Coding plans are now the default way to use frontier models for a lot of people, and the gap between the frontier models and open-weight models is narrowing each week — Kimi 2.6 and MiMo Pro score 54 on the Artificial Analysis Index, compared to 57/60 for Opus-4.7/GPT-5.5.

It is obvious that coding plans are the cheapest way to access the most intelligent models. But how does the pricing compare?

Blended subscription rate*

Provider	$ / M blended tokens
MiniMax 2.7	$0.004
Kimi 2.6	$0.047
GLM 5.1 (Lite)	$0.065
Codex (GPT-5.5)	$0.080
MiMo V2.5-Pro	$0.141
Claude Pro (Opus 4.7)	$0.744

* Subscription cost ÷ monthly tokens delivered, on a blended Claude-Code workload. See the methodology note at the bottom.

The usage these plans provide is intentionally obfuscated and likely played around with depending on supply and demand. I make an effort here to measure and snapshot what you actually get on each plan. This data is from May 1st, 2026.

I proxy each request through a server that logs, and measure input tokens, thinking tokens, output tokens, and calculate the price it would cost to use these models directly from the API.

Two notable absences: DeepSeek v4 doesn’t offer a subscription plan — only pay-as-you-go API access. Gemini has yet to launch a serious coding plan; Code Assist on the OAuth path is request-capped at the free tier and meters by request count rather than tokens.

Subscription	$/mo	5h cap	Weekly cap	Sessions/wk	Monthly cap	API-$ / sub-$	Monthly tokens	Tokens / sub-$
Claude Pro — Opus 4.7	20.00	$4.75	$38	8.0×	$152	7.6×	26.9 M	1.35 M
Codex (GPT-5.5)	20.00	$22	$134	6.0×	$536	26.8×	250 M	12.5 M
Kimi 2.6 (for Coding)	20.00	$8.50	$40	4.7×	$160	8.0×	423 M	21.1 M
GLM 5.1 (Lite)	18.00	$7.70	$34.60	4.5×	$138	7.7×	275 M	15.3 M

Two more plans don’t fit the rolling-$-cap pattern, so I split them out. MiniMax meters request count instead of tokens, and MiMo gives a flat monthly credit pool with no rolling 5h or weekly windows.

Subscription	$/mo	Weekly cap	Monthly cap	API-$ / sub-$	Monthly tokens	Tokens / sub-$
MiniMax 2.7	20.00	45,000 req	$675 (at 30K tok/req)	33.8×	5,400 M	270 M
MiMo V2.5-Pro (token-plan-sgp)	14.08	—	200 M credits / $14.08	1.0×	100 M	7.1 M

Speed

Opus 4.7 in Claude is the most expensive option in subscription plans, but possibly one that costs lesser in user frustration. I personally reach for it again and again for two reasons:

It understands my intention and direction I want to go in and gets working on it.
It is also the fastest to execute as is seen from the table below.

Provider	n	TTFT avg	Out avg	TPS avg	TPS max
anthropic	23	2244 ms	1306	82.3	159.6
zai-glm	324	5097 ms	5659	72.9	107.5
xiaomi-mimo	10	2791 ms	1182	62.9	82.6
minimax	26	2048 ms	787	53.9	88.1
kimi-coding	148	3848 ms	863	50.9	89.8
openai-chatgpt	159	1566 ms	815	46.1	54.5

A quick practical tip to maximize usage. I used Claude Code as the harness and in using Kimi and MiniMax API’s, I didn’t have a faster, smaller model to work with. I instead routed every Haiku level call to deepseek-v4-flash. I loaded $2 for all of these experiments and I still have most of it.

A caveat on scope. Something that I haven’t tested for is using these subscriptions outside of coding harnesses. Most of them support being used in OpenClaw/Hermes setups other than Claude Code. While I was able to use the APIs of all the chinese providers for testing purposes, your mileage may vary and they may strike your account with a ban.

A note on blended usage. From proxy logs, an average Claude Code call breaks down as 92.4% cache_read · 5.2% output · 2.4% fresh input. Each provider prices these axes differently, so I apply each one’s per-axis pricing to the same workload — all the numbers above sit on identical request shapes.