MiniMax AI (August 2026 Review)

MiniMax AI is worth trying for routine coding and agent work when you can check the result cheaply: M3.1 Flash costs $0.15 per million input tokens and $0.60 per million output tokens. It is the better buy than paying more for Claude or GPT when it passes your existing tests without extra retries or review; keep the incumbent when difficult changes, costly mistakes, or an already-paid subscription erase that saving. MiniMax's current API rates.

What MiniMax AI Actually Is

MiniMax is an AI model company with several products that share models and billing but do not behave like one unified assistant. M3.1 Flash is the latest language model for coding and agents, software that uses tools to complete tasks. M3 remains available. MiniMax Code is the coding and work app; H3 generates video with audio; Speech 2.8 handles voice; Music 3.0 remains a separate music model. The creative canvas previously called Hub now lives at MiniMax Design, where the official Hub address redirects. Those products serve different jobs; Flash does not replace the media models or the creative app.

Since October 19, 2026, M3.1 Flash has been live on the MiniMax API, M Plan, and MiniMax Code. It replaces the retired M3.1 Flash Preview model name, and MiniMax Code now presents M3.1-Flash as its default model on the download page, superseding the earlier M3 description here. M3 remains a separately listed API model. The launch announcement says Flash's weights will be released later and a Pro model is planned; neither is a reason to assume a downloadable Flash deployment is available today. Launch announcement, model availability, MiniMax Code.

MiniMax AI vs the Alternatives at a Glance

MiniMax's strongest case for this reader is inexpensive, repeatable coding work. Its media stack is useful additional coverage. Prices and model availability were verified against the makers' live pages on October 20, 2026.

The choice flips on the work you accept, not the breadth of the model catalog. A founder with a queue of small fixes does not need to use H3 or Speech to justify Flash. A buyer who already gets those fixes done within a Claude or ChatGPT subscription should compare the extra cost of adding MiniMax against the subscription they would actually cancel or downgrade.

Who MiniMax Is For and Who Should Skip It

MiniMax fits technical builders and creative operators who can turn low unit prices into repeatable workflows. It is a weaker default for buyers who judge a tool mainly by the smoothness of one consumer app.

Choose MiniMax for high-volume model and media work

A solo technical builder can use M3.1 Flash through MiniMax Code or a supported coding client for bounded bug fixes, test additions, routine refactors, and document-to-tool workflows. These are good evaluation candidates because a test suite or a simple acceptance checklist can expose bad output before it reaches customers.

A funded founder or mid-market CTO should look at MiniMax when repository context and cost control matter together. MiniMax's current docs explicitly list a 1,000,000-token context window for Flash and M3; context is the material a model can consider in a request, not a guarantee that it will reason correctly over every file. The right trial is one expensive, repeatable route with an acceptance rubric. Text-generation docs.

A creative operator should still shortlist H3 when reference-driven video and native audio are central. Flash changes the coding choice; it does not turn the language model into H3, Speech, or Music.

Choose ChatGPT for one general-purpose workspace

ChatGPT remains the simpler purchase when you already use its workspace and included coding tools. Go is $8 monthly and Plus is $20 monthly; OpenAI lists Codex access on the web, CLI, IDE extension, and iOS under Plus. A MiniMax subscription is not automatically cheaper: its regular Go renewal is higher than ChatGPT Plus. The potential saving comes from Flash's metered API usage or the amount of useful work your plan covers. OpenAI pricing.

Choose Claude when coding judgment matters more than unit price

Keep Claude when it resolves your difficult tasks with fewer failed changes or reviewer minutes. Pro costs $20 month-to-month or $200 paid annually, equivalent to about $16.67 monthly; Anthropic rounds the annual monthly figure to $17 on its page. M3.1 Flash deserves a measured trial for routine volume, but there is no basis here to declare a universal quality winner. Compare accepted patches, retries, tool errors, and review time on the same repository. Claude pricing.

Choose Runway when the deliverable lives in a video editor

Runway is the clearer fit when a creative team wants a video-first product, model choice, and a credit workflow. Standard costs $15 monthly or $12 monthly with annual billing and includes 625 monthly credits, which Runway equates to 52 seconds of Gen-4.5. Runway pricing. MiniMax H3 can be less expensive per generated second, but raw generation cost does not buy a mature review and editing process. The broader AI video generator comparison is the better map when video is the only buying question.

The Four MiniMax Capabilities That Matter

MiniMax earns a place on the shortlist through four concrete workflows: long-context agent work, reference-driven video, coordinated creative production, and low-cost voice generation.

1. MiniMax M3.1 Flash and Code make routine agent work inexpensive

M3.1 Flash takes over the lead coding recommendation here. MiniMax documents image and video input alongside text, tool use, and adjustable thinking depth. Its effort setting defaults to max; lower settings can reduce reasoning output and latency on simpler work, while thinking cannot be switched off. Start with the files the task needs and adjust effort against your acceptance checks. Model and thinking documentation.

M3 remains available, including its documented hosted context up to 1M tokens with a guaranteed minimum of 512K. Keep an existing M3 integration until the Flash comparison shows a benefit; the release does not announce an M3 shutdown. M3 model page.

MiniMax reports 73.7% on DeepSWE v1.1 and 61.7 on OSWorld V2 for M3.1 Flash; these are MiniMax's own benchmark claims, not our tests. Release results.

MiniMax Code is the ready-made surface for Flash. Its current page shows M3.1-Flash selected and describes access to local files, tools, a browser, skills, and plugins. M Plan includes Code without another subscription; a separate Subscription Key connects the plan to supported clients such as Claude Code, Codex, Cursor, or OpenCode, sharing the same usage limits. Using the Claude Code client with MiniMax does not mean using a Claude model. Code, M Plan FAQ.

Here is the workflow that makes the economics measurable:

1. 1

   Choose one repository task

   Use a bounded issue with existing tests, a known failure, and a reviewer who can accept or reject the result. A repository-wide migration is too broad for the first comparison.

2. 2

   Keep the setup constant

   Connect M3.1 Flash to the same supported coding client used for the incumbent model. Preserve tool permissions, test commands, timeouts, and the acceptance rubric. Record the thinking setting as well.

3. 3

   Price the complete request

   For an illustrative request with 200,000 uncached input tokens and 20,000 billed output tokens, Flash costs $0.042: $0.030 input plus $0.012 output. The same request costs $0.084 on M3 Standard, so Flash halves the token bill. At 700,000 input and 30,000 output, Flash costs $0.123 versus M3's $0.492. These calculations use the current rates and exclude tool charges and additional requests in an agent run. API pricing.

4. 4

   Count accepted outcomes

   Record model spend, elapsed time, retries, failed tool calls, changed files, and reviewer minutes. Divide total cost by accepted tasks. Cheap output that needs repair is not cheap work.

2. MiniMax H3 combines references, video, and native stereo audio

MiniMax H3 is the standout reason to look beyond the chat and coding market. It accepts a multimodal brief and generates a 4 to 15 second video at 24 FPS with 32 kHz stereo audio. The reference mode allows up to 9 images, 3 videos, 3 audio clips, and 12 files total, which is enough to specify a subject, visual language, motion reference, and voice reference in one job. MiniMax H3 model card.

A product-launch shot can be structured without pretending that a single text prompt will carry the whole brief:

1. Provide the product stills as reference images and a short movement reference as video.
2. Add the narration or voice character as an audio reference.
3. Choose a 4 to 15 second duration and a supported ratio. Text-to-video requires an explicit ratio; image-to-video follows the input image.
4. Generate at 768P for the selection pass, then move the accepted shot to 2K. The supported durations, modes, and ratio rules are in the H3 API reference.

The meter is direct. A 10-second 768P output costs $0.80. The same duration at 2K costs $1.30, and a 15-second 2K output costs $1.95. A 10-second 2K generation that also consumes a 10-second reference video costs $2.60 because MiniMax bills both output duration and reference-video duration at $0.13 per second. Audio references are free, the first 5 image references are free, and later images cost $0.04 each. These are calculations from the current H3 rates.

MiniMax links ComfyUI workflows for its H3 Base checkpoints from the official model card. That makes the 768P Base checkpoints usable in a node workflow, but it does not make the full hosted system local. The distinction becomes the central limitation later in this review.

3. MiniMax Hub, now Design, coordinates a creative brief across media

MiniMax Design, the current destination of the official Hub address, continues the creative-canvas role. It turns separate models into one production workflow. Its public workflow starts with a brief, breaks the work into script, storyboard, image, video, music, voiceover, and editing tasks, and can launch copy, image, video, and audio agents in parallel.

The practical workflow is a small campaign, not a one-shot masterpiece. Give Design a product brief, target audience, approved claims, reference assets, and deliverable list. Let the copy agent draft the script and captions, the image agent establish visual candidates, the video agent build short sequences, and the audio agent prepare voiceover or music. Review each stage before the next model spends more credits. Export selected assets to the professional tool that will finish them.

That last review step matters. Agent orchestration reduces tool switching, but it does not remove art direction, rights clearance, brand checks, or edit decisions. Design is most valuable when it turns a repeated campaign recipe into a reusable skill. It is less compelling when the team already has a mature editor-centered workflow and only needs one generation model.

4. MiniMax Speech 2.8 makes narration costs easy to predict

MiniMax Speech 2.8 supports native sound tags and voice cloning from a 10-second sample according to its release page. The synchronous API requires text shorter than 10,000 characters; the asynchronous route accepts up to 1M characters. Those limits belong to the speech APIs and are unaffected by the Flash launch. Speech API, asynchronous pricing and limit.

For a 50,000-character training narration, speech-2.8-turbo costs $3 and speech-2.8-hd costs $5. A rapid voice clone adds $1.50 on first use, while a designed voice costs $3. The workflow is straightforward: secure a properly authorized voice sample, generate a short approval clip, lock pronunciation and pace, then send the full script through the asynchronous endpoint. These examples use the current speech rates.

Music 3.0 remains part of the wider MiniMax ecosystem, but access changed independently of Flash: from August 20, 2026, paid music and lyrics APIs stopped accepting new users, while existing paying users can continue. The free music APIs were discontinued; MiniMax directs new users to Audio or the Music 3 model weights. Image-01 remains $0.0035 per image, so 100 outputs cost $0.35. Selection, rights, consistency, and finishing remain the expensive parts of the outcome. Music access notice and image rates.

MiniMax Pricing: Every Current Tier

MiniMax pricing is inexpensive at the model layer and complicated at the account layer. The public system has M Plan subscriptions, separate prepaid credits, pay-as-you-go API balance, Audio Subscriptions, and legacy Video Packages. Every rate below was checked on October 20, 2026. The old minimax.io/pricing address now redirects to the API pricing overview; subscriptions live on the separate M Plan page.

M Plan subscriptions

These are regular renewal prices in USD, not first-month promotional prices. MiniMax M Plan.

The monthly selector currently offers 50% off the first month: Go $11, Explore $27.50, and Build $66, then the regular renewal price. MiniMax says that offer ends October 31, 2026 (UTC+8). The annual amounts above are paid up front; the effective monthly values are annual totals divided by 12. Current offer and billing controls.

Token Plan is closed to new purchases. Existing Plus, Max, and Ultra subscribers can keep their original price and benefits while auto-renewal remains on, and their text model now includes Flash. An upgrade moves them onto M Plan terms and cannot be reversed back to Token Plan. The current migration notice says legacy Token Plan does not include H3; Explore and Build do. Check the account's retained benefits before giving up a legacy subscription. Token Plan migration notice.

The prepaid credit packs remain Starter at $5 for 5,000 credits, Growth at $25 for 25,000, and High Volume at $100 for 100,000. Purchased credits remain valid for one year. They are not the same thing as the standard pay-as-you-go API balance. Credit packs.

The quota language deserves as much attention as the prices. Non-video usage is controlled by 5-hour and weekly windows; video uses only the weekly window. Unused included quota does not carry into the next billing cycle, and peak-hour capacity can tighten. MiniMax recommends pay-as-you-go for production. The current tiers advertise relative usage rather than the old billions-of-tokens, simultaneous-agent, or daily-video estimates. M Plan FAQ.

M3.1 Flash and M3 pay-as-you-go API

M3.1 Flash uses flat pricing of $0.15 input, $0.60 output, and $0.015 cached input per million tokens. The published table does not apply M3's long-input price step to Flash. A cache read reuses previously processed input at the lower rate; it does not discount new input or output. API pricing.

M3 remains separately priced. At Standard priority and no more than 512K input tokens, it costs $0.30 per million input tokens, $1.20 output, and $0.06 prompt-cache reads. Above 512K input, those rates become $0.60, $2.40, and $0.12 respectively. M3 rates.

M3 Priority is 1.5 times Standard. At or below 512K input, it costs $0.45 input, $1.80 output, and $0.09 cache reads per million. Above 512K, it costs $0.90 input, $3.60 output, and $0.18 cache reads. Do not apply those M3 prices or service-tier assumptions to Flash. Standard and Priority.

H3 and media API rates

MiniMax H3 costs $0.08 per output second at 768P and $0.13 per output second at 2K. Regenerating an existing 768P result to 2K costs $0.05 per output second, with the original input materials billed again. The first 5 reference images are free; later images cost $0.04 each during generation and $0.025 each during regeneration. Reference audio is free. Reference video is metered per second at the selected output resolution's rate. Video rates.

Speech-2.8-turbo costs $60 per million characters, and speech-2.8-hd costs $100. Rapid voice cloning is $1.50 per voice and Voice Design is $3 per voice. For existing paying music API users, the listed Music-3.0 rate remains $0.15 for up to 5 minutes and lyrics generation or editing is $0.01 per song; new users cannot buy access through those APIs. Image-01 is $0.0035 per image, API-vlm is $0.01 per request, and server-side web search is $0.01 per request. Media and tool rates.

Every Audio Subscription tier

MiniMax still lists these five fixed speech subscriptions plus custom pricing. RPM means requests per minute; TPM means tokens per minute. Audio subscription table.

• Starter: $5 monthly, $13.50 for 3 months, or $48 yearly; 100,000 audio points/month, 10 RPM, and 10 voice slots.
• Standard: $30 monthly, $81 for 3 months, or $288 yearly; 300,000 points, 50 RPM, and 100 voice slots.
• Pro: $99 monthly, $267 for 3 months, or $950 yearly; 1.1M points, 200 RPM, and 250 voice slots.
• Scale: $249 monthly, $672 for 3 months, or $2,390 yearly; 3.3M points, 500 RPM, and 500 voice slots.
• Business: $999 monthly, $2,697 for 3 months, or $9,590 yearly; 20M points, 800 RPM, and 800 voice slots.
• Custom: negotiated capacity, unlimited RPM/TPM, priority updates, and negotiated service guarantees.

The subscription page names points rather than promising a universal character conversion. Compare the intended workload in the billing console before treating the point total as a fixed number of narration minutes.

Every legacy Video Package tier

MiniMax's one-month Video Packages are Standard at $1,000 for 3,760 video points and 20 RPM, Pro at $2,500 for 9,920 points and 30 RPM, Scale at $4,500 for 18,900 points and 40 RPM, and Business at $6,000 for 26,780 points and 50 RPM. Custom pricing offers negotiated points and unlimited RPM/TPM.

These packages support the Hailuo series and do not support H3 yet, according to the current Video Packages table. Unused package quantity resets to zero when the pack expires. A buyer drawn in by H3 should not assume that a large Hailuo pack discounts H3 usage.

The Limitations That Change the Decision

MiniMax has four limitations serious enough to change who should buy it and how it should be deployed.

H3's open weights are partial and territorially restricted

MiniMax H3 is open-weight only at the Base layer. The released Base checkpoints produce 768P. H3-Context-IR, which MiniMax describes as critical to final quality, remains a hosted service. H3-Regenerate-2K is also not open, and the initial release does not include the model's sparse-attention implementation. MiniMax's own SGLang commands specify 4 GPUs for each task checkpoint. H3 model card.

The license is an even larger constraint for the site's US-first audience. The standard H3 Community License defines the United States, European Union, United Kingdom, and South Korea as excluded territories. Organizations in those regions can apply for formal authorization, and the hosted API remains globally available. Commercial products and services generating more than $20M in annual revenue also require prior written authorization. H3 license, MiniMax licensing Q&A.

That means "downloadable" is not the same as "ready for a US production deployment." Use the hosted API for evaluation, and get the correct authorization before building a local product around the weights. The wider open-weight model guide explains why files, license, hardware, and a reproducible serving path are four separate gates.

M Plan is not a production capacity guarantee

MiniMax says M Plan is designed for individual, interactive developer use and recommends pay-as-you-go for production. Its 5-hour and weekly limits for non-video usage, weekly-only video limit, non-carrying quota, and dynamic peak throttling can suit a personal coding agent. They are not a capacity commitment for a customer-facing service. Plan limits.

Use M Plan to learn the model and support interactive work. Move a production route to pay-as-you-go with explicit timeouts, retries, observability, and spend caps.

Billing is fragmented across products

MiniMax's low rates sit behind several ledgers. M Plan quota, purchased credits, pay-as-you-go balance, audio points, and video points are not one interchangeable pool. Video Packages exclude H3, and the current M Plan page lists personal tiers rather than the old Team offering. A finance owner cannot estimate the monthly total from the Go, Explore, and Build cards alone.

Build the forecast by workload: Flash or M3 input/output, H3 output and reference duration, speech characters, image count, and any specialist subscription. Add music API spend only if you are an eligible existing paying user. Then assign each line to its actual balance. This is tedious, but it prevents a cheap headline rate from hiding stranded credits or a second bill.

The strongest performance evidence is vendor-reported

The release results are useful for choosing an evaluation, not for skipping one. Flash's launch announcement says weights will be released later, despite open-weight wording elsewhere on MiniMax's site. M3's own model page also mixes open-weight positioning with a promise of future full open sourcing. Treat release files, a license, and a usable serving path as requirements before planning private infrastructure. Flash release status, M3 access section.

The Kling AI review applies the same standard to video: the model earns a place through accepted shots and controllable workflows, not its best launch demo. MiniMax deserves the same discipline.

MiniMax AI Verdict

M3.1 Flash is the better buy for routine coding and agent work when the output is easy to verify and the total cost per accepted task stays below your Claude or GPT route. Start with small fixes, test generation, repetitive refactors, and bounded tool workflows. You do not need a mixed-media project to make that decision.

The spend example is concrete: 100 single requests at the illustrative 200K-input and 20K-output workload cost $4.20 on Flash before tools and retries. That can be a useful alternative to another monthly subscription. It is not a promise that 100 agent tasks cost $4.20: one task may send many requests, and thinking output counts toward usage. Token rates, thinking behavior.

Keep paying for Claude or GPT when your existing workflow solves ambiguous bugs, broad architectural changes, or consequential work with materially less correction. Also keep it when its included usage already covers the routine queue and you need its wider app features: adding Flash then creates another bill rather than removing one.

For a founder's next working week, choose a bounded recurring task, run it through Flash with the same checks as the incumbent, and compare accepted output plus review time. Keep customer-facing traffic on pay-as-you-go. Buy M Plan for interactive use only after its usage windows fit your working pattern; don't use the retired Plus illustration as a current break-even calculation.

H3, Speech, Music, and the Hub-to-Design creative workflow remain separate reasons to choose MiniMax. Start H3 at 768P, pay for 2K after a shot survives review, and include reference-video duration in the estimate. A US, EU, UK, or South Korean organization should use the H3 API or secure formal open-weight authorization before local deployment. Do not plan self-hosted Flash around weights that have only been promised.

MiniMax AI FAQ

How much does MiniMax AI cost?

M Plan regular renewals are $22/month for Go, $55 for Explore, and $132 for Build, with annual prices of $220, $550, and $1,320. Flash API pricing is $0.15 input, $0.60 output, and $0.015 cached input per million tokens. The first-month promotion and its deadline are in the pricing section. M Plan, API rates.

Is MiniMax AI better than ChatGPT?

Flash can be the better buy for routine coding and tool work if it passes the same checks at lower total cost. ChatGPT is the better purchase when its workspace and included coding usage already solve the job. This review verifies published information; it does not claim a head-to-head performance test.

Is MiniMax AI safe to use?

MiniMax H3 applies automated moderation to submitted text, images, and video, but its model card says false positives and false negatives remain possible. Keep sensitive material out of an evaluation until the relevant privacy, security, and contract terms meet your requirements, and enforce permissions outside the model.

Is MiniMax worth it?

MiniMax is worth it when a measured Flash, M3, H3, or speech workflow saves money after retries and human review while meeting the acceptance bar. It is not worth adding several MiniMax surfaces for occasional chat that a simpler subscription already handles.

Can MiniMax H3 run locally?

The 768P H3 Base checkpoints can run locally, and ComfyUI provides native workflows. The official MiniMax SGLang example uses 4 GPUs, while the quality-critical Context-IR and 2K regeneration stages remain hosted. The standard open-weight license also excludes the US, EU, UK, and South Korea unless formal authorization is granted.

Want the current model picks mapped to the work they fit? Get the AI tools map for business owners, free.

Get the AI tools map for business owners

A plain-English map of which AI models and tools fit which job, updated as prices and capabilities change. Free to subscribers.