How the M Plan Has Powered My Three Journals

I run three WordPress journals — one on Chinese history, one on Jewish textual scholarship, one on science and technology — and I have been on a MiniMax subscription since the Token Plan era. M Plan, the relaunched version that went live on 30 September 2026, is the first subscription I have used that genuinely changed how each of those sites gets made. The change is not just price. The change is that one subscription covers a chat model, image generation, audio generation, an Anthropic-protocol bridge into Claude Code and Cursor, and the MiniMax Code agent that drives much of the work end-to-end. This post is a personal read on what M Plan has actually been worth across the three sites, written for anyone considering whether the bundle pays off in practice and not just on paper.

What M Plan replaces for me

The honest comparison is not between M Plan and ChatGPT Plus or M Plan and Claude Pro. The honest comparison is between M Plan Explore at $55 a month and the basket of subscriptions and credit accounts I had before it.

For most of 2025 and the first half of 2026, my monthly AI bill ran roughly as follows. ChatGPT Plus at $20. Claude Pro at $20 (or $17 on annual). Anthropic API credits for whatever Cursor was burning through, somewhere between $20 and $40 a month. A separate image-generation credit bucket, billed in dollars per render, usually $20-30 a month. An occasional Veo 3 or Sora 2 video render, also per-render. Music generation, where I had it, was a separate tier. A separate web-search subscription for the rare deep-research task where I wanted more than the chat model’s own search. The sum settled in the $120-150 range most months, and there was no integration story between any of them — each vendor had its own chat surface, its own settings, its own memory, its own quirks.

Switching to M Plan Explore at $55 a month, or about $46 on annual billing, replaced most of that. The chat model is M3.1 Flash Preview on all three sites, with web search bundled. The image generation is included, on a shared credit pool. The audio generation is included. The H3 video is included for the occasional illustrative clip. The Subscription Key replaces the Anthropic key in Cursor, Claude Code and Cline. The MiniMax Code agent desktop app replaces the IDE-style coding surface I used to drive from a separate tool, all on the same quota. The bill is now a single line on the monthly statement instead of five or six.

That is the value argument in one paragraph. The rest of this post is the same value argument told site by site, because the three journals I run are different enough that M Plan’s bundle buys different things on each of them.

The three journals in one paragraph each

The science and technology journal is the place this post is appearing on, and it is the simplest of the three in terms of what M Plan does for it. The workflow is: take a question, build a draft using M3.1 with web research and source verification, generate a featured image in the site’s pale-blue-and-cream house style via the matrix image tool, write the post into a markdown source of truth, build a post.html from the markdown, run a deploy script that pushes the post and the image up to the WordPress instance over SSH, set the category, set the featured-image meta, publish, and purge the cache. The whole pipeline runs end to end on a single M Plan key. The chat model drafts the section bodies, walks the source list, formats the table, suggests the headline. The image tool gives me three featured-image variants per post; I pick one and ship it.

The Chinese-history journal is the most bespoke of the three because the illustrations are mine. I paint the featured images myself rather than generate them, which means the AI part of the workflow is text and verification rather than image. The drafts go through several layers of cross-source verification in Chinese and English before I publish, because the editorial standard on the site is high and the readers notice when a source is paraphrased carelessly. M3.1’s web search helps me pull from the Chinese-language sources, and the Browser skill means I can hand the model a specific URL and ask it to quote, paraphrase or check a passage. The plan’s value on this site is the text quality and the verification reach, not the image.

The Jewish-textual-scholarship journal sits between the other two. The visual house style is specific enough that the model needs a stable reference and gentle settings; the model gets there with the right configuration. The text work on that site is heavy on Hebrew and Yiddish quotation, careful translation, and close attention to manuscript references. M3.1 handles the Hebrew and Yiddish acceptably for editorial use, with a human proofreading pass on every translation before it ships. The agent’s Browser skill is again useful for the rare manuscript-image inspection.

How M Plan shows up in the nekud workflow (this site)

This journal is the most metric-heavy of the three, and the place where the cost case for M Plan is straightforward to make.

Research is the first thing. Each post starts with a question and the model returns a standfirst draft (about 100–130 words), a numbered key-findings list, a section outline, and a working source list with numbered references. The source list is what the model is best at in this workflow. The model is good at remembering the canonical reference for a given claim, less good at the obscure reference, and pretty good at flagging which sources are load-bearing and which are decorative. The web research is built into the same model call because M Plan includes web search in the M3.1 token budget. There is no separate research tool to buy.

Drafting is the second thing. Each section is its own paragraph block with one H2 ahead of it. The model writes at a measured, source-led register with footnoted claims; the closest analogue in published journalism is the science section of a serious newspaper. M3.1 is good at that register for the kind of post this site runs. The model is also good at recognising when a paragraph is doing too many jobs and asking it to be split. The refusal rate on substantive claims is low, the rewriting rate is high, and the failure mode I see most often is the model paraphrasing a source rather than quoting it. The post-publish step is always a human pass on the paraphrase.

The featured image comes third. For this site, I generate via the matrix image tool in three variants at the house palette (pale blue, grey, cream, iris, with a single ochre accent), and pick. The generation is included in the M Plan image credit pool. The image is processed locally into 1600×1000 with cream padding, mirrored to the OneDrive backup directory alongside the post source, uploaded to the site over SCP as part of the same deploy script that pushes the post.html. The whole thing is one subscription and one credential chain.

The deploy is the fourth thing, and is the closest the workflow comes to “agent” in the canonical sense. The MiniMax Code agent reads a deploy script, runs it, watches the SSH responses, recovers from the unexpected `wp post create` output format, and reports a clean summary. The Cursor and Claude Code work also runs on the same M Plan Subscription Key, with the same M3.1 underneath. I have not had to swap the model underneath the agent in any session since M Plan came out.

For this site specifically, M Plan Explore is the right tier. The image and audio credit draw is modest, the video draw is occasional, and the cross-tool key drop into Cursor and Claude Code is what closes the gap. Build is overkill; Go is sufficient for the chat work but not for the bundled image budget when I publish at the rate this site ships.

How M Plan shows up in the huaxia (Chinese history) workflow

The Chinese-history site is the place where M Plan’s value is the most depends-on-the-question.

The text work is heavy. Every post is drafted from primary sources where possible, with Chinese-language citations when the English-language literature is too thin. The model is, in my experience, reliable at reading modern Chinese and at translating modern Chinese into English with editorial register. It is less reliable on classical Chinese (文言文), where the canonical translations are sometimes the only safe output, and a careful post will quote rather than translate. The Agent’s Browser tool is what closes the gap on obscure citations: I can hand the agent a URL to a specific page on a Chinese university press or archive and ask it to confirm or quote a passage, and the model comes back with a result close enough to be useful.

The reference integrity on the Chinese-history site is non-negotiable. The model’s web research sometimes cites a source that does not exist or paraphrases more confidently than the underlying citation justifies. The post-publish workflow on this site therefore includes a careful human pass on every claim that has a Chinese-language citation, and a smaller but still real human pass on every claim that has an English-language citation. M Plan’s value here is not the absence of risk; it is the throughput. The same work without M Plan would be roughly twice the time per post.

The illustrations on this site are my paintings, not AI-generated. The workflow for images is: draft the post, settle the text, paint a featured image, scan it at high resolution, process it into 1600×1000 with the standard house palette, deploy. The painting is the slowest single step; the AI pipeline helps in that it lets me focus the text on text rather than on visual mock-ups. M Plan’s image credit budget is, on this site, mostly unused — and that is fine, because the value is in the bundled text and verification rather than the bundled image.

The other thing this site uses from M Plan is the audit trail. The model retains session memory of the editorial choices on a given post, so when I come back to a series of posts a month later, I can ask the model which direction the editorial standard has drifted and have it flag the relevant sentences. The audit work is what makes consistency across a long-running series possible, and the audit work is what the Token Plan / M Plan era bought me.

How M Plan shows up in the masoretes (Jewish text) workflow

The Jewish-textual-scholarship site is the place where M Plan is doing the most volume per post.

The text work on this site is heavy on Hebrew and Yiddish quotation, careful translation, manuscript reference and the editorial use of community-standard romanisation. M3.1 handles modern Hebrew acceptably well, with the standard caveat that theological and poetic language needs a human pass; Yiddish is weaker and I check every line; classical Hebrew (especially Biblical Hebrew) is reasonable for editorial paraphrasing but not for academic citation. The general pattern is that the model gives me a serviceable first draft and the human pass lifts it to the publication standard.

The illustrations on this site use the AI pipeline. The house style is a specific watercolour-over-ink look, not pure ink and not the pale-blue-and-cream style of nekud. The right pipeline is the matrix image tool with a fixed reference image and a moderate reference-strength value, configured per session; the output is two or three variants per post, of which I pick the strongest. The relevant engineering detail is that the model is good at the watercolour look but only when given a stable reference; without one, the output drifts to photoreal. The M Plan image credit budget covers the volume for a site that publishes less often than the science journal but publishes with more elaborate images.

The cross-tool key drop is also useful on this site. The Hebrew and Yiddish text work is often quoted from a primary source that I keep open in a browser tab. The agent’s Browser skill lets me hand the model a URL to a specific manuscript image or a specific chapter at Sefaria and have it extract, translate or comment on the passage. The model is reliable enough at that workflow to be worth the credit cost.

The thing that ties the three sites together on M Plan is the Subscription Key. One login, one billing line, three journals, three voices, and the same model underneath.

Workflow patterns that work across all three sites

Four patterns recur across the three journals, and they are the patterns that justify M Plan over a cheaper subscription on a single site.

The first is the source-first drafting pattern. Every post starts with a numbered reference list rather than with a topic sentence. The model works back from the references to the prose, which forces the verification to happen up front rather than as a post-publish fix-up. The pattern works on M3.1 and would not work on a smaller-context or weaker-citation model.

The second is the cross-tool key drop. The M Plan Subscription Key goes into Claude Code when I want a particular style of agent behaviour, into Cursor when I want IDE-style editing on a long-running codebase, into Cline when I want the same model inside an extension-based surface. The three tools get the same model and the same quota underneath, and the human-facing surface changes. The pattern is what makes it cheap to switch between three different agent surfaces without paying three different API bills.

The third is the image-variants pattern. For every post, the matrix image tool returns three variants in the site’s house palette. The human picks one. The pattern is what makes it possible to publish at the volume the science journal runs at without each post demanding a bespoke image brief from me. The pattern depends on a stable reference image and a consistent prompt frame, and the M Plan pool is what covers the credit cost.

The fourth is the audit-trail pattern. The post-publish memory on the model is what lets me ask “is this post in the same register as the last six?” or “have I drifted on the citation style?” or “what’s the dominant source for this series?” with the model answering from the session history. The audit pass is a few minutes per post, and it is what makes consistency across a long-running series possible.

Honest limits of M Plan on this kind of work

There are two honest limits worth flagging.

The first is that M3.1 Flash Preview is not the absolute frontier reasoning model. The deepest pass on a difficult translation, the most cautious citation policy on a contested historical claim, the most careful editorial read on a complex theological argument — these still want a top-tier reasoning model underneath, and M Plan does not sell access to one. The pragmatic answer is to do those tasks by hand rather than pay separately for a more capable model on a per-call basis. M Plan makes the everyday work very cheap; the cost of the exceptional work is the cost of doing it by hand.

The second is that the shared credit budget is a constraint, not just an economy. The image budget is shared with the text budget; the audio budget is shared with both. A week’s worth of heavy image generation means the text budget shrinks accordingly, and the same is true if I spend the week generating audio or video rather than text. For a low-volume week this is invisible. For a high-volume week with several posts and several image variants per post, the constraint shows up.

Neither limit is enough to make M Plan the wrong choice. For most of what I do across the three sites, the bundle is the cheapest path. For the things outside the bundle, the cost is real but is the cost of doing things carefully rather than cheaply.

Bottom line

M Plan landed at the right moment. The Token Plan had been good value; M Plan is materially better value because the capabilities it bundles are the capabilities I was already buying separately. The three sites it powers are different — a science journal, a Chinese-history journal, a Jewish-textual-scholarship journal — and the bundle works differently on each, but it works on all three. The chat model is the constant. The image and audio credits are shared. The Subscription Key drops into the rest of my toolchain. The agent runs the deploy scripts. The bill is one line a month.

For anyone running a single writing workflow — a blog, a newsletter, a publishing pipeline of any kind — M Plan Explore at $55 (or about $46 on annual billing) is the simplest version of the value argument in this post. The case gets stronger when the single workflow turns into several, which is when the shared capability bundle starts to mean what it claims to mean.

References

1. MiniMax, “FAQ - Models,” MiniMax API Docs, accessed 14 October 2026, https://platform.minimax.io/docs/m-plan/faq
2. minimax-ai.chat, “MiniMax M3.1-Flash-Preview: Release, Access & Benchmarks,” minimax-ai.chat Models, October 2026, https://minimax-ai.chat/models/minimax-m3-1-flash-preview/
3. Margrop, “Dual-Track Adoption: Native Client & Third-Party Agents,” Margrop Blog, October 2026, https://blog.margrop.net/en/post/minimax-m-plan-launch-and-agent-integration-guide/
4. MiniMax, “Pricing - MiniMax M Plan,” MiniMax, accessed 14 October 2026, https://minimax.io/m-plan