Large Models and the Cost per Million Tokens

Every time a new AI model is released, the comment sections erupt into debates over usage costs. "Chinese models are an order of magnitude cheaper," "Anthropic's pricing is predatory," "Subscriptions are better value than APIs," and so on. These debates usually lead nowhere; the disagreement stems not just from personal preference, but from varying use cases based on individual experience. People often compare apples to oranges: prices are published in different currencies, some rates are promotional and set to expire within a month, third-party aggregators can charge significantly more (or less) than the original developers, and benchmark scores vary depending on who ran the test, which hardware they used, and what methodology was applied.

Working with models has become everyday business, yet these debates have zero impact on actual monthly invoices — you still have to calculate your own total cost of ownership (TCO).

For example, a Wall Street Journal article examines a lawsuit filed by Washington resident Carl Kahn against Anthropic. According to Kahn, the actual usage limits for "Max" plans are significantly lower — by factors of five and twenty — than what is advertised on their website. He claims that a single five-hour session with Claude Code consumed 15% of his weekly limit. The lawsuit seeks class-action status on behalf of all users who purchased these plans since April 2025; the outcome remains uncertain.

The same article highlights the opposite extreme: an American company reportedly spent $500 million on Claude tokens in a single month simply because they failed to implement usage limits for their employees. It is a cautionary tale that is amusing to everyone except the participants. Neither the company nor Kahn knew what access would actually cost them until the bills arrived.

Rather than relying on third-party summary tables, we have compiled the data for this article directly from official provider pages, technical reports, and documentation. We analyzed over a dozen models across three different platforms, identifying base rates, cache pricing, long-context premiums, and benchmark scores. We also cross-referenced these with our own internal statistics for hosting open-weights models, which, as it turns out, often differ significantly from public leaderboards.

Why Comparing Prices is Harder Than It Was a Year Ago

A year ago, pricing discussions usually revolved around three or four model families. Today, there are dozens of families, each containing several tiers that vary in cost by multiples. The scale of this expansion is best illustrated by Alibaba; their Model Studio price list contained 194 items in the text generation section alone at the time of our check, including third-party models. The sidebar for that same section listed 222 units (summing 215 Qwen models and seven third-party ones), meaning even their own page is internally inconsistent.

Anthropic publishes rates in USD, OpenAI's European site shows prices in EUR, and Alibaba’s price list toggles between regions — meaning rates for mainland China, the US, and the EU differ. To compare them, everything must be converted to a single currency at current exchange rates. The Alibaba data used below is from their international section.

For instance, Sonnet 5 currently starts at $2 per million input tokens and $10 per million output tokens. However, this is a promotional rate valid until August 31, 2026; after that, rates will rise to $3 and $15 respectively — a 1.5x price hike for the API. Notably, this information is tucked away in a footnote below the table rather than in the primary pricing row.

Claude prices are from Anthropic; DeepSeek prices are from DeepSeek's own documentation; all others are from the Alibaba Model Studio price list.

The input cost difference between qwen-turbo and Fable 5 is a factor of 200, making direct comparison meaningless as they serve different use cases. The lower bound actually drops to zero for open models like GLM-4.7-Flash and GLM-4.5-Flash, which are released entirely for free along with their weights. This scale of difference explains why Chinese models dominate discussions about cost savings. However, whether that cheapness translates to better results is a question we will address below.

One Model, Two Different Prices

The DeepSeek rows in the table above are intentionally duplicated because the same model is sold in two places at different prices. Through Alibaba Model Studio, deepseek-v4-pro costs $2.40 per million input tokens, whereas in DeepSeek’s own pricing, the same model costs $0.435 (cache miss) — a 5.5x difference. A similar multiplier applies to output: an aggregator charges $4.80 compared to $0.87 directly. For Flash, the gap is smaller, at approximately 40%.

The consistent markup only exists on these rows; cache pricing behaves differently for every model. For Flash via an aggregator, a cache hit costs $0.04 per million tokens versus $0.003 directly — a 14x difference compared to the 40% gap on standard input. For Pro, the same operation costs $0.20 instead of $0.33; here, the gap reaches 55x despite only a 5x difference in standard input. Anyone choosing a platform based solely on token price while hoping for massive cache savings will find that the second half of the calculation works very differently from the first.

A similar pattern occurs with glm-5.2. The list prices match almost exactly between platforms at $1.40 (input) and $4.40 (output), but they diverge on cache hits: $0.26 at Z.ai versus $0.28 via Alibaba.

The reason for this markup is simple: platforms like Alibaba provide a single interface and unified billing for models from various developers without requiring separate contracts with each one, all while the aggregator takes its cut. The practical takeaway here concerns the comparison process itself: before comparing prices, you must verify whose price list you are looking at. Reviews that list DeepSeek as a single line item without specifying the provider can be off by orders of magnitude. In fact, in our first draft of this article, we accidentally used aggregator figures for DeepSeek.

Cache Changes the Math More Than Model Selection

The discount for cache hits on DeepSeek is so massive that it renders the price difference between adjacent models in the lineup almost negligible. A million input tokens in a cache hit costs about $0.003 for Flash and slightly more than $0.003 for Pro, whereas those same tokens cost $0.14 and $0.435 respectively if they miss the cache — meaning a 50x to 120x difference. Such gains only occur when your prompt contains significant repetitive data, such as long system instructions or large chunks of documentation passed in every call.

DeepSeek's mechanism is quite transparent: disk caching is enabled by default for everyone; no specific request parameter is required, and there is no charge for writing to the cache, meaning caching only works in your favor. However, you shouldn't rely on it as a guaranteed cost-saving measure; the documentation describes the system as "best effort" and does not promise 100% hit rates. Additionally, caching only reduces input costs; output tokens are re-billed for every call.

Due to the "sliding window attention" mechanism, a cached fragment is treated as an independent, complete unit; a cache hit only occurs if there is an exact match. Simple overlaps are not enough. Fragments are preserved at request boundaries (e.g., at the end of user input or model response), when common prefixes are detected across multiple calls, and at fixed token intervals.

This approach prevents situations where a long prompt fails to cache entirely because it didn't reach a specific boundary. You should also keep "TTL" in mind: unused cache is automatically cleared, typically within a few hours to several days, meaning infrequent workloads will pay full price every time. To track your actual hit rate, monitor the prompt_cache_hit_tokens and prompt_cache_miss_tokens fields in the API response.

Anthropic implements similar logic but with modest proportions. Reading from the cache at Anthropic costs one-tenth of a standard input ($1.00 per million for Fable 5), while writing to the cache is 25% more expensive than standard input ($12.50). Because of this write cost, you only see a benefit starting from the second call to the same content; this rule applies across the entire lineup from Haiku 4.5 to Fable 5.

The depth of discounts varies wildly among other developers. For kimi-k2.7-code, a cache hit costs $0.19 compared to $0.95 for standard input; for glm-5.2, it is $0.26 versus $1.40 at Z.ai — a 5x saving in both cases. Currently, Z.ai has announced that cache storage is "temporarily free" across its lineup, though they will likely charge for it eventually. For qwen3.6-max-preview, the ratio almost exactly mirrors Anthropic's logic: reading costs $0.13 when input is $1.30, while writing costs $1.625.

What the Documentation Shows That Price Lists Don't

Price lists usually state cost per million tokens, which leads readers to assume a flat rate for all cases. However, some Alibaba models increase their rates as the prompt length grows — a detail often omitted in brief reviews. For qwen3-coder-30b-a3b-instruct, there are four pricing tiers: for input up to 32k tokens, that "million" costs $0.45; then it jumps to $0.75, $1.20, and finally $2.40 for inputs between 256k and 1M tokens. The top tier is over five times more expensive than the bottom. Output prices escalate even faster, jumping from $2.25 to $14.40 — a sixfold increase.

For qwen3.6-max-preview, the rates mentioned above ($1.30 input / $7.80 output) apply up to 128k tokens; for the 128k–256k range, they rise to $2.12 on both input and output — a roughly 1.5x increase. qwen3-coder-480b-a35b-instruct has even steeper scaling: input doubles from $4.50 to $9.00 at the final tier, while output jumps fourfold from $22.50 to $90.00 in a single step. The output-to-input ratio stays at 5x for the first three tiers but hits 10x on the last one, meaning long-context workloads with verbose responses become much more expensive than the base price suggests.

DeepSeek does not care about prompt length; both V4 models handle up to a 1M token context and 384k token output at a flat rate, which is a significant advantage for long-context tasks. However, both support "reasoning mode" by default unless explicitly disabled. This can lead the model to generate thousands of invisible reasoning tokens before its short answer; all these tokens are billed at the output rate.

Alibaba's reasoning mode impacts costs in two ways: it increases the number of output tokens and changes the per-token rate, while input remains the same. The table below shows the multiplier applied to different models when enabling or disabling reasoning mode.

While most models in the table show a 3x multiplier, qwen-turbo is only 2.5x. Furthermore, reasoning mode has a compounding effect: because you get more tokens and each token is more expensive, the total cost discrepancy is much larger than the simple multiplier suggests. There is no universal rule; most models fall between 1x and 3x. For qwen3-32b, both rows are identical ($0.16 input / $0.64 output) regardless of mode, whereas for qwen3-235b-a22b (index 2507), the multiplier is 2.5x instead of 3x. The model size doesn't explain this: qwen3-32b sits right between the two qwen3-14b/235b models that have a 3x multiplier.

DeepSeek also differs in its rate limits, where Flash allows 2,500 concurrent requests compared to 500 for Pro. For interactive workloads, hitting these rate limits can become a bottleneck before the budget even becomes an issue.

Context window and output speed are other parameters missing from most price lists. On the Qwen page, qwen3.6-max-preview is listed with a 262k token window, while its successor Qwen3.7-Max has a 1M window — matching most of its competitors. According to Artificial Analysis, the model outputs about 50 tokens per second (ranking 119th out of 186). For chat or code editors, this latency is often more noticeable than a few cents' difference in token pricing.

The names deepseek-chat and deepseek-reasoner correspond to the standard and reasoning modes for deepseek-v4-flash, which were deprecated on July 24, 2026. DeepSeek reserves the right to change prices at any time and recommends topping up your account based on actual usage by checking their pricing page regularly.

Benchmarks: Different Testbeds, Different Scores

Benchmark scores are often reported as a single number, implying they are inherent properties of the model. In reality, these numbers depend entirely on the specific hardware, test sets, and configurations used during evaluation. This becomes obvious when looking at the same model across different benchmarks.

SWE-bench Pro consists of 731 tasks from real-world repositories; a "pass" is only recorded if new tests pass and existing ones don't fail. Scale's measurement shows Claude Opus 4.6 solving 51.9% of tasks, whereas Moonshot claims a score of 53.4 for the same model in its Kimi-K2.6 card. DeepSeek's technical report claims 57.4% — a 4.5 to 5.5 point lead over Scale. Since Scale reports a confidence interval of ±3.6 points, Moonshot’s score falls within the margin of error (meaning half the gap is just noise), but DeepSeek's result sits entirely outside that range.

Reasoning modes don't explain the rest of the gap; both Moonshot and DeepSeek claim their measurements are "maximum possible" scores, yet they still differ by 3.9 points. Scale marks its row with an asterisk, indicating it used mini-swe-agent instead of the standard SWE-Agent methodology.

The same occurs with other developers: DeepSeek evaluates V4-Pro in Max mode at 55.5%, but Alibaba's comparative table (published for the Qwen3.7-Max launch) places that same model at 59% — a 3.5 point difference. This isn't necessarily an error; it simply means the test conditions were different.

DeepSeek describes its test configuration and methodology in detail: a custom internal tool with two agents, a shell, and file editing capabilities, capped at 50 interaction steps and a 512k token context. Scale uses its own wrapper released on GitHub alongside the official GitHub task set; this provides a more representative view of all models but is limited by its dataset (for example, the latest Claude models weren't available when the data was collected).

Even these comparisons are relative, as noted in the footnotes. Some models were tested with cost limits and 50-iteration caps, while others had unlimited budgets and 250 iterations. The asterisked rows (including Opus 4.6) were evaluated via mini-swe-agent rather than the standard SWE-Agent. This affects both ranking and task selection; on a private repository set, Opus 4.1 dropped from 22.7 to 17.8 points, while GPT-5 fell from 23.1 to 14.9, essentially swapping places.

The K2.6 score in DeepSeek's report (58.6) matches Moonshot’s own card exactly, as does GLM-5.1 (58.4) in Z.ai's publication. Such precision across different sources suggests that researchers often copy results from published tables rather than re-verifying them — especially since DeepSeek admits it couldn't access certain models via API for testing.

For the models in our pricing table, the situation is as follows:

These values were collected on July 22, 2026, and none were obtained through a single unified evaluation process. DeepSeek's score was taken in "Max" mode, whereas the pricing table above doesn't specify a mode — and mode affects both the result and output token count (the final score).

The 80% score for Fable 5 is actually from BenchLM, which uses an average of five runs, whereas Moonshot’s announcement listed it as 80.3%. The 62.1 value for glm-5.2 comes from a Z.ai publication that compared it against GPT-5.5 (58.6). Competitors are often cherry-picked for specific benchmarks; for example, Kimi K2.7 Code was touted by Moonshot as having improvements across three proprietary sets (Kimi Code Bench v2, Program Bench, and MLS Bench Lite) but failed to provide results on any independent benchmarks. Alibaba claimed first place for qwen3.6-max-preview across six benchmarks, including SWE-bench Pro, though it only provided relative improvements over its predecessor rather than absolute numbers.

In summary: most benchmark scores in published tables are likely copied from the authors' own announcements rather than independently verified, and many "top" rankings rely on proprietary datasets that aren't available to the public.

Cost per Task

Artificial Analysis runs models on its own testbed and calculates a weighted average cost per task based on its own index, finally placing price and performance in the same coordinate system without mixing third-party data.

*Based on Version 4.1 Index, data as of July 22, 2026. Tokens per task estimated by multiplying time by speed. "Reliability" is the AA-Omniscience index (0 = equal correct/incorrect; negative values indicate more incorrect answers).

The cost per task for DeepSeek V4 Pro and Fable 5 differs by nearly 70x, even though their intelligence indices are within a factor of 1.5x (60 vs 44). This is the root of the "cheap Chinese models" argument — but that savings isn't free. A $0.04 cost per task sounds convincing until you see that DeepSeek is the only model with negative reliability, meaning it produces more incorrect answers than correct ones. You will have to manually verify its output, and those labor costs are never included in any price list.

This isn't a geographic issue; GLM-5.2 maintains a reliability of 4 while remaining both the second cheapest and the fastest. Both models are Chinese and significantly cheaper than US flagships, yet their reliability scores put them on opposite sides of the scale. You have to evaluate each model individually.

Generation time consists of two variables: token count and throughput (tokens/sec). DeepSeek outputs about 33k tokens per task — most being invisible reasoning tokens — at 69 tokens/sec, totaling 8 minutes. Kimi K3 is slower at 33 tokens/sec, taking 10 minutes for 20k tokens. GLM-5.2 is the most efficient, outputting 35k tokens in just over 3 minutes thanks to its 179 tokens/sec speed. Verbosity and speed are decoupled; you must evaluate them separately.

Note that "Generation Time" excludes time-to-first-token (TTFT) and overhead, so actual latency is higher. For batch overnight processing, the difference in minutes doesn't matter, but for an interactive assistant, token price matters less than responsiveness.

The Claude Fable 5 entry in the original table refers to a hybrid setup where some requests fall back to Opus 4.8; this measures the entire system, not just the raw model. This is actually a more honest way to calculate costs for production environments.

Data was collected on July 22, 2026, and no results were obtained through a single unified procedure. DeepSeek's score was captured in "Max" mode, whereas the pricing table lists it without a specific mode — and mode affects both accuracy and total token count (the final bill).

The Fable 5 score of 80% comes from BenchLM in Anthropic's System Map, which averages five runs. Meanwhile, Moonshot claimed a score of 80.3% for its K2.7 model in an announcement, showing that discrepancies exist even before comparing competitors. The glm-5.2 value of 62.1 comes from Z.ai, which compared it against GPT-5.5 (58.6). Opus 4.8 didn't even participate in that specific benchmark, though its successor did.

In the K2.7 Code announcement, Moonshot claimed improvements across three proprietary sets (Kimi Code Bench v2, Program Bench, and MLS Bench Lite), but provided no results on independent benchmarks. Alibaba's Qwen blog claimed first place on six benchmarks, including SWE-bench Pro, though it only offered relative improvements rather than hard numbers for its own proprietary tests (QwenClawBench and QwenWebBench).

Subscriptions: The Hidden Cost of "Unlimited"

Direct chat pricing works differently. Anthropic offers a $17/month Pro plan (billed annually at $200 upfront, or $20 monthly), followed by the Max plan starting at $100. Max offers a 5x or 20x capacity increase over Pro, with 20x costing $200 (billed monthly). Team plans offer Standard ($20/mo annual) and Premium ($125/mo) seats, which can be mixed within a team.

The Pro plan includes Claude Code, Cowork, Design, and Science — meaning agentic tools are available even at the entry-level paid tier. These consume limits fastest, as we will see below.

OpenAI uses EUR on its European page; these must be converted to USD before comparison. Their tiers include a €8 Go plan, a €23 Plus plan, and a €103 Pro plan. Business plans cost €21 per user (annual) or €26 (monthly). OpenAI's Pro tier is split into two levels that differ only in usage volume, as the feature sets are identical. Unlike Anthropic, OpenAI does not offer annual billing or multi-month prepayments for its consumer plans.

Chinese companies rarely use subscription models. DeepSeek uses a pure pay-as-you-go API model, recommending users to top up as they go. Z.ai publishes token rates and releases GLM-5.2 weights under the MIT license, making subscriptions largely irrelevant for those using open models.

Anthropic's Max plan and OpenAI's Pro tier both offer 5x or 20x quota increases, but they use different baselines: Anthropic scales from the Pro plan, while OpenAI scales from Plus. Neither company discloses the exact size of that baseline — a lack of transparency that was at the heart of the lawsuit mentioned earlier.

OpenAI does not allow users to pay for extra capacity within a tier; there is no mechanism to bypass limits by paying more. Anthropic, however, allows Pro and Max subscribers to use "credits" to work beyond their plan's volume via API rates — a feature introduced in May 2026.

OpenAI’s Go plan may include advertisements. While standard for consumer-grade apps, this is a notable shift for professional productivity tools.

Model names in official documentation have drifted from those used in benchmarks. OpenAI's documentation details reasoning capabilities: Medium and High levels are included in Plus, while Extra High (and the Pro mode) is reserved for Pro, Business, and Enterprise tiers. The Free and Go plans do not get GPT-5.6 Sol; they receive GPT-5.5 Instant, with a separate "Reasoning" button that allows ten messages every five hours on a different model.

When Subscriptions No Longer Cover the Flagship Model

On July 19, 2026, Anthropic ended its promotion where Fable 5 usage was included in the weekly subscription limit and moved to a separate pay-as-you-go model. While Pro and Team users still have access, they must now pay for Fable 5 usage on top of their subscription.

This "overage" mechanism isn't new; Anthropic’s credit documentation was published in May 2026, describing credits as a feature for Pro and Max users to handle usage spikes. The July transition simply applied this existing mechanism to Fable 5.

As part of the transition, Anthropic issued $100 in credit value to Pro and Standard Team users (capped at $2,500 per organization). These credits are on a separate balance that can be used for any model and expire on September 17.

Max and Premium seats in Team/Enterprise plans were excluded from this promotion; Fable 5 remains included within the subscription up to half of the weekly limit. Consequently, a $17 or $20 subscription price no longer defines your total cost — the flagship model has become an "add-on" expense that must be tracked separately in team budgets.

Enterprise plans were always structured this way: Anthropic's documentation lists a $20 base per seat plus API usage rates. For lower tiers, the true cost of using high-end models only became apparent with this change.

Limits: The Invisible Part of the Price

Anthropic promises Pro users at least 5x more capacity per session than Free users, though they specify that "capacity" is calculated based on peak hours and depends heavily on prompt length, chat history, and model choice. A session lasts five hours before the counter resets. Anthropic does not disclose exactly how many messages constitute a "Free" tier limit.

The multipliers in the high-end plan names refer to this 5-hour window: Max 5x provides five times the Pro limit, and Max 20x provides twenty times more.

Weekly limits are not explicitly defined in documentation; they exist as an unstated "black box." Pro users have one limit for all models, while Max users have two (a general limit and a specific Sonnet limit). The weekly counter resets at a fixed time unique to each account. Anthropic also reserves the right to impose additional monthly or weekly caps at their discretion.

The only absolute number provided in OpenAI's documentation is "160 messages every three hours" for GPT-5.5 Instant on Plus and Go plans. Free tier limits are described as a five-hour window with "dynamic" constraints based on system load and user behavior. Reasoning mode quotas aren't specified at all, except to note that Pro tiers have lower limits than the standard "hundreds of messages" tier.

When a Plus limit is hit, users are downgraded to GPT-5.5 Instant mini; when reasoning mode limits are reached, they fall back to GPT-5.4 Thinking mini. You'll only know you've been downgraded by the drop in response quality.

Token consumption is shared across all platforms — web, desktop, mobile, and Claude Code "eat from the same pot." Because Claude Code performs dozens of iterative steps (reading files, running commands, parsing output) per task, it can burn through limits extremely fast. Following complaints about users leaving agents running 24/7, Anthropic introduced specific weekly limits for agents in July 2025.

The Impact of Geopolitics on Model Selection

On June 12, 2026, following export control orders, Anthropic cut off access to Fable 5 and Mythos 5. The order applied to any foreign national, including employees with foreign passports, regardless of their location. There was no grace period.

Because Anthropic could not verify citizenship in real-time, they shut down access for everyone. While targeted users were the intent, everyone lost access. Control over Fable 5 was lifted on June 30 and it returned on July 1, but Mythos 5 — designed for cybersecurity professionals — remains restricted to a limited number of trusted US organizations according to Reuters. This resulted in three weeks of downtime for one model and ongoing uncertainty for the other; neither subscriptions nor contracts could prevent these outages.

Anthropic maintains separate supported country lists for their API and web interface, which do not always match. More importantly, they reserve the right to deny service to any organization that is "primarily owned" (directly or indirectly) by entities in a non-supported country. Simply renting a server in an approved country is not enough to bypass these rules.

A similar situation is unfolding in China. On July 7, 2026, Reuters reported that China's Ministry of Commerce held meetings with Alibaba, ByteDance, and Z.ai regarding restricting foreign access to the most powerful models — including those yet to be released or open-weights versions. Discussions included treating model "leaks" as national security violations. By July 21, Financial Times reported that China is considering adding advanced AI models to its export-controlled technology list, which would also ban the export of training data and the downloading of weights by foreigners.

The risk isn't just about the models; it’s about who owns the developers. In April, Beijing pressured Meta to cancel its acquisition of a Chinese startup, Manus, and in June introduced stricter controls on deals involving foreign investors and technologies. Anthropic's "ownership" clause and China's "investor" rules are two sides of the same coin.

Even users of Chinese models face risks. On July 8, 2026, CNBC reported that US lawmakers are investigating the use of Chinese models within American companies, sending inquiries to firms like Cursor and Airbnb. While there is no outright ban on usage yet, several agencies have implemented their own restrictions. Taiwan's cabinet banned government departments from using DeepSeek as early as February 3, 2025, citing data leakage risks and censorship concerns.

Open weights change the risk profile even further. Models like GLM-5.2, DeepSeek-V4, and the Kimi line are released under open licenses; once a user has downloaded the weights, it is technically difficult for regulators to "un-distribute" them. The US lawmakers acknowledge this, noting that while files are freely available online, bans face significant free speech hurdles. This makes open-weights models effectively immune to foreign mandates — though the "price" of this independence varies wildly based on model accuracy (e.g., DeepSeek vs. GLM-5.2).

At HOSTKEY, our demand for hosting open-weights models doesn't match public leaderboards. Four-fifths of our orders are for DeepSeek-R1 (the 14B parameter version), with the remaining fifth being qwen3-coder (30B); that covers everything.

None of the "flagship" open-weights models from our pricing table appear in this list, as customers typically rent compact versions that can run on affordable GPU servers. For these, token pricing is irrelevant; you pay for GPU hours, and your primary constraint is whether the model fits on the hardware.

However, there are exceptions. Qwen3.6-Max-Preview exists only as a hosted service without open weights — a deliberate shift from Alibaba's previous practice. If a model lives solely on a third-party platform, it can be restricted alongside that platform.

As one investigator noted, the world will be built on whichever models are cheapest and most accessible — and today, those are Chinese models. Every technical team faces this choice, often without the weight of high-level geopolitics behind them.

Therefore, before committing to a model, you must verify:

• What is the developer's jurisdiction and which regulations must they follow?
• Are the weights truly open, and does the license permit self-hosting?
• Is there a viable "failover" model in a different jurisdiction that has been tested under load, not just "kept in mind"?
• Will your own legal entity pass the provider's compliance checks, regardless of where your server is located?

Conclusions

The data in this analysis was gathered on July 22, 2026. Many figures are time-sensitive: deepseek-chat and reasoner were deprecated on July 24; Anthropic’s promo credits expire August 2; and the introductory price for Sonnet 5 ends August 31 (after which input rises from $2 to $3 per million tokens).

During our research, we fell into the exact trap we warned against. In our first draft, we used Alibaba's pricing for DeepSeek, which was 5.5x higher than DeepSeek’s own documentation. We also nearly misattributed a K2.6 benchmark score to K2.7 by following a trend in third-party summaries. Both errors were easy to make because the numbers looked "plausible" and appeared in existing, pre-packaged tables.

This analysis does not cover enterprise custom contracts, data training costs, or the infrastructure overhead (electricity/hardware) of self-hosting open models. We also did not dive into regional data privacy regulations or analyzed only one Chinese aggregator.