JOURNAL / 2026.10.20

MiniMax M3.1 Flash and Mercury 3: two ways to lower an agent's cost

MiniMax and Inception launch models for fast, affordable work. Comparing their prices also requires looking at default reasoning settings, waits between tools, and the evidence behind capability claims.

October 19 brought MiniMax M3.1 Flash and Mercury 3 from Inception. Both are available through APIs and aim to bring agent capabilities to applications with tight budgets. MiniMax promises improvements in coding, automation, and computer use; Inception introduces a diffusion model generating more than a thousand tokens per second. They share a practical opportunity: consulting the model several times during an assignment could become more viable. The available evidence does not yet establish a winner.

There are concrete changes from their predecessors. MiniMax's pricing halves the input and output rates of standard M3 for inputs up to 512,000 tokens, and Flash keeps a flat rate. Mercury expands context from 260,000 to one million tokens compared with Mercury 2.5, whose announcement already reported more than a thousand tokens per second. Its announced advance combines capability and context while preserving speed; fast generation is not a new discovery with this release.

  • MiniMax M3.1 Flash: $0.15 per million input tokens and $0.60 for output; cached input has a separate rate.
  • Mercury 3: $0.10 for input and $0.35 for output per million tokens during the launch promotion; regular rates are $0.20 and $0.70, respectively.

Figures come from MiniMax's official prices and Inception's model catalog, checked today. They exclude external tools and do not imply that both models consume the same tokens. That catalog describes Mercury 3 as a text model with tools and structured outputs. MiniMax's guide documents text, image, and video inputs: for an agent interpreting screenshots, modality may decide the choice before price does.

The setting that changes a comparison

MiniMax always reasons and uses max effort unless configured otherwise. Effort can be reduced, but reasoning cannot be disabled. Mercury 3 supports settings from instant to high; its effort documentation sets low as the default, whereas Mercury 2.5 used medium.

Changing only the model name can also change how much it deliberates. This matters both when comparing providers and when migrating within a family. I would explicitly set effort and the response budget, then find the fastest configuration that meets the required accuracy. Matching the setting's name does not guarantee matching internal work either.

Diffusion adds another difference. The original Mercury report from 2025 explains refining multiple tokens in parallel; it helps explain the family, rather than providing a technical evaluation of Mercury 3. The current streaming guide distinguishes refined text blocks from an optional mode showing complete intermediate steps. An application displaying those drafts must replace them rather than concatenate them as final fragments. It should also avoid turning a draft into an executable instruction.

Where savings could matter

I find this especially promising for short cycles: look up a record, choose the next tool, and check its response. Saving time at each decision could improve the experience without producing longer answers. But a slow browser, a remote query, or a retry can dominate total time. Tokens per second describe only part of the journey.

Capability claims need the same caution. MiniMax reports improvements on several agent tests; Inception claims to compete with Luna and Haiku and includes a partner's test on 50 web tasks. These are pieces of evidence from providers and selected collaborators with different scopes; they are not a controlled comparison between these two releases. I would not transfer their results to a new workflow without checking.

My decisive test would run the same assignments with identical tools and acceptance criteria, counting failures, corrections, the complete bill, and time to an acceptable result. I would separate the first visible response from actual completion. The valuable advantage would be being able to verify more within the same budget, provided those checks detect errors.

There is also an availability boundary: although MiniMax's guide uses the term “open-weight,” its announcement says weights will follow later. Today I would treat it as service access; a promise of weights does not yet allow evaluating a self-hosted deployment.

Primary sources accessed October 20, 2026. We have not run these models or reproduced their evaluations.

← Back to journal