MiniMax M3.1 Flash Is 50% Off on B.AI Until 2 November: What the Deal Really Cuts

On 26 October 2026, B.AI announced a limited time benefit: its latest MiniMax model, MiniMax M3.1 Flash, is available at 50% off across the B.AI web platform, Android app, and API. The event runs from 26 October at 17:00 to 2 November at 16:59 Singapore time. That is the deal as reported, and it is the reason this model is worth a look this week. But a percentage off means little until you know which price it applies to, and what a finished agent task actually costs.

MiniMax M3.1 Flash is the model that ran anonymously as Space Bunny before its official launch on 19 October 2026. MiniMax says that early version processed more than 200 trillion tokens while developers speculated about its identity. The model is positioned as a lightweight, agent focused model for coding, computer use, and cybersecurity work, with a 1 million token context window. It is not a general flagship meant to beat every model on every test. It is a fast, cheap model meant to run many agent steps without a large bill. That positioning is what makes a 50% deal interesting, and also what makes the fine print worth reading.

What is actually 50% off?

B.AI said the 50% discount applies to MiniMax M3.1 Flash on B.AI itself: web, Android app, and API. That is a platform offer, not a MiniMax direct price cut. MiniMax's own official API rates, as listed in its pricing documentation and repeated across independent pricing checks, are $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens. Those official rates are already about half the previous MiniMax M3 rates for short context requests, which were $0.30 input and $1.20 output per million tokens.

Do not assume the B.AI discount simply halves the official MiniMax rates. B.AI has its own pricing, credits, and billing, and the reported announcement does not publish a B.AI token price in the coverage we checked. The safe way to use this deal is to open B.AI, find the M3.1 Flash price shown at checkout for your exact use, and compare it with the official rates in the table below. If the B.AI price is clearly lower for the same work, the deal is real for you. If it is not, the headline percentage is not saving you anything.

That example task cost is arithmetic on the official rates, not a B.AI price. One million input tokens at $0.15 plus 100,000 output tokens at $0.60 per million equals $0.21. It shows why this model class is cheap for bulk work, but it does not include any B.AI fee, credit markup, or discount. Check those on B.AI itself.

Why is this model cheap, and where does the cost hide?

MiniMax attributes the low price to efficiency, not to a smaller model alone. The company claims output above 200 tokens per second, with a median above 170 tokens per second, and inference about three times faster than the previous model, thanks to a more aggressive sparse attention design. Independent measurement is more cautious: Artificial Analysis measured about 105 tokens per second at maximum effort, well below the claimed peak. Speed claims from a vendor are best treated as a ceiling, not as the speed you will get on your own task at your own settings.

The bigger cost trap is thinking. MiniMax says M3.1 Flash always reasons before answering, that thinking depth can be lowered but not disabled, and that thinking tokens count toward the output limit. That means the effort setting is a cost setting. A low effort run can be dramatically cheaper than a maximum effort run on the same task. One independent test by AIHubMix on a single billing function fix found the difference clearly: at low effort, the fix used 889 output tokens and cost about $0.0006, while at maximum effort it used 6,673 output tokens and cost about $0.0045, with both passing the same hidden tests. That is one test, not a guarantee, but it shows the pattern: start low, and only raise effort when the low effort answer fails.

MiniMax also claims overall usage cost for agentic tasks is 70% below its predecessor. Treat that as a company level workflow claim, not as a price you can bank on. The tariff cut from M3 to M3.1 Flash is 50% on input and output for short context, and 75% on cached reads. The extra saving to reach 70% on a full workflow depends on fewer wasted tokens, lower effort settings, and caching working in your favour. Your workflow may or may not get there.

What should you test during the deal window?

First, test the exact job you want to buy this model for. Pick one coding task, one agent workflow, or one computer use task, and run it at low effort first. Note the tokens, the time, and whether the answer is usable. Then run the same task at a higher effort only if the low effort version fails. That two step test tells you your real cost per finished job, which is the only number that matters when a discount window is short.

Second, check the limits that matter for agents: context window, tool use, and whether the B.AI route supports the features your workflow needs, such as structured output or multi step tool calls. A cheap model that cannot call your tools is not a saving. Third, check the end date in your own time zone. The window ends 2 November at 16:59 Singapore time. Singapore is 3 hours ahead of Pakistan time, so that is 13:59 Pakistan time on 2 November, and early morning in the United States on 2 November after the clocks change. Do not leave a bulk purchase to the last hour, when a checkout problem has no time to be fixed.

Fourth, watch the quality caveats that independent coverage has flagged. VentureBeat, citing Artificial Analysis, reported that M3.1 Flash showed a high tendency to give incorrect answers rather than admit uncertainty. That is a warning to verify outputs, especially for code, cybersecurity, and any task where a confident wrong answer costs more than a slow right one. Use the model for drafts, first passes, and bulk work you can check, and keep a stronger model for final decisions until you have measured its error rate on your own tasks.

Who should use this deal, and who should skip it?

This deal fits developers and small teams running many agent steps, coding assistance, or bulk analysis where speed and low unit cost matter more than top tier reasoning. It fits prototyping, where a 50% window lets you test a workflow cheaply before committing to a larger spend. It fits anyone already using B.AI, because the discount is on that platform, not a reason alone to switch platforms.

Skip it if you need the model to admit uncertainty reliably, if your workflow needs a feature B.AI does not expose for this model, or if the B.AI checkout price for your use is not actually lower than the official MiniMax rate. Also skip a bulk credit purchase if you have not yet run the low effort test above. Credits bought in a discount window are only a saving if you use them on work that passes your quality check.

Key takeaways:

• B.AI announced MiniMax M3.1 Flash at 50% off from 26 October 17:00 to 2 November 16:59 Singapore time, on web, Android app, and API.
• MiniMax's official rates are $0.15 input, $0.60 output, and $0.015 cached input per million tokens.
• The B.AI discount is a platform offer, so verify the B.AI checkout price against the official rates before buying.
• Thinking tokens count, and effort level changes cost sharply, so start at low effort and raise it only when needed.
• Verify outputs, because independent coverage flagged a tendency to answer confidently rather than admit uncertainty.

Sources: B.AI announcement of the MiniMax M3.1 Flash 50% limited time benefit, 26 October 2026, as reported by ChainCatcher; MiniMax M3.1 Flash pricing documentation, $0.15 input, $0.60 output, $0.015 cached input per million tokens; MiniMax launch coverage on Space Bunny, 1 million token context, and speed claims, October 2026; AIHubMix independent effort test, 19 October 2026.

Related reading

• TopUseAI Blog
• Open Source LLM API Pricing
• Best Local Coding Model VRAM