Skip to main content
BenchLM
Data

MiniMax M3 vs Step 5 Preview

Updated November 5, 2026. Rank says Step 5 Preview is ahead. Price, access, and your workload can each overturn that. Public scores include evidence status and uncertainty.

Share or export
Share on XLinkedInAPI/MCP

Decision reading

Step 5 Preview has the higher public point estimate, 56.81 versus 49.3. Their conditional score ranges overlap. These ranges do not establish rank confidence. 6 results are shared. Category rows resting on Estimated evidence or different benchmark sets are marked directional and do not name a winner.

Model A
MiniMax logo
MiniMax M3

MiniMax

49.3/100

Supported · Public rank #100

90% interval 34.0–64.5

Model B
StepFun logo
Step 5 Preview

StepFun

56.81/100

Estimated · Public rank #60

Conditional range 47.1–66.5

Shared results
6
MiniMax M3 only
22
Step 5 Preview only
18
Like-for-like categories
2 / 8
Supported: MiniMax M3 · Estimated: Step 5 Preview. Conditional ranges do not establish rank confidence.How the comparison works

Which one for your work

Recommendations appear only when a shared evidence basis or an explicit operating constraint supports the call. Secondary and unsupported use cases stay disclosed below the initial list.

  • Chat turn cost

    1K fresh input + 500 output tokens

    MiniMax M3

    MiniMax M3 has the lower estimated token cost for this stated workload. Costs use the listed standard API rates.

    Confidence: listed-rates

  • Cache-heavy agent loop cost

    200K cached + 20K fresh input + 10K output tokens

    MiniMax M3

    MiniMax M3 has the lower estimated token cost for this stated workload. Costs use the listed standard API rates.

    Confidence: listed-rates
Show secondary and unsupported calls

  • Repository review cost

    50K fresh input + 3K output tokens

    MiniMax M3

    MiniMax M3 has the lower estimated token cost for this stated workload. Costs use the listed standard API rates.

    Confidence: listed-rates

  • Coding work

    Code generation, repair, and software-engineering tasks

    Not enough matched evidence

    Step 5 Preview is scored on Estimated evidence for coding, so the reading is directional rather than like-for-like.

    Confidence: limited

  • Agentic work

    Tool use, computer use, and multi-step task completion

    Not enough matched evidence

    Step 5 Preview is scored on Estimated evidence for agentic, so the reading is directional rather than like-for-like.

    Confidence: limited

  • Long documents

    Prompts that approach the documented context limit

    No clear pick

    The documented context windows are equal.

    Confidence: documented

Which one for a specific job

Choose a job from the LLM Selector's task catalog to see the category row it rests on, under the same basis rules as the table below. A directional row stays directional; choosing a job never creates a winner.

The same task catalog as the LLM Selector. Each job names the evidence surface it rests on; nothing here adds a new score.

37.4MiniMax M348.8Step 5 Preview

Directional only · BenchAlign v5.8

Step 5 Preview has the higher coding point estimate. Conditional score ranges do not establish rank confidence.

Coding scores combine specific tasks and setups. Match the editor, harness, and effort to your workflow.

Open the coding leaderboardShortlist models for this job
Same basis rules as the category table below

What is actually comparable

Shared results can support a head-to-head reading. Results present for only one model describe coverage, not superiority.

2 categories rest on Estimated evidence or different evidence sets. Those rows remain visible for coverage context but do not name a winner.

Shape of the matched evidence

Only shared public evidence is shown. Sparse evidence stays a ruled list rather than being closed into a radar shape.

Too few matched category axes support a radar. The ruled list below shows only shared benchmark results; positions use each benchmark’s normalized display scale when available.

Bars run 0–100 on each benchmark’s normalized display scale

Category results, on a stated basis

Each row shows the public-lane category score for both models: the BenchAlign v5.8 lane for agentic, coding, and knowledge, and the weighted public rows elsewhere, with the same rank each model page reports. A row is like-for-like only when both scores rest on Supported evidence or the same weighted set. Directional and not-comparable rows remain visible, but they never receive a winner.

Multimodal

Like-for-like
MiniMax M3
59.0
#48/58
Step 5 Preview
73.3
#32/58
Basis
Provisional lane · 2 vs 2 weighted rows
Reading
Step 5 Preview leads

Knowledge

Like-for-like
MiniMax M3
54.9
Supported · #60/177
Step 5 Preview
61.0
Supported · #44/177
Basis
BenchAlign v5.8 lane · 2 vs 2 public rows
Reading
Step 5 Preview leads · intervals overlap

Agentic

Directional only
MiniMax M3
39.3
Supported · #78/135
Step 5 Preview
55.3
Estimated · #42/135
Basis
BenchAlign v5.8 lane · 10 vs 13 public rows
Reading
Directional only

Coding

Directional only
MiniMax M3
37.4
Supported · #75/156
Step 5 Preview
48.8
Estimated · #47/156
Basis
BenchAlign v5.8 lane · 10 vs 6 public rows
Reading
Directional only

Reasoning

Not comparable
MiniMax M3
82.0
Unranked · 2 rankable rows
Step 5 Preview
82.5
Unranked · 2 rankable rows
Basis
Provisional lane · 0 vs 0 weighted rows
Reading
Not comparable

Multilingual

Not comparable
MiniMax M3
Not ranked
Step 5 Preview
Not ranked
Basis
Provisional lane · 0 vs 0 weighted rows
Reading
Not comparable

Instruction following

Not comparable
MiniMax M3
93.8
#6/157
Step 5 Preview
Not ranked
Basis
Provisional lane · 0 vs 0 weighted rows
Reading
Not comparable

Math

Not comparable
MiniMax M3
Not ranked
Step 5 Preview
Not ranked
Basis
Provisional lane · 1 vs 0 weighted rows
Reading
Not comparable

Ranks count the models scored in each category’s lane, so the agentic, coding, and knowledge denominators (BenchAlign v5.8) differ from the provisional-lane categories. Unranked scores sit on the provisional lane’s scale but never name a winner. Like-for-like rows within 0.5 points read as a practical tie.

Supported evidence per lane · bars run 0–100Methodology

What each workload costs

Three fixed token mixes turn per-token rates into comparable decisions. Each scenario states context fit and whether cached input had to fall back to the published list-input rate.

Chat turn

1K fresh input + 500 output tokens

MiniMax M3
$0.0009
Fits in one request
Step 5 Preview
$0.00235
Fits in one request

MiniMax M3 has the lower modeled cost

Costs use the listed standard API rates.

Repository review

50K fresh input + 3K output tokens

MiniMax M3
$0.0186
Fits in one request
Step 5 Preview
$0.0581
Fits in one request

MiniMax M3 has the lower modeled cost

Costs use the listed standard API rates.

Cache-heavy agent loop

200K cached + 20K fresh input + 10K output tokens

MiniMax M3
$0.03
Fits in one request
Step 5 Preview
$0.057
Fits in one request

MiniMax M3 has the lower modeled cost

Costs use the listed standard API rates.

Cached input falls back to the list input rate only where a cached rate is unpublished

Specification differences

Sourced differences are shown directly. Missing facts stay explicit instead of being inferred from a model name or family.

Cached-input rate

A missing cached-input rate falls back to the listed input rate only in the stated workload estimate.

MiniMax M3

$0.06 per 1M cached input tokens

Step 5 Preview

$0.05 per 1M cached input tokens

StepFun pricing and rate limits

Reasoning profile

MiniMax M3

Non-Reasoning

Step 5 Preview

Reasoning

Weight access

MiniMax M3

Open Weight

Step 5 Preview

Open Weight

License

MiniMax M3

Open Weight

Step 5 Preview

Open Weight

Release date

MiniMax M3

2026-06-01

Step 5 Preview

2026-09-20

If you already use one of these models

Deployment change
The models list different providers, so authentication, endpoint behavior, limits, and feature support may change.
Quality signal
Step 5 Preview has the higher public point estimate, 56.81 versus 49.3. Their conditional score ranges overlap. These ranges do not establish rank confidence.
Workload cost
Repository review: $0.0186 vs $0.0581. Cache-heavy agent loop: $0.03 vs $0.057.
Context tradeoff
Both models list 1M.
Run the same representative tasks against both endpoints before changing production traffic.

Questions

Which is better, MiniMax M3 or Step 5 Preview?

Step 5 Preview has the higher public point estimate, 56.81 versus 49.3. Their conditional score ranges overlap. These ranges do not establish rank confidence. The page therefore keeps the decision tied to the specific documented workload.

Which is better for coding, MiniMax M3 or Step 5 Preview?

Step 5 Preview scores higher for coding on the public lane, 48.8 to 37.4. Step 5 Preview is scored on Estimated evidence for coding, so the reading is directional rather than like-for-like. BenchLM does not name a winner for coding from a directional reading; read the shared benchmark rows directly and test the models on the same task set.

Which is better for agentic tasks, MiniMax M3 or Step 5 Preview?

Step 5 Preview scores higher for agentic tasks on the public lane, 55.3 to 39.3. Step 5 Preview is scored on Estimated evidence for agentic tasks, so the reading is directional rather than like-for-like. BenchLM does not name a winner for agentic tasks from a directional reading; read the shared benchmark rows directly and test the models on the same task set.

Which costs less, MiniMax M3 or Step 5 Preview?

For the stated presets, chat costs $0.0009 on MiniMax M3 and $0.00235 on Step 5 Preview; repository review costs $0.0186 and $0.0581; the cache-heavy agent loop costs $0.03 and $0.057. Costs use the listed standard API rates.

Which has the larger context window, MiniMax M3 or Step 5 Preview?

Both models list the same context window, 1M.

Self-host vs API cost

Estimates at 50,000 req/day · 1000 tokens/req average.

MiniMax M3
API / mo$1,125
Self-host / mo$13,981
Break-even818M/day
Step 5 Preview
API / mo$2,775
Self-host / mo$27,962
Break-even617M/day
Model the full break-even

Benchmark evidence

The full public result ledger is available for audit without forcing a wide desktop table onto a phone.

Browse raw public benchmark evidence46 rows

Agentic

  • Terminal-Bench 2.1

    MiniMax M366.0%
    Source
    Step 5 Preview85.0%
    Source

    Step 5 Preview leads this result

  • BrowseComp

    MiniMax M383.5%
    Source
    Step 5 Preview88.7%
    Source

    Step 5 Preview leads this result

  • OSWorld-Verified

    MiniMax M370.1%
    Source
    Step 5 Preview—

    Not directly comparable

  • MCP Atlas

    MiniMax M374.2%
    Source
    Step 5 Preview85.6%
    Source

    Step 5 Preview leads this result

  • Claw-Eval

    MiniMax M374.5%
    Source
    Step 5 Preview—

    Not directly comparable

  • BankerToolBench

    MiniMax M376.1%
    Source
    Step 5 Preview—

    Not directly comparable

  • ResearchClawBench

    MiniMax M319.8%
    Source
    Step 5 Preview—

    Not directly comparable

  • OSWorld 2.0

    MiniMax M34.6%
    Source
    Step 5 Preview—

    Not directly comparable

  • Terminal-Bench 2.1 (Vals)

    MiniMax M353.6%
    Source
    Step 5 Preview—

    Not directly comparable

  • MazeBench

    MiniMax M30%
    Source
    Step 5 Preview—

    Not directly comparable

  • Terminal-Bench 4.0

    MiniMax M3—
    Step 5 Preview33.30%
    Source

    Not directly comparable

  • CyberGym

    MiniMax M3—
    Step 5 Preview84.7%
    Source

    Not directly comparable

  • AutomationBench

    MiniMax M3—
    Step 5 Preview44.0%
    Source

    Not directly comparable

  • Toolathlon-Verified

    MiniMax M3—
    Step 5 Preview74.1%
    Source

    Not directly comparable

  • JobBench

    MiniMax M3—
    Step 5 Preview59.0%
    Source

    Not directly comparable

  • APEX-Agents

    MiniMax M3—
    Step 5 Preview37.8%
    Source

    Not directly comparable

  • DRACO

    MiniMax M3—
    Step 5 Preview83.3%
    Source

    Not directly comparable

  • Agents' Last Exam

    MiniMax M3—
    Step 5 Preview29.5%
    Source

    Not directly comparable

  • Hermes Index

    MiniMax M3—
    Step 5 Preview32.8%
    Source

    Not directly comparable

  • Hermes Bench

    MiniMax M3—
    Step 5 Preview69.2%
    Source

    Not directly comparable

Coding

  • SWE-bench Verified

    MiniMax M380.5%
    Source
    Step 5 Preview—

    Not directly comparable

  • SWE-bench Pro

    MiniMax M359%
    Source
    Step 5 Preview—

    Not directly comparable

  • Terminal-Bench 2.1

    MiniMax M366.0%
    Source
    Step 5 Preview85.0%
    Source

    Step 5 Preview leads this result

  • NL2Repo

    MiniMax M342.1%
    Source
    Step 5 Preview—

    Not directly comparable

  • VIBE V2

    MiniMax M350.1%
    Source
    Step 5 Preview—

    Not directly comparable

  • SVG-Bench

    MiniMax M363.7%
    Source
    Step 5 Preview—

    Not directly comparable

  • KernelBench Hard

    MiniMax M328.8%
    Source
    Step 5 Preview—

    Not directly comparable

  • OpenHarmony Bench

    MiniMax M348.4%
    Source
    Step 5 Preview—

    Not directly comparable

  • LiveCodeBench (Vals)

    MiniMax M382.2%
    Source
    Step 5 Preview—

    Not directly comparable

  • SWE-bench (Vals)

    MiniMax M375.0%
    Source
    Step 5 Preview—

    Not directly comparable

  • DeepSWE

    MiniMax M3—
    Step 5 Preview67.7%
    Source

    Not directly comparable

  • SciCode

    MiniMax M3—
    Step 5 Preview58.9%
    Source

    Not directly comparable

  • ProgramBench

    MiniMax M3—
    Step 5 Preview80.5%
    Source

    Not directly comparable

  • sweMarathon

    MiniMax M3—
    Step 5 Preview72.7%
    Source

    Not directly comparable

  • MLS-Bench Lite

    MiniMax M3—
    Step 5 Preview40.5%
    Source

    Not directly comparable

Reasoning

  • CritPt

    MiniMax M3—
    Step 5 Preview20.9%
    Source

    Not directly comparable

Multimodal

  • OfficeQA Pro

    MiniMax M345.1%
    Source
    Step 5 Preview60.3%
    Source

    Step 5 Preview leads this result

  • OmniDocBench 1.5

    MiniMax M391.6%
    Source
    Step 5 Preview—

    Not directly comparable

  • MMMU-Pro

    MiniMax M378.1%
    Source
    Step 5 Preview76%
    Source

    MiniMax M3 leads this result

  • VideoMMMU

    MiniMax M384.6%
    Source
    Step 5 Preview—

    Not directly comparable

  • Video-MME (with subtitle)

    MiniMax M385.4%
    Source
    Step 5 Preview—

    Not directly comparable

Knowledge

  • GPQA Diamond (Vals)

    MiniMax M392.7%
    Source
    Step 5 Preview—

    Not directly comparable

  • MMLU-Pro (Vals)

    MiniMax M384.2%
    Source
    Step 5 Preview—

    Not directly comparable

  • GPQA-D

    MiniMax M3—
    Step 5 Preview93.5%
    Source

    Not directly comparable

  • HLE

    MiniMax M3—
    Step 5 Preview46.5%
    Source

    Not directly comparable

Math

  • USAMO 2026

    MiniMax M385.7%
    Source
    Step 5 Preview—

    Not directly comparable

46 public results · 6 shared

Related comparisons

Compare API pricingRead the methodologyOpen the model selector

Watch MiniMax M3 vs Step 5 Preview

One weekly email when material rank, price, or benchmark evidence changes make this matchup worth revisiting.

Read a sample issue

Join 6,100+ subscribers.

Last updated November 5, 2026