MiniMax Takes Its Time

If You Only Read One Thing

Three extra correct answers can change a model’s job. Plotly’s MiniMax test finds a capable analyst with an inconvenient waiting time; Mike Taylor’s personal benchmark shows how to turn actual work into a model comparison. The selection problem is subtle: the best model for memorable failures may differ from the best model for the work that fills the day.

MiniMax Earns the Background Job

MiniMax’s new model belongs on the shortlist for analysis that can run in the background. Its strongest independent result comes with enough waiting to make an interactive deployment a different proposition.

M3.1 Flash launched October 19 through MiniMax’s API, subscription plan and coding application. List prices are $0.15 per million input tokens and $0.60 per million output tokens. The company identifies the earlier anonymous Space Bunny as a preliminary version; downloadable weights remain a promise. MiniMax’s claimed 70% reduction in agent-task costs is a vendor comparison with its predecessor, not a universal workload result.

Plotly’s October 20 test gives that distinction substance. M3.1 Flash passed 41 of 43 data-analysis questions, up from Space Bunny’s 38. It discovered all five deliberately hidden patterns, versus two previously. But median completion time increased from 36 seconds to three minutes, and the run’s estimated list-price cost rose from $0.82 to $4.09. Those are published estimates, not the bill from the temporarily free endpoint.

Think of an analyst who checks more explanations before submitting a report. More investigation can reveal a missing observation or an unexpected pattern, but the extra work still takes time. Plotly recorded 45 model calls per question against 15 for the preview. The evidence supports a changed investigation pattern; it does not isolate which training change caused the better answers.

The practical payoff is a more specific deployment choice. A scheduled report can tolerate several minutes if the additional investigation prevents a misleading conclusion. The same behavior inside a conversational dashboard can make every follow-up feel slow. Token price alone cannot decide between those applications.

There is also a strong competitive objection: MiniMax is not the cheapest model at this level on Plotly’s exam. GPT-6 Sol matched its score for $2.61; GPT-6.1 Sol answered all 43 correctly at that cost. Separately, Artificial Analysis places M3.1 Flash at 43 on its broader intelligence index. Neither comparison establishes universal frontier parity.

This extends yesterday’s discussion of efficiency budgets with a concrete model-selection boundary. The likely opportunity is unattended investigation where answer quality matters more than immediate response. The deciding test is accepted analytical output within the application’s waiting-time limit, measured on the same questions and tools.

Past Failures Need a Second Sample

A collection of failed AI tasks can reveal a better model while giving a distorted picture of the savings from switching. The missing ingredient is how often those difficult tasks occur in ordinary work.

Mike Taylor’s personal-benchmark account, updated October 16, describes a reproducible way to preserve useful comparisons. He collects ten real assignments with their original context, compares model outputs, and translates preferences into explicit checks. His examples span dashboards, writing and presentations. The accompanying public artifact makes this more substantial than advice to “try models on your own tasks.”

The published run data also shows why an overall impression is insufficient. In the August 22 runs, GPT-5.6 Luna passed nine of nine applicable dashboard checks, while Claude Fable 5 passed eight. On the writing task, Luna passed five of ten and Fable eight. These are historical, author-run results under Taylor’s criteria, not a ranking of today’s models or independently established quality.

That reversal is the useful evidence. One model can satisfy an analytical display’s requirements and still miss the standards for prose. The comparison belongs to a task and its acceptance criteria. A family name, a general intelligence score or an attractive average cannot preserve that distinction.

This extends the reader’s existing second-model review and compound-learning system. Reviews decide whether the current output survives. A reusable task collection tests whether a different worker can produce acceptable output from the same starting materials. Preserving the inputs and judgments makes past work useful for the next model decision, rather than merely adding another instruction to the agent.

The boundary is selection bias: a sample built from memorable failures overrepresents difficult work. Like testing a delivery service only on parcels that previously went missing, it answers a valuable question about recovery while saying little about typical performance. My inference is that model selection needs both the failure collection and a separate sample of routine assignments, with their actual frequency retained.

The cost is maintaining those cases and checking the judges against human decisions. Taylor’s artifact supports the method’s inspectability; it does not establish a causal productivity gain. The adoption test is whether a substitution chosen on archived cases also preserves acceptance on the next unseen batch of ordinary work. Success only on the archive would be learning the test.

The Contrarian Take

Everyone says: A larger intelligence score is the best starting point for choosing the model that will be easiest to work with.

Here’s why that’s wrong (or at least incomplete): Correctness and interaction cost are separate properties. Artificial Analysis’s new AA-PersonalityBench compares default communication across 900 questions without a system prompt. That makes verbosity, directness and other behaviors visible alongside capability. But a default style is not an immutable property: the real application adds instructions, tools and history. My inference is that a deployment comparison should retain the actual operating prompt and measure how much correction the output requires. A pleasant unprompted answer can still be an expensive starting point for a tightly specified deliverable.

Under the Radar

A displayed score can hide a changing denominator. Taylor’s public benchmark data records three of five applicable presentation checks for the historical Luna run and twelve of fifteen for Fable. Those fractions should not be treated as identical tests merely because both become percentages. The visible artifact makes the discrepancy inspectable; it does not establish why the applicable counts differ. Comparable criteria matter before an average becomes a routing rule.
Model cost can exclude the machinery that judged it. The same benchmark artifact explicitly labels a recorded token-cost estimate as excluding judge and subagent calls, tools and some premiums. That is useful accounting disclosure. A cheaper worker might require more assessment, so the missing expense could change a close comparison. The artifact supports checking cost scope, not claiming that its listed model expense equals the total workflow bill.

Quick Takes

Mercury 3 adds a different speed option. Inception’s October 19 release claims 1,068 output tokens per second and expands context to one million tokens. Its diffusion approach generates text through iterative refinement rather than strictly one token at a time. These are vendor results; faster generation does not establish faster successful agent work. The technical test differs from today’s News distribution angle: completion quality and end-to-end waiting time under the same workload. (Source)
OpenCode’s incident reaches the credential boundary. OpenCode’s October 17 statement says a breach affected 1.4% of Console users, specifically a subset using external providers. It describes an attempt to train another model on captured outputs, known as distillation. That motive remains a vendor allegation. The operational implication is that an agent interface can concentrate access to several model accounts. Changing the model does not remove exposure through the service holding its credentials. (Source)
Microsoft separates discovering agents from constraining them. Its October 19 announcement expands local-agent discovery to 60 types and introduces Guardian Agents within Project Perception. The proposed controls include preserving reads while restricting writes and requiring approval for consequential actions. Discovery establishes that an agent exists; it does not prove every action is intercepted. The missing deployment evidence is which execution paths the controls actually cover. (Source)

The Thread

A successful routing policy changes the evidence used to evaluate the next model. Once routine tasks move to an inexpensive worker, the premium model receives a harder remainder. Its observed success rate can fall even if its capabilities have not changed. Conversely, a cheap model can look excellent because the dispatcher protects it from difficult assignments. This is a selection effect created by deployment itself, distinct from yesterday’s question of where efficiency savings go. MiniMax’s background-analysis opportunity and Taylor’s task-specific results make that effect practical: comparisons need to retain which work each model was offered, including rejected or escalated assignments. My inference is that an enduring model-selection advantage comes from preserving the incoming workload distribution alongside outcomes. Otherwise every successful substitution makes the next before-and-after comparison less trustworthy. A stable sample of incoming tasks, evaluated before routing, would distinguish a changing model from a changing job.

Predictions

New predictions:

I predict: By October 30, MiniMax will publish downloadable M3.1 Flash weights in an official model repository. The launch explicitly promises weights, which is stronger evidence than extrapolating from an unrelated release; it still does not establish a delivery date. (Confidence: medium; Check by: 2026-10-30)

2026-10-20 · 03:29 AM ET