coding-subscription on tomrochette.com

Kimi Code

tom@tomrochette.com (Tom Rochette) — Sat, 26 Sep 2026 00:00:00 +0000

Kimi Code is Moonshot AI’s developer coding subscription, selling the K-series models through a CLI, desktop app, and VS Code extension under a membership quota. The same membership is sold at a yen ladder on kimi.com and a dollar ladder on kimi.ai, and the entry tiers do not convert between them. Facts below verified as of 2026-09-26.

What it is
#

Moonshot AI sells Kimi Code as the developer service inside a Kimi membership, with CLI, desktop, IDE, and third-party requests drawing on one shared quota. Third-party harnesses connect through OpenAI-compatible and Anthropic-compatible endpoints on api.kimi.com. Model access is tiered: K2.7 Code for every paying member, K3 at full 1M context from the mid tier up, with RMB-billed Extra Usage as overflow.

Status
#

Active and restructuring quickly: K3 shipped into Kimi Code on July 17, 2026, and by late September a restructure had renamed tiers and dropped the weekly quota window for new members. The CLI is published on GitHub under MoonshotAI, with third-party coverage putting it at 3.2k stars in July 2026. Independent guides multiplied through 2026, which I read as real adoption.

Strengths
#

The $19 entry ($15 billed annually) undercuts Claude Pro and Cursor Pro while including a 1M-context frontier-class model at the mid tier.
The yen ladder starts at ¥49 per month, well below the $19 international entry at any recent exchange rate.
Anthropic-compatible endpoints mean Claude Code and OpenCode work without new tooling.

Cautions
#

Coding quota shares one pool with chat, Deep Research, and other Kimi features, so non-coding usage drains your coding allowance.
The plan ladder moved twice in 2026, so quota rules at signup may not persist.
A February 2026 checkout experiment offered personalized first months between $0.99 and $11.99, so two subscribers on one tier can pay different prices.
One four-month user quit, reporting that higher tiers bought less inference per dollar, not more.

Pricing
#

On kimi.ai: Moderato $19, Allegretto $39, Allegro $99, Vivace $199 per month, annual effective $15/$31/$79/$159 (as of 2026-09-26). On kimi.com: Andante ¥49, Moderato ¥99, Allegretto ¥199, Allegro ¥699 per month (as of 2026-09-26). Under the new ladder, Go has no coding quota, Plus and above include Kimi Code, and Pro and above unlock K3 at 1M context. Extra Usage is pay-as-you-go overflow with a ¥25 minimum top-up, generally non-refundable.

Price history
#

Date	Plan	Change	Source
2026-07-17	Moderato to Vivace	Listed $19/$39/$99/$199 with time-limited promos at $15/$31/$79/$159; Kimi and Kimi Code separation announced	https://www.digitalapplied.com/blog/kimi-code-k3-hands-on-setup-plans-cache-2026
2026-09-26	Same tiers	Promos became standard annual pricing; new plans released at unchanged prices, weekly window dropped for new members	https://www.kimi.ai/help/membership/membership-pricing

Compared to
#

MiniMax Coding Plan (../minimax-coding-plan/index.md) meters tokens across modalities instead of time windows, cheaper at the top end but carrying 2026 billing-change baggage. NanoGPT (../nanogpt/index.md) fits when open-weight breadth matters more than one lab’s flagship. Claude Code or Cursor remain the picks when reliability per hour beats price per hour.

Bottom line
#

Recommended for cost-sensitive daily coders who want K3’s 1M context at $15 to $39 per month and accept quota churn. Not for teams needing contract-stable plan terms after two restructures in one year. I will claim something arguable: the ¥49 China-market ladder proves this product profits at roughly a third of the international price, which makes $19 a regional tax, not a cost.

Changes
#

2026-09-26 - Created.

References
#

https://www.kimi.com/code/en - official product page; surfaces and the K3 1M-context claim.
https://www.kimi.com/code/docs/en/kimi-code/membership.html - membership docs; new ladder (Go/Plus/Pro), weekly window removal, Extra Usage terms.
https://www.kimi.ai/help/membership/membership-pricing - dollar pricing $19/$39/$99/$199 monthly, $15/$31/$79/$159 annual, as of 2026-09-26.
https://www.kimi.com/en/help/membership/membership-pricing - yen pricing ¥49/¥99/¥199/¥699 and shared credit pool rules, as of 2026-09-26.
https://www.digitalapplied.com/blog/kimi-code-k3-hands-on-setup-plans-cache-2026 - July 17, 2026 snapshot of listed vs promo prices and the separation banner.
https://aihackers.net/value/deals/kimi-haggle/ - documented February 2026 first-month offers of $0.99 to $11.99, marked historical.
https://www.noemititarenco.com/blog/kimi-code-plan-after-4-months-of-use-honest-review/ - skeptical four-month user review and cancellation rationale.

MiniMax Coding Plan

tom@tomrochette.com (Tom Rochette) — Sat, 26 Sep 2026 00:00:00 +0000

The product sold as the MiniMax Coding Plan became the MiniMax Token Plan on June 1, 2026, a usage-based subscription on platform.minimax.io that meters the M-series models through a Subscription Key. MiniMax converted a liked flat coding plan into token billing overnight, without notice, and the resulting trust deficit is still the product’s largest liability. Facts below verified as of 2026-09-26.

What it is
#

MiniMax (a Hong Kong-listed Chinese AI lab) sells Plus, Max, and Ultra tiers, with quotas shared across text, image, and speech models including M3 and M2.7. Usage runs through a Subscription Key separate from the pay-as-you-go API key, and Anthropic-compatible endpoints let Claude Code-style harnesss plug in directly. Quota uses a 5-hour rolling window plus a weekly window, with no carryover. Credits packages at 1,000 per dollar cover overflow once subscription quota runs out.

Status
#

Active at company scale but turbulent: H1 2026 revenue hit $117M (up 283% year over year), ARR passed $800M by August, and July token consumption ran 20 times January’s. The stock fell more than 80% from its HK$1,330 peak to HK$298 by September 21, 2026, and JPMorgan downgraded it in June, citing M3’s lack of pricing power. The June 2 apology and compensation package followed developer complaints after the billing switch, with a refund portal on June 3.

Strengths
#

One token pool covers code, images, and speech, which suits mixed-agent workloads other coding plans exclude.
Credits overflow means a task does not hard-stop at the quota edge, it just starts spending Credits.
Even the top tier ($132) undercuts Western $200 plans by more than a third, on an open-weight model family.

Cautions
#

An open GitHub issue from June 3, 2026 documents quota draining with zero API calls and an unauditable cache discount, with Plus reportedly exhausted in 4-5 hours of agent work.
The FAQ states the plan does not support refunds and that MiniMax itself recommends pay-as-you-go for production.
Dynamic rate limiting tightens weekday peak hours (15:00-17:30), cutting Plus to roughly 3-4 concurrent agents.
Plus rose from $20 to $22 within three months of launch, an early sign that sticker prices here move fast.

Pricing
#

As of 2026-09-26: Plus $22/month, Max $55/month, Ultra $132/month, all with 5-hour rolling and weekly windows. Typical peak-hour capacity is 3-4 (Plus), 4-5 (Max), and 6-7 (Ultra) agents. Credits cost $5, $25, or $100 at 1,000 per dollar, valid 365 days, and a 10% referral discount applies at checkout.

Price history
#

Date	Plan	Change	Source
2026-06-01	Coding Plan -> Token Plan	Flat per-task Coding Plan replaced by usage-based Token Plan at the M3 launch, without advance notice	https://news.aibase.com/news/28699
2026-06-03	Token Plan Plus	Observed at $20/month, $200/year	https://github.com/MiniMax-AI/MiniMax-M2.7/issues/47
2026-09-26	Token Plan Plus/Max/Ultra	Listed at $22/$55/$132, with Plus up $2 against June	https://platform.minimax.io/docs/guides/pricing-token-plan

Compared to
#

Kimi Code (../kimi-code/index.md) sells time-windowed quota of one lab’s models and fits steady daily coders better than bursty multimodal users. NanoGPT (../nanogpt/index.md) has no quota windows at all, removing the mid-sprint lockout risk at the cost of per-token metering. MiniMax’s own pay-as-you-go API is the right choice for production, by the company’s own documentation.

Bottom line
#

Recommended for tinkerers who want cheap, broad token volume across modalities and will watch the usage bar. Not for anyone needing auditable billing, refunds, or production reliability. I will claim something arguable: until the quota denominator and per-call billing history are published, the usage bar is a trust position rather than a meter, and the June apology does not change that.

Changes
#

2026-09-26 - Created.

References
#

https://platform.minimax.io/docs/guides/pricing-token-plan - current tiers $22/$55/$132, quota windows, agent counts, Credits packages, as of 2026-09-26.
https://platform.minimax.io/docs/token-plan/faq - Subscription Key mechanics, no-refund policy, production guidance, peak-hour rate limiting.
https://github.com/MiniMax-AI/MiniMax-M2.7/issues/47 - critical billing-bug report: passive quota drain, unverifiable cache discount, $20 Plus in June 2026.
https://news.aibase.com/news/28699 - the June 2, 2026 apology, compensation package, and Coding Plan to Token Plan switch.
https://eu.36kr.com/en/p/3993817731234817 - critical business reporting: developer backlash, JPMorgan downgrade, market value collapse, H1 2026 financials.

OpenCode Go

tom@tomrochette.com (Tom Rochette) — Sat, 26 Sep 2026 00:00:00 +0000

OpenCode Go is a $10/month subscription from the OpenCode (Anomaly) team that bundles access to a curated set of open coding models, usable from OpenCode or any compatible coding agent. Facts below verified as of 2026-09-26.

What it is
#

Go sells model access, not an editor or a harness. You subscribe, copy an API key, and point any agent at OpenAI-compatible, Anthropic-compatible, or Responses endpoints under opencode.ai/zen/go/v1. The lineup is 32 open-weight coding models as of 2026-09-26 (Grok 4.7/4.6, GLM-5.3 family, Kimi K3, Qwen3.x, DeepSeek V4, MiniMax M3, MiMo, GPT 6 Luna), and only one member per workspace can hold a subscription.

Status
#

Active and changing fast. The OpenCode repo shows about 208K GitHub stars as of 2026-09-26, the Go docs were last updated 2026-09-25, and the plan has grown from a three-model Beta in March 2026 to 32 models. Churn is constant: GPT 6 Luna arrived 2026-09-22 and Space Bunny Free appeared as a limited-time unlimited model. Go converts $10 into up to $60 of metered usage per month, and that 6x ratio is the entire value proposition.

Strengths
#

The $60 monthly usage value costs $10, up to 6x leverage if you max it.
Models are benchmarked for agentic coding before inclusion, per the team.
Works with any agent; Claude Code, Codex, Pi, jcode, and Kilo Code CLI are validated clients.
Dollar-denominated windows make the bill predictable, free models stay available after limits, and “Use balance” falls back to your Zen balance instead of hard-stopping requests.
Most models run with zero retention and no training on your data.

Cautions
#

This is no longer general API access.
Since September 2026, clients must send coding-agent traffic, a real user agent, and an x-opencode-session header, which the community reads as whitelisting non-agent use.
Limits are dollar-value, so flagship tiers (Grok, Kimi K3, GLM-5.3, GPT 6 Luna) include only $15/month, with a $3 5-hour window.
The catalog and limits change monthly, the $5 first-month promo is gone, and usage estimates assume heavy prompt caching, so your real request counts will not match.
Muse Spark Contributor models train on your prompts and are region-limited.

Pricing
#

$10/month, cancel any time, as of 2026-09-26. Each model carries a monthly dollar limit (mostly $60, some $30, $15 for flagships), with windows at 20% per rolling 5 hours (so $12 on a $60 model) and 50% weekly ($30). Exceeded limits fall back to free models, or to your Zen balance if you enable it. Top-ups draw on the shared Zen balance, where card fees are passed at cost (4.4% + $0.30).

Price history
#

Date	Plan	Change	Source
2026-03-12	Go (Beta)	Baseline: $10/month, three models (GLM-5, Kimi K2.5, MiniMax M2.5), $60 usage value	https://help.apiyi.com/en/opencode-go-subscription-worth-it-review-en.html
2026-08-24	Go	First-month $5 promo removed, flat $10/month	https://codingplan.org/en/plans/opencode-go
2026-09-02	Go	$15 monthly limit became the standard for several flagship models (community-reported)	https://www.reddit.com/r/opencode/comments/1vo9j8l/opencode_go_15_is_now_the_standard/
2026-09-24	GLM-5.3-Flash	Monthly limit doubled to $60 (docs confirm $60 by 2026-09-25)	https://www.reddit.com/r/opencode/comments/1wc1cwe/glm_53_flash_now_gets_twice_the_limits_on/

Compared to
#

OpenCode Zen (../opencode-zen/index.md) is the sibling pay-per-use gateway: no subscription, huge catalog including Claude and GPT, but per-token prices that can exceed OpenRouter’s. Choose Go when you want a capped, predictable bill for open models. OpenRouter (../openrouter/index.md) has hundreds of models with routing and fallback at passthrough prices plus a 5.5% credit fee; choose it when you need frontier or long-tail models Go does not carry. Direct DeepSeek API beats Go unless you fully use the $60 cap, per a community breakdown that pegs Go at 33% cheaper only at full usage.

Bottom line
#

Recommended for heavy users of open coding models who will actually burn the $60 monthly value. Not for anyone needing general-purpose API access, closed frontier models, or light usage. My disagreeable claim: at roughly 50% usage Go is about break-even with direct API pricing, so most subscribers are partly paying for slack they never consume.

Changes
#

2026-09-26 - Created.

References
#

https://opencode.ai/go - product page, $10/month framing, model snapshot, GitHub star count (200).
https://opencode.ai/docs/go - full model/limit table, 20%/50%/100% windows, validated clients, header requirements (200).
https://codingplan.org/en/plans/opencode-go - $5 first-month promo removal on 2026-08-24, catalog churn timeline (200).
https://help.apiyi.com/en/opencode-go-subscription-worth-it-review-en.html - March 2026 Beta state, dollar-window mechanics, critical take (200).
https://www.reddit.com/r/opencode/comments/1w9pyvq/opencode_go_is_no_longer_general_api_access/ - community reaction to session-header requirement (direct fetch 403; content read via search).
https://www.reddit.com/r/opencodeCLI/comments/1tqx9u1/why_opencode_gos_deepseek_v4_pro_is_33_cheaper/ - full-usage price math vs DeepSeek API (content read via search).

coding-subscription on tomrochette.com

Kimi Code

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Price history
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#

MiniMax Coding Plan

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Price history
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#

OpenCode Go

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Price history
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#

coding-subscription on tomrochette.com

Kimi Code

What it is #

Status #

Strengths #

Cautions #

Pricing #

Price history #

Compared to #

Bottom line #

Changes #

See also #

References #

MiniMax Coding Plan

What it is #

Status #

Strengths #

Cautions #

Pricing #

Price history #

Compared to #

Bottom line #

Changes #

See also #

References #

OpenCode Go

What it is #

Status #

Strengths #

Cautions #

Pricing #

Price history #

Compared to #

Bottom line #

Changes #

See also #

References #

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Price history
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Price history
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Price history
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#