The LLM API market in 2026 has produced a phenomenon that puzzles many developers: running the same customer-service chatbot at 10M tokens a month costs about $140 on a US frontier model, but only about $30 or even $4 on a Chinese model — an order of magnitude apart. It is not that Chinese models cut corners; the entire cost structure is different. This piece lays out the prices, capabilities, and scenario costs of mainstream Chinese and US LLMs side by side and does the math by actual usage. No narrative, just data.
One thing up front: every price below is a summary of public quotes as of July 2026, sourced from each vendor's official pricing page. Different hosting platforms and time points produce different numbers. This is a representative comparison, not a hands-on benchmark; capability assessments are based on public reviews. By the time you read this, prices will likely have updated — treat the official latest pricing as the source of truth, not this article.
1. Why LLM selection in 2026 is about cost-performance
Picking an LLM used to be simple: check the leaderboard, pick the highest Elo. That no longer works in 2026, for three reasons.
First, the capability gap among frontier models is narrowing. GPT-5, Claude Opus 5, and Gemini 3 are tightly bunched on most benchmarks; no single model dominates across the board. When capability differences shrink, price becomes the deciding factor.
Second, Chinese models have entered the arena at rock-bottom prices. DeepSeek, Qwen, and Kimi are not "a bit cheaper" — they are an order of magnitude cheaper. Fortune and other outlets are already discussing how "Chinese labs are beating US labs on cost" (China's Moonshot/DeepSeek/Z.AI beating US labs on cost). This is not a marketing gimmick; it is a real API pricing gap.
Third, compliance and data security have become hard constraints. DoorDash was probed by US lawmakers for using the Chinese model Kimi K2.6, which shows that model selection is no longer purely a technical decision — it involves supply-chain compliance. Cheap does not mean free to use; this is the new variable in 2026 model selection.
So the selection logic has shifted: it is no longer "pick the most powerful" but "pick the most cost-effective model that meets your capability needs and whose compliance risk is acceptable." This article breaks it down along those lines.
2. The lineup: price overview (data as of 2026-07)
Here are the API prices for mainstream Chinese and US models. Prices are per million tokens (input/output), in USD. Sourced from a summary of public quotes as of July 2026; each vendor's official pricing page is authoritative, and different hosting platforms and time points produce different numbers.
| Model | Vendor | Country | Input ($/M tokens) | Output ($/M tokens) | Positioning |
|---|---|---|---|---|---|
| GPT-5 | OpenAI | US | 10 | 30 | Flagship |
| Claude Opus 5 / Opus 4.6 | Anthropic | US | 5 | 25 | Flagship |
| Claude Sonnet 5 | Anthropic | US | 2 | 10 | Promo through 2026-08-31, then 3/15 |
| GPT-4.1 | OpenAI | US | 2 | 8 | Production workhorse |
| Gemini 3 Flash | US | ~3 | ~3 | 10M tokens ~$30 | |
| Grok 4.5 | xAI | US | 2 | 6 | Cheapest production-grade (score 70+) |
| GPT-5 Nano | OpenAI | US | — | — | Budget tier, see official pricing |
| Claude Haiku 4.5 | Anthropic | US | — | — | Budget tier, see official pricing |
| DeepSeek-V3 | DeepSeek | China | 0.27 | 1.10 | DeepInfra hosted, ~7x cheaper than GPT-4.1 |
| Qwen3.7 Flash | Alibaba | China | 0.03 | 0.13 | Cheapest |
| Kimi K2.6 | Moonshot | China | — | — | Open-source Modified MIT, see official pricing |
| GLM | Zhipu | China | — | — | See official pricing |
Two details stand out. First, the US camp itself has three price tiers: flagship (GPT-5 $10/$30, Opus 5 $5/$25) is the top tier, production workhorses (Sonnet 5 $2/$10, GPT-4.1 $2/$8, Gemini 3 Flash ~$3, Grok 4.5 $2/$6) are the middle tier, and budget models (GPT-5 Nano, Haiku 4.5) are the bottom tier. Within the middle tier, prices are close, but the gap to the top tier is 3-5x. Second, Chinese models are priced in a different league. DeepSeek-V3 at $0.27/$1.10 is about 7x cheaper than GPT-4.1 and 10-14x cheaper than Sonnet 5 at promo pricing (by total cost). Qwen3.7 Flash at $0.03/$0.13 is the cheapest model in the entire table.
Note the "—" cells. GPT-5 Nano and Claude Haiku 4.5 are budget-tier models whose pricing floats by tier, so no fixed number is given. Kimi K2.6 is open-source under Modified MIT and is self-hosted or third-party-hosted, so pricing depends on the platform. GLM is similar. For these models, check the official or hosting-platform latest quote; do not treat this table as a contract.
3. Scenario cost comparison: doing the math by actual usage
Unit price alone is not intuitive enough; monthly cost by scenario is what matters. Take a customer-service chatbot as the representative scenario: 10M tokens per month, estimated at an 80% input + 20% output ratio (customer service has far more input than output). The GPT-5.2 ~$140 and Gemini 3 Flash ~$30 figures are BenchLM case references; the rest are calculated from public pricing. Representative comparison, not a hands-on benchmark; official pricing is authoritative.
| Model | Input price | Output price | Monthly cost (8M input + 2M output) | Relative to GPT-5 |
|---|---|---|---|---|
| GPT-5 | $10 | $30 | ~$140 | 100% |
| Claude Opus 5 | $5 | $25 | ~$90 | 64% |
| Claude Sonnet 5 (promo) | $2 | $10 | ~$36 | 26% |
| GPT-4.1 | $2 | $8 | ~$32 | 23% |
| Gemini 3 Flash | ~$3 | ~$3 | ~$30 | 21% |
| Grok 4.5 | $2 | $6 | ~$28 | 20% |
| DeepSeek-V3 | $0.27 | $1.10 | ~$4.4 | 3% |
| Qwen3.7 Flash | $0.03 | $0.13 | ~$0.50 | 0.4% |
The last row is the most striking: running the same 10M-token customer-service chatbot, GPT-5 costs $140 while Qwen3.7 Flash costs $0.50 — a 280x difference. DeepSeek-V3 at $4.4 is 3% of GPT-5. Even within the US camp, Grok 4.5 ($28) is 5x cheaper than GPT-5 ($140), and Sonnet 5 at promo pricing ($36) is nearly 4x cheaper.
But do not look at price alone. This table assumes "capability is sufficient" — a customer-service chatbot does not need GPT-5's reasoning ceiling; it needs stability, speed, and low cost. If your scenario is complex reasoning, code generation, or long-document analysis, the capability gap among frontier models will show, and a cheap model may "save money but fail the task." So the first step in selection is not scanning the price table; it is defining your capability floor, then picking the cheapest model that meets it.
4. One by one: each model's best range
GPT-5: the flagship ceiling, powerful but expensive
OpenAI's flagship at $10/$30, the most expensive in the table. GPT-5 remains top-tier in complex reasoning, code generation, and multimodal understanding. Best for scenarios demanding peak capability where budget is not a constraint: financial analysis, legal reasoning, complex code architecture design. The cost is steep — 10M tokens a month runs $140, over 30x what DeepSeek-V3 costs. If your scenario is not "GPT-5 or nothing," using it is burning money.
Claude Opus 5 / Opus 4.6: Anthropic's flagship, strong on long context and agents
$5/$25, half the price of GPT-5. Anthropic's Opus line has a reputation for long-context understanding and agent tasks (multi-step reasoning plus tool use). Best for scenarios requiring ultra-long context windows and complex agent workflows. The tradeoff is that the $25 output price is still high for production models, making high-output scenarios costly.
Claude Sonnet 5: the best value in the US camp during the promo window
$2/$10, promo through 2026-08-31, then moving to $3/$15. During the promo window, Sonnet 5 is one of the best-value models in the US camp: capability at roughly 80-90% of Opus 5, at 40% of the price. Monthly cost $36, 26% of GPT-5. Best for medium-to-high-load scenarios that need strong capability but have budget limits. Note that after the promo ends, cost rises to about $54/month (at the same usage), at which point the value proposition weakens and reevaluation is needed.
GPT-4.1: the production workhorse, stable and reliable
$2/$8, OpenAI's production workhorse. Less capable than GPT-5 but sufficient for most production scenarios, at 23% of the price. Monthly cost $32. Best for scenarios that do not need flagship-level reasoning but require stability and a mature ecosystem: API product backends, enterprise applications, content generation.
Gemini 3 Flash: Google's value play
~$3 input and output at the same price, 10M tokens for about $30. Google's Flash line has always gone for "good enough and cheap." Same-price input and output simplifies billing; monthly cost $30 is close to GPT-4.1 but with simpler accounting. Best for scenarios with balanced input-output ratios and Google ecosystem integration (search, multimodal). The BenchLM customer-service case — Gemini 3 Flash at $30 vs GPT-5.2 at $140 — is its sweet spot.
Grok 4.5: the cheapest production-grade model (score 70+)
$2/$6, the cheapest production-grade model in the US camp (benchmark score 70+). Monthly cost $28, cheaper than GPT-4.1. Best for budget-sensitive scenarios that still need a production-grade capability guarantee. The tradeoff is a less mature ecosystem than OpenAI/Anthropic, with fewer tooling and integration options.
DeepSeek-V3: China's cost-performance benchmark
$0.27/$1.10, hosted on DeepInfra. About 7x cheaper than GPT-4.1 and 10-14x cheaper than Sonnet 5. Monthly cost $4.4, the second-cheapest model in the table. On capability, DeepSeek-V3 approaches GPT-4.1 on most benchmarks, with a strong reputation for reasoning and code. Best for scenarios with extreme budget sensitivity, high volume, and near-production-grade capability needs: large-scale content generation, customer service, data-labeling assistance. The tradeoff is that it is hosted on a third-party platform (DeepInfra), so data compliance must be self-assessed; deployments outside mainland China may be affected by network latency.
Qwen3.7 Flash: floor price, pick it only at extreme scale
$0.03/$0.13, the cheapest in the table. Monthly cost $0.50, 0.4% of GPT-5. The Qwen Flash line is positioned as lightweight and fast, best for ultra-large-scale scenarios where per-call quality does not need to be extreme: batch classification, simple Q&A, data extraction. If your monthly volume is at the 100M-token level, Qwen3.7 Flash's cost advantage scales to a staggering degree. The tradeoff is that complex reasoning is weaker than frontier models; it is not suited for high-difficulty tasks.
Kimi K2.6: the open-source Chinese contender, a compliance variable
Made by Moonshot, open-source under Modified MIT license. Kimi's uniqueness is that it is simultaneously a "Chinese model" and an "open-source model" — you can self-host it and keep data in-house, but this also means it was swept into the compliance controversy by the DoorDash incident. Best for scenarios requiring open-source control and willingness to self-host, but supply-chain compliance risk must be assessed before selection — especially for companies serving the US market or under US regulation.
5. A decision tree: which one for you
Don't pick by hype; pick by the job. Here is a path to slot yourself into.
You need flagship-level reasoning (complex reasoning, legal or financial analysis, high-difficulty code architecture) and budget is not a constraint. Pick GPT-5 or Claude Opus 5. GPT-5 is the capability ceiling but the most expensive; Opus 5 is half the price and strong on long context and agents. Choose by scenario preference.
You need strong capability but have a limited budget, chasing value within the US camp. Pick Claude Sonnet 5 (during the promo window) or GPT-4.1. Sonnet 5 at promo $2/$10 is the best deal; note the price hike after 2026-08-31.
You want a production-grade capability guarantee, are budget-sensitive, and are staying in the US camp. Pick Grok 4.5 ($2/$6, cheapest production-grade with score 70+) or Gemini 3 Flash (~$3, same input-output price simplifies billing).
Your budget is extremely tight, volume is high, and capability needs are mid-to-upper. Pick DeepSeek-V3 ($0.27/$1.10, 7x cheaper than GPT-4.1). Assess data compliance before deploying.
You have ultra-large-scale volume and per-call quality does not need to be extreme. Pick Qwen3.7 Flash ($0.03/$0.13, cheapest in the table).
You need open-source control and are willing to self-host. Pick Kimi K2.6 (open-source Modified MIT). But you must assess supply-chain compliance risk.
Two common combos. One, flagship for hard jobs, cheap models for volume: GPT-5 or Opus 5 handles complex reasoning and high-value tasks, while DeepSeek-V3 or Qwen Flash handles simple tasks in bulk, routing by task difficulty. Two, US models for compliance-sensitive scenarios, Chinese models for non-sensitive ones: use GPT-4.1 or Grok 4.5 for US-facing or regulated work, and DeepSeek-V3 for internal tools and non-sensitive scenarios — compliance and cost, both covered.
6. Four pitfalls: pricing units, compliance risk, capability floor, hosting differences
First, pricing units differ, so don't compare raw numbers. Some models have the same input and output price (Gemini 3 Flash), while others have cheap input and expensive output (GPT-5 $10/$30). Your actual cost depends on the input-output ratio: in customer service, input far exceeds output, so models with high output prices lose out; in content generation, output dominates, so low input prices help. Estimate your input-output ratio first, then convert to total cost — don't just look at the input or output price alone.
Second, compliance risk is the new variable in 2026. DoorDash was probed by US lawmakers for using Kimi K2.6, showing that using Chinese models is not just a technical decision but a supply-chain compliance issue. Companies serving the US market or under US regulation must clear legal review before selection: the compliance risk of Chinese models (DeepSeek, Qwen, Kimi, GLM) needs assessment, no matter how high the technical cost-performance. This is not fear-mongering; there is already a precedent.
Third, capability gaps widen in extreme scenarios. Cheap models close the gap with flagships on simple tasks, but pull apart on complex reasoning, long-chain agents, and multimodal understanding. A customer-service chatbot on Qwen Flash is fine, but financial reasoning on Qwen Flash may "save money and produce errors." Define your capability floor first, then pick the cheapest model that meets it — don't save money by picking a model that can't do the job.
Fourth, the same model can cost differently across hosting platforms. DeepSeek-V3's $0.27/$1.10 is the DeepInfra hosted price; the official DeepSeek API or other platforms may quote differently. Price, latency, uptime, and data policies all vary by host. Shop around before committing to an API; don't assume one quote applies across all platforms.
7. Why Chinese models are so much cheaper
Many people ask: are Chinese models this cheap because of some trick? The answer is a different cost structure, not cutting corners.
First, training efficiency. DeepSeek and other Chinese vendors have invested deeply in algorithmic optimization, training near-frontier models with less compute, resulting in lower training costs than US peers at the same tier.
Second, engineering optimization. Inference-side optimizations (quantization, batching, KV cache optimization) have drastically reduced the per-token cost of inference, and these savings flow directly into API pricing.
Third, market strategy. Chinese vendors use low prices to capture market share and developer ecosystems; the model itself may not be profitable, with revenue coming from companion cloud services. This is the same logic as US giants subsidizing model pricing with cloud revenue, just more aggressive.
Fourth, open-source competition. Kimi K2.6 is open-source under Modified MIT, and DeepSeek also has open-source versions. Open-source models drag down the market's pricing benchmark, making it harder for closed-source models to charge a premium.
But cheap comes with tradeoffs. Compliance risk (the DoorDash incident), data-policy uncertainty, network latency for overseas deployment, and capability discounts in English-language scenarios are all hidden costs to factor into selection. Cost-performance is not just about the API unit price; it is about total cost of ownership.
References
- Each vendor's official pricing page (OpenAI / Anthropic / Google / xAI / DeepSeek / Alibaba / Moonshot / Zhipu), per a summary of public quotes as of July 2026
- DeepInfra hosted DeepSeek-V3 pricing
- Fortune: China's Moonshot/DeepSeek/Z.AI beating US labs on cost
- BenchLM customer-service cost case (GPT-5.2 ~$140 vs Gemini 3 Flash ~$30)
- Artificial Analysis LLM Elo leaderboard and public benchmarks