[6][6][7][8][10]; DeepSeek, Mistral, Moonshot and AWS rows from the pages listed in data/pricing.csv; our calculation. n/a: the request does not fit the window.

Three models with identical short-context prices show how the tier changes the ranking. Claude Sonnet 5.5, gpt-6.1-sol and Gemini 4 Argon all cost $0.084 for a 32K request. At 990K tokens Sonnet 5.5 and Argon still cost $2.00, while gpt-6.1-sol costs $3.99 and Gemini 3.1 Pro Preview $3.996. The threshold convention matters at the boundary: at exactly 200,000 tokens grok-4.7 already charges its long rate and costs $0.824, while Gemini 3.1 Pro still charges its short rate and costs $0.424.

Chart 2 makes the cliffs visible. Because every tier re-prices the whole request, adding one token at the threshold raises the bill by a median of 98.2% across the 19 tiered offerings, with a range of 96.3% to 400%. Claude Haiku 5.5 is the extreme case: a 100,000-token request costs $0.011 and a 100,001-token request $0.055. We can express the cliff as an equivalent prompt length. A gpt-5.4 request of 272,001 tokens costs as much as 550,002 tokens would at the short rate; for Gemini 3.1 Pro the 200,001-token request equals 406,002 short-rate tokens; for qwen3.7-plus the 256,001-token request equals 784,003.

Where a task can be split, the cliff becomes an arbitrage. Two requests just below the threshold cost a median 48.9% less than one request at twice the threshold, rising to 66.2% for qwen3.7-plus and 79.0% for Claude Haiku 5.5. The surcharge on a full 990K request is a median 99.5% over short-tier rates, 200% for qwen3.7-plus and 566.2% for qwen3.7-flash, whose three-tier ladder starts at 32K tokens. Host choice is a lever of the same size: the same Qwen3.7 Plus model costs $1.1976 for a 990K request on Alibaba’s tiered schedule and $0.3194 on Together’s flat rate.

4.3 RQ3: Discounts shrink the surcharge in dollars, not in proportion #

Every tiered schedule in our sample re-prices cached tokens with the tier as well, so caching and batching scale the short and long sides of the bill together.

For a 990K request the median surcharge is $1.99 at standard rates. With 90% of the prompt served from cache, it falls to a median $0.307, a reduction of 80.5% in dollars across the 15 tiered models with a published cache price. In relative terms the surcharge barely moves, from a median 99.5% to 97.5%; for Claude Haiku 5.5 it stays at 400%, because the page states that cached reads count towards the threshold and are billed at the higher rate[9]^[7]. Batch processing cuts the dollar surcharge by the batch discount (50% for most models, 20% for grok-4.3) for the 14 models where a long-context batch price exists, and leaves the median percentage at 99.5%. OpenAI publishes no long-context batch price for gpt-5.5-pro, and qwen3.7-plus lists no batch discount. Discounts make long requests affordable; only staying below the threshold removes the tier.

Caching also changes the comparison with retrieval. A cost model for persistent memory shows that long-context inference incurs a per-turn charge that grows with context length even under caching, so a fact-based memory system becomes cheaper after about ten turns at 100K tokens[23]^[18]. Our tier data adds a second reason: under a tiered schedule the per-turn charge also jumps once the history crosses the threshold.

The last question is what the window is worth when it is used. The cost of filling a window, divided by the tokens a model reliably uses, is far above the list price.

Across the ten NoLiMa models with current prices, the median ratio of claimed to effective context is 314, with a range of 16 to 10,000. GPT-4.1 claims one million tokens and holds 85% of its baseline up to 16K, so its $2.00 list price becomes $125 per million reliably used tokens when the window is filled. GPT-4o moves from $2.50 to $40, Gemini 2.5 Flash from $0.30 to $150, and Llama 4 Scout, with a claimed ten million tokens and an effective length of 1K, from $0.17 to $1,700. For three models NoLiMa reports an effective length below 1K, so their ratios are lower bounds. The benchmark choice matters: RULER puts Llama 3.1 70B’s effective length at 64K, a ratio of 2.0, while NoLiMa’s harder, non-literal task puts it at 2K, a ratio of 64.

5. Discussion #

Implications for practitioners. Compare providers at the prompt length you actually send, not at the headline rate. For workloads that stay under 100K tokens, almost every tier is invisible (Qwen3.7 Flash re-prices at 32K) and the cheapest base rate wins. For document-scale prompts, a flat-rate model can cost half as much as a tiered model with the same short-context price, as the Sonnet 5.5 and gpt-6.1-sol comparison shows. Build the threshold into prompt assembly: a retrieval system that caps context at 200K tokens on Gemini 3.1 Pro, 199,999 on xAI or 272K on OpenAI’s tiered models avoids a cliff worth roughly a doubled bill. Where tasks decompose, splitting long inputs saves about half the cost before any quality effect is considered.

The usable-context data argue for retrieval over brute-force context in most question-answering work. Long-context models outperform retrieval when resourced sufficiently, but retrieval is far cheaper, and routing between the two maintains comparable performance at significantly lower computation cost[24]^[19]. LaRA’s comparison finds no single winner: the right choice depends on model size, context length, task type and chunk characteristics[25]^[20]. Adding more retrieved passages first helps and then hurts because of hard negatives[26]^[21], and an order-preserving retrieval design can beat a full long context with far fewer tokens[27]^[22]. Performance also degrades with input length even when retrieval is perfect, by 13.9% to 85% within claimed lengths[28]^[23]. These findings match the earlier hub analysis of needle-in-a-haystack and its successors[29]^[24], which showed that synthetic retrieval tests overstate usable context, and the long-standing evidence that models use the middle of long inputs poorly[30]^[25].

flowchart TD
    S[Prompt length known] --> T{Below every tier threshold?}
    T -->|Yes| A[Choose on base rate and capability]
    T -->|No| U{Task can be split?}
    U -->|Yes| V[Split below threshold, about half the cost]
    U -->|No| W{Prefix repeats across requests?}
    W -->|Yes| X[Cache it, surcharge falls in dollars only]
    W -->|No| Y{Needs whole-document reasoning?}
    Y -->|No| Z[Retrieve, keep context short]
    Y -->|Yes| F[Prefer a flat-rate long-context model]

Why providers tier. The serving literature gives a cost rationale: KV memory and prefill compute grow with prompt length[15]^[12][16]^[13]. The mechanism-design view suggests a second rationale: discrete service tiers let a seller screen buyers by their time preferences, and two-tier pricing captures 26–66% more profit than the best single tier in the authors’ calibration[3]^[4]. Our data cannot separate the two. Google prices its newest Argon model flat while keeping tiers on Gemini 3.1 Pro and 2.5 Pro, and Anthropic applies a tier only to its smallest current model, which suggests that tiers are a commercial choice as much as a cost pass-through.

Limitations and threats to validity. The snapshot covers one day; several prices are promotional with announced increases (Gemini 4 Argon rises to $4 input on 1 December 2026, the Gemini 3.x Flash models in January 2027) and Mistral Large 4 was on sale. List prices ignore negotiated discounts, regional uplifts, priority and flex tiers, DeepSeek’s off-peak halving and cache storage fees. The capability index covers 213 of 365 listings and 49 of 67 first-party rows, and it includes a long-context reasoning evaluation, so it may absorb part of the window’s value; a capability score without long-context items could restore a small positive elasticity. Effective-context data exist only for an older cohort, so the usable-token figures describe GPT-4.1, Llama 4 and Gemini 2.5, not the 2026 flagships, and synthetic benchmarks differ from realistic long-context tasks[31]^[26][32]^[27]. Token counts themselves are reported by providers: per-token billing creates an incentive to misreport[33]^[28], and hidden reasoning tokens can be inflated without detection under current audit frameworks[34]^[29].

What would change the conclusion. A one-million-token model priced above an otherwise identical 200K model would contradict RQ1. A switch between flat and tiered schedules would change RQ2 within one pricing cycle, and an effective-context benchmark for the current flagships could narrow the usable-token gap sharply.

6. Conclusion #

RQ1 Finding: Long context is not priced into the base per-token rate; it is charged through request-level tiers. Measured by the capability-controlled elasticity of input price to window size = -0.231 (95% CI -0.546 to 0.083, n = 213), and by the share of one-million-token first-party models with a surcharge = 32.7%. This matters for the series because the cost of AI now depends on how a request is shaped, not only on which model serves it.

RQ2 Finding: Tiers create cliffs that re-price the whole request. Measured by the median cost increase from one extra token at the threshold = 98.2% (range 96.3% to 400%) and the median surcharge on a 990K request = 99.5%. This matters for the series because prompt-length discipline is now a first-order cost lever, comparable to model choice.

RQ3 Finding: Caching and batching shrink the surcharge in dollars but not in proportion, and advertised windows overstate usable context by orders of magnitude. Measured by the median dollar reduction of the surcharge under 90% caching = 80.5%, the median relative surcharge under caching = 97.5%, and the median claimed-to-effective context ratio = 314. This matters for the series because the hidden cost of long context lies in paying for tokens the model does not reliably use.

The next question for the series is how routing between retrieval and long context should be priced when both the tier schedule and the effective context are known in advance.