Back to all guides
weekly-ranking11 min readNovember 16, 2026

2026 W48 Cheapest AI API Weekly Ranking: Four Models at CNY 0.80, Three at USD 0.10

The November 16 re-verification of eleven vendors found zero price movement, making W48 a clean market snapshot. Four domestic models share the CNY 0.80 per million input floor (Alibaba Qwen3.8-Flash, Alibaba Qwen3.8-Omni-Flash, Volcengine Doubao Seed 2.1 Lite, Zhipu GLM-5.3-Flash) and three dollar models share the USD 0.10 floor (OpenAI GPT-6 Luna, Anthropic Claude Haiku 5.5, Meta Muse Ember), all with 1M context. Includes a 100M-tokens-per-month cost table, a cache-hit cost table showing where the real savings live, DeepSeek valley-hours scheduling, and the tokenizer caveat that shifts every cross-model comparison.

AI API price comparison looks like one question and is really three: what is the headline rate, what does your prompt length trigger, and how much lands on cache. The November 16 re-verification of eleven vendors found zero price movement, so this week's ranking reflects the market rather than a promotion expiring. Four domestic models sit at CNY 0.80 per million input, three dollar models at USD 0.10, and the details between those two floors decide the bill.

W48 2026 weekly snapshot (Nov 16)

The November 16 re-verification of eleven vendors' official pricing pages found zero price movement. The cheapest domestic input tier is CNY 0.80 per million, held jointly by Alibaba Qwen3.8-Flash, Qwen3.8-Omni-Flash, Volcengine Doubao Seed 2.1 Lite and Zhipu GLM-5.3-Flash. The dollar floor is USD 0.10, held jointly by OpenAI GPT-6 Luna, Anthropic Claude Haiku 5.5 and Meta Muse Ember, all with 1M context. Two tiers were added this round: Google Gemini 4 Hydrogen at USD 1.50 / 7.50, which lands 25 percent below the Gemini 4 Carbon promotional rate, and Claude Mythos 5.2 at USD 8 / 40.

The CNY 0.80 input floor: four models, three different reasons

The output prices across those four differ by one cent, but the reasons behind the tie differ, and that decides whether you can count on the rate next month. Zhipu's official table lists a standard rate only, with no discount row, so CNY 0.80 is the price you can plan around. Alibaba Qwen3.8-Omni-Flash carries an explicit CNY 0.10 cached-input rate, the cheapest domestic cache read. Doubao's CNY 0.80 covers text input only, with audio priced separately at CNY 12 per million, which changes the arithmetic immediately for any voice workload.

The USD 0.10 floor and the 100K prompt cliff

All three dollar models carry a prompt-length cliff. Claude Haiku 5.5 prices at USD 0.10 / 0.50 for prompts up to 100K tokens, and the entire request reprices to USD 0.50 / 2.50 above that. Meta Muse Ember's cliff sits at 200K tokens, where the request is charged at USD 0.30 input and USD 1.50 output. These are not footnotes; they are the difference between a USD 30 monthly bill and a USD 300 one for the same workload. DeepSeek is the only model in the library with no cliff, because its valley pricing halves the peak rate outright.

Valley pricing: the only schedule-aware discount left

DeepSeek remains the only vendor here with a time-based rate card. DeepSeek V4.1 Flash bills at USD 0.15 input and USD 0.60 output during off-peak hours and USD 0.30 / 1.20 at peak, with cached input at USD 0.003 off-peak, the lowest cached-input price in the library. V4 Pro follows the same structure at USD 0.66 / 1.98 in the valley. Peak covers weekdays 01:00-04:00 and 06:00-10:00 UTC, and the full rate applies across weekends and Chinese public holidays. For batch work that can be scheduled this is worth roughly half the bill; for interactive traffic it is worth nothing.

100M tokens per month: what each tier actually costs

The figures below cost 100M tokens per month at a 50/50 input-output split, converted at USD 1 = CNY 7.16. Xiaomi MiMo-V2.6-Flash runs about CNY 150, the four CNY 0.80 domestic tiers land between CNY 175 and CNY 180, the three dollar-floor models about CNY 215, Google Gemini 4 Hydrogen about CNY 3,220, and Claude Mythos 5.2 about CNY 17,180. Read this as an anchor rather than a quote: cache hit ratios, prompt lengths and batch discounts all move the number.

The cache lever: where the real savings live

If 100M tokens per month all hit cache, the bill collapses. DeepSeek V4.1 Flash at the USD 0.003 valley cache rate costs roughly CNY 2.15, Meta Muse Ember at USD 0.004 follows closely, and GPT-6 Luna and Claude Haiku 5.5 at USD 0.01 land near CNY 7. Domestically, Qwen3.8-Omni-Flash and MiniMax M3.1-Flash at CNY 0.10 cost about CNY 10, Doubao Seed 2.1 Lite at CNY 0.16 about CNY 16, and GLM-5.3-Flash at CNY 0.23 about CNY 23. Anthropic's current tiers price cached input at 2.5 percent of the standard rate and OpenAI at 5 percent, so any workload with a stable system prompt or a large retrieved-document prefix should be engineered for cache hits first.

The tokenizer caveat that changes every comparison

Anthropic's pricing page states that Claude 4.7 and later models, plus Claude Mythos Preview, use a new tokenizer that produces roughly 30 percent more tokens from the same text, while Claude Sonnet 4.6 and earlier keep the old one. A comparison built on character counts therefore understates the newer model's true cost by about a third. Two further traps sit alongside it: MiniMax M3 input cost doubles once a request crosses 512K tokens, and xAI reprices the entire request at the long-context tier once the prompt exceeds 200K tokens rather than charging only for the excess.

When to stop optimizing and start picking a plan

Metered pricing only wins while your usage is irregular. The moment usage becomes predictable, a subscription wins, because the subscription converts a variable bill into a fixed one. The current domestic entry points are Alibaba Cloud Token Plan Lite at CNY 39 per month for 20,000 credits, Baidu Qianfan from CNY 9.9, and the IDE coding subscriptions starting at Cursor Plus USD 8. The break-even arithmetic depends on how steady your daily token volume is, not on any single price. If your monthly usage swings by a factor of three, stay on metered pricing; if it is stable within twenty percent, a plan will usually be cheaper.

Frequently Asked Questions

What is the cheapest AI API per million tokens right now?

Two floors exist at once. The cheapest domestic meter price is CNY 0.80 per million input tokens, held jointly by Alibaba Qwen3.8-Flash, Alibaba Qwen3.8-Omni-Flash, Volcengine Doubao Seed 2.1 Lite and Zhipu GLM-5.3-Flash, all with 1M context. The cheapest dollar price is USD 0.10 per million input, held jointly by OpenAI GPT-6 Luna, Anthropic Claude Haiku 5.5 and Meta Muse Ember, also with 1M context. On output the picture changes: the lowest domestic output is CNY 2.00 on Xiaomi MiMo-V2.6-Flash, and the lowest dollar output is USD 0.50 on the same three dollar models.

Which models have the cheapest cached-input rates?

DeepSeek V4.1 Flash has the lowest cached-input rate in the library at USD 0.003 per million during off-peak hours and USD 0.006 at peak. Meta Muse Ember follows at USD 0.004, with USD 0.012 once the request exceeds 200K tokens. OpenAI GPT-6 Luna and Anthropic Claude Haiku 5.5 both sit at USD 0.01 for prompts up to 100K tokens. Domestically, Alibaba Qwen3.8-Omni-Flash and MiniMax M3.1-Flash are both CNY 0.10, Volcengine Doubao Seed 2.1 Lite is CNY 0.16 and Zhipu GLM-5.3-Flash is CNY 0.23. Anthropic's current tiers price cached input at 2.5 percent of the standard input rate.

Is the cheapest model also the one with the lowest total cost?

No, and this week's ranking is a good example. Xiaomi MiMo-V2.6-Flash is the cheapest domestic option at CNY 1 input and CNY 2 output, which at 100M tokens per month works out to roughly CNY 150, while GLM-5.3-Flash at CNY 0.80 input works out to roughly CNY 180. The gap is small at this volume, but it reverses at higher volumes and on workloads that generate more output than input. Cache hit ratio and prompt-length cliffs usually move the bill more than the headline input price does.

What are the current promotions worth watching this week?

Three offers are live as of November 16. Tencent TokenHub offers 50 percent off Hy4 metered pricing, bringing input down to CNY 3 per million, valid until December 31. Tencent's Token Plan points system offers 2.5-fold pricing on the same model, effectively CNY 1.50 per million input, but that window closes on November 17. Baidu Qianfan's ERNIE 5.5 Flash preview runs at 50 percent during the day and 25 percent overnight, valid until November 30. For reference, the Alibaba Cloud Token Plan Pro tier has been cut to CNY 312 per month from CNY 499 since November 11.

How does the new Anthropic tokenizer affect cost comparisons?

Anthropic's pricing page states that Claude 4.7 and later models, plus Claude Mythos Preview, use a new tokenizer that produces roughly 30 percent more tokens from identical text, while Claude Sonnet 4.6 and earlier retain the old tokenizer. A comparison against an older Anthropic model that is based on character counts therefore understates the newer model's true cost by about a third. The same caution applies to the new Claude Mythos 5.2 tier added on November 16. Normalize by actual token counts before comparing.

cheapest-apiweekly-rankingw48qwen3.8-flashglm-5.3-flashdoubao-seed-2.1-litegpt-6-lunaclaude-haiku-5.5gemini-4-hydrogendeepseek-valleycache-hit

Want more AI pricing comparisons?

Check our price changelog for daily updates on AI model pricing, new releases, and deprecations.