Cost strategy · verified June 2026

The West's heaviest AI subs all cost ~$200. China's cost a tenth — or less.

Four Western labs cluster at $200–300/mo for their top consumer plan. China's equivalents are $10–80/mo (or pure metered API, 20–40× cheaper per token for ~90–95% of the capability). A $200 sub is a great dev seat — but you can't run a product on it. Below: both questions, scored. See also the Max-20x calculator and more dispatches.

Executive summaryclick to expand/collapse

The Western $200+ club (top individual plans): Claude Max 20x $200, ChatGPT Pro $200, Google AI Ultra $200 (down from $250), SuperGrok Heavy $300. One price band.

China, same agentic-coding workflow: Zhipu GLM Coding $10/$30/$80 (runs inside Claude Code; GLM-5.2, 1M context, trained without Nvidia); Moonshot Kimi $19→$199 (Kimi Code + Agent Swarm up to 300 subagents); DeepSeek & Qwen — no premium tier at all, just metered API.

On the API — where you ship a product — DeepSeek V4-Pro 0.435/0.87, V4-Flash 0.14/0.28, Qwen-Plus 0.4/1.2, GLM-4.7 0.6/2.2 per 1M — vs Claude Opus 4.8 $5/$25 and Fable 5 $10/$50. Roughly 20–40× cheaper per token.

Where it lands: your dev seat → a $10–30 Chinese, Claude-Code-compatible plan does most of what a $200 Western sub does; a product you ship → the cheap stack wins by an order of magnitude; pay Western-frontier prices only for the hardest tasks. The frontier is real. The price floor is being set in China.

Time-sensitive: Fable 5 is free on Claude subs only through Jun 22, then converts to API-rate credits (~2× faster burn than Opus). From Jun 15, programmatic/Agent-SDK usage draws a separate API-rate pool.

The Western $200 club vs China's coding plans

Top consumer/coding plans. Western flagships cluster at $200–300; China's Claude-Code-compatible coding plans run $10–80, and DeepSeek/Qwen skip subscriptions entirely.

Plan	Price	Top model	Notes
Claude Max 20x Anthropic	$200/mo	Opus 4.8	Fable 5 free through Jun 22
ChatGPT Pro OpenAI	$200/mo	GPT-5.5 Pro	20× Plus limits, 1M context
Google AI Ultra Google	$200/mo	Gemini 3.1 Pro · Deep Think	20× Pro; down from $250
SuperGrok Heavy xAI	$300/mo	Grok 4.3 + Grok 4 Heavy	multi-agent; max rate limits
Zhipu GLM Coding	Lite $10 · Pro $30 · Max $80 /mo	GLM-5.2 (1M ctx)	runs inside Claude Code on a quota system; trained without Nvidia; MIT open weights
Moonshot Kimi membership	$19 / $39 / $99 / $199 /mo	Kimi K2.6	Kimi Code + Agent Swarm up to 300 parallel subagents; Modified-MIT open weights
DeepSeek	no premium tier — free app + metered API	V4-Pro / V4-Flash	MIT open weights; OpenAI- & Anthropic-compatible API
Alibaba Qwen	no premium tier — free Qwen Chat + metered API	Qwen 3.7 Max / Qwen-Plus	DashScope / Model Studio; global endpoint

Dev-seat value — what each ~$200 subscription is worth

API-equivalent inference value at 1B tokens/mo (7:2:1) ÷ price = leverage. Higher leverage ⇒ it arbitrages a more expensive API; not a quality ranking.

Subscription	$/mo	Top model	API-equiv @1B	Leverage
ChatGPT Pro OpenAI	$200	GPT-5.5	$4,350	21.8×
Claude Max 20x Anthropic	$200	Claude Opus 4.8	$3,850	19.3×
MiniMax Agent Plus MiniMax	$20	MiniMax-M3	$264	13.2×
GLM Coding Max Zhipu / Z.ai	$80	GLM-5.1	$902	11.3×
Google AI Ultra (top) Google	$200 $99.99 entry · was $250	Gemini 3.1 Pro	$1,740	8.7×
Doubao Professional ByteDance	$69	Doubao Seed 2.0 Pro	$364	5.3×
Kimi Vivace Moonshot	$199	Kimi K2.6	$702	3.5×
SuperGrok Heavy xAI	$300	Grok 4.3	$640	2.1×

Product-runtime cost — metered API per 1M tokens

Serving real traffic. Last column adds ~$120/mo of Tavily search. Sorted cheapest blended cost first.

Model	in $/M	cache $/M	out $/M	@1B/mo	+ Tavily
DeepSeek V4-Flash	0.14	0.014*	0.28	$66	$186
DeepSeek V4-Pro	0.435	0.003625	0.87	$177	$297
Qwen-Plus	0.4	0.04*	1.2	$228	$348
MiniMax-M3	0.3	0.12	1.2	$264	$384
Doubao Seed 2.0 Pro	0.47	0.047*	2.37	$364	$484
GLM-4.7	0.6	0.06*	2.2	$382	$502
Grok 4.3	1.25	0.2	2.5	$640	$760
Kimi K2.6	0.95	0.16	4	$702	$822
GLM-5.1	1.4	0.26	4.4	$902	$1,022
Qwen 3.7 Max	2.5	0.25*	7.5	$1,425	$1,545
Gemini 3.1 Pro	2	0.2	12	$1,740	$1,860
Claude Opus 4.8	5	0.5	25	$3,850	$3,970
GPT-5.5	5	0.5	30	$4,350	$4,470
Claude Fable 5	10	1	50	$7,700	$7,820

* cache rate estimated. DeepSeek V4-Flash ($66/mo) is ~59× cheaper than Opus; V4-Pro ~22× cheaper than Opus and ~44× cheaper than Fable 5. At these prices the ~$120 search layer is ~40% of the DeepSeek bill — see the build-vs-buy search dispatch.

The takeaway — tiered routing

Default traffic

Cheap, capable stack — DeepSeek/Qwen/GLM + your own search. Order-of-magnitude cheaper for ~90–95% of the capability.

Hardest tasks

Escalate to Opus / Fable selectively, cap tokens, cache hard. Pay frontier prices only where they change the outcome.

Your dev seat

Flat-rate sub — a $10–30 Chinese Claude-Code-compatible plan, or a $200 Western sub for frontier-tier personal coding.