MiniMax

Tier A · Strong with clear operating rules.

MiniMax M3.1 Flash

The cheapest model in this roster, and the one whose newer benchmark set produced two unrunnable scenes.

API pricing

USD per million tokens · MiniMax API Standard
Input
$0.15
Output
$0.6
Cached input
$0.015
Context window
1M
Pricing details and specifications

Try the active visual tests

Open the generated scene to test it, or compare that prompt with another model.

7 runs

Concourse Divide

Run date: 2026-10-27

Generation provenance: unrecorded

Superbench 2.1 brief · tool-free single-message authoring, no repair turn

Visible agent rerouting, collision avoidance, and recovery from blocked connectivity.

Ember Glider

Run date: 2026-10-26

Generation provenance: unrecorded

Superbench 2.1 returning brief · original 1.x run protocol

Sunset gliding journey testing flight energy management, checkpoint flow, and atmospheric scene craft.

Lightkeeper's Atelier

Run date: 2026-10-27

Generation provenance: unrecorded

Superbench 2.1 brief · tool-free single-message authoring, no repair turn

Material differentiation, dynamic light, geometric shadows, and art direction.

Low-Poly Tower Defense

Run date: 2026-10-26

Generation provenance: unrecorded

Superbench 2.1 returning brief · original 1.x run protocol

Diorama tower defense testing economy balance, wave design, placement rules, and combat readability.

Neon Drift

Run date: 2026-10-26

Generation provenance: unrecorded

Synthwave time-trial racing testing drift physics, lap timing, ghost replay, and unlock progression.

Wrecklight Salvage

Run date: 2026-10-27

Generation provenance: unrecorded

Superbench 2.1 brief · tool-free single-message authoring, no repair turn

Three-phase game progression with persistent physical cargo and a visible late-stage consequence.

Editorial verdict

Where this model fits

Back to all benchmarks →

MiniMax M3.1 Flash is the Flash generation of MiniMax M3, released on October 19, 2026. It keeps M3 one-million-token context and native multimodality, adds tunable thinking depth, and drops to a flat $0.15 input and $0.60 output per million tokens — the cheapest entry in this ranking. Our own evidence is now seven visual builds: the three October 26 Superbench 1.x core briefs — Ember Glider, Low-Poly Tower Defense and Neon Drift — plus the four new Superbench 2.1 contracts from October 27 — Skyhook Cargo, Concourse Divide, Lightkeeper’s Atelier and Wrecklight Salvage. The 1.x set is clean: three working artifacts, zero console errors, one defect found and fixed before publication. The 2.1 set is the weaker half. Wrecklight Salvage is a working three-phase salvage game whose honest victory rule, gate cycle and review fixture all behaved as specified. Concourse Divide boots and simulates correctly but fails the required South-only recovery. Skyhook Cargo never advances its physics at all, and Lightkeeper’s Atelier throws on boot and never draws its scene. Both failures are published as evidence rather than repaired.

Best for

  • Cheap high-volume agent and build work
  • Long-context coding and whole-repository work at a 1M window
  • Multimodal inputs — text, image and video
  • Routing executor jobs where call count, not peak quality, drives cost

Watch out

  • Two of the four new Superbench 2.1 contracts do not run: Skyhook Cargo freezes its simulation and Lightkeeper’s Atelier throws before it renders
  • Concourse Divide fails its required recovery, so the crowd does not clear the station inside 90 simulation seconds
  • The 30 fps target on the two 3D briefs was never measured on real GPU hardware
  • Neon Drift shipped a boot-time crash that only the browser pass caught
  • Low-Poly Tower Defense balance was tuned against a near-optimal scripted player
  • Provider benchmark numbers are launch-reported and were not reproduced here

Why it is ranked here

  1. MiniMax launched M3.1 Flash on October 19, 2026 at a flat $0.15 input, $0.60 output and $0.015 cached input per million tokens — roughly half of M3 on every line, and the cheapest model in this roster.
  2. MiniMax reports 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents’ Last Exam, 76.2 on BankerToolBench and 82% on CyberGym, with output above 200 tokens per second. Those are vendor numbers and this ranking does not reproduce them.
  3. On our own prompts the model delivered three complete, self-contained browser builds from isolated sessions, each scoped to one brief and one output path with no reference implementation.
  4. After the verification loop all three artifacts load from file:// with zero console errors and zero page exceptions, and the flight, economy and lap mechanics respond to the exposed hooks.
  5. The October 27 Superbench 2.1 set is the weaker half of the evidence. One of four briefs — Wrecklight Salvage — produced a game whose gate cycle, towing, tether release and dock settle rule all behaved as specified. Concourse Divide produced a legible 48-traveler station with honest counts, but the crowd never cleared the concourse inside the required 90 simulation seconds after a South-only reopening, and a passage stayed stuck in its amber closing state with travelers trapped inside it.
  6. The other two 2.1 briefs are failed runs rather than weak ones. Skyhook Cargo freezes its own simulation, because the fixed substep constant collides with the canvas-height variable, so no input moves the trolley or the cable. Lightkeeper’s Atelier aborts on boot with a strict-mode reference error and never draws its scene. Both are published as failed-run evidence, exactly as delivered.
  7. A tier reflects that delivery record plus the cost position, and it is an editorial routing choice rather than a numeric benchmark score. The placement stays below S and the record now points in one direction: on the 1.x briefs this model ships working games, and on the newer 2.1 contracts it needs a browser check before anyone relies on it.

Evidence

2026-11-10

Superbash editorial model ranking

Takeaway: MiniMax M3.1 Flash is currently placed in Tier A.

The November 2026 editorial roster places MiniMax M3.1 Flash at rank 10.

Open source →
2026-10-19

MiniMax M3.1 Flash: Redefining Agent Performance and Cost

Takeaway: M3.1 Flash halves the price of M3 and holds frontier-adjacent agent and coding scores.

The launch post states the flat $0.15 input, $0.60 output and $0.015 cached price per million tokens, the DeepSWE, OSWorld V2, Agents’ Last Exam, BankerToolBench and CyberGym figures, and the 200+ tokens per second output rate.

Open source →
2026-10-27

MiniMax M3.1 Flash Superbench run record

Takeaway: Seven visual builds, including the four new Superbench 2.1 contracts, with two failed runs published as evidence.

The run record lists both delivery protocols, the per-brief reviewer conclusions and the limits of the delivery, including that the model version was read from the harness route rather than returned by the provider.

Open source →

Local benchmark run

Seven visual builds, including the four new Superbench 2.1 contracts

Superbash Learn sourceOctober 27, 2026Run record and published artifacts →
Sources, pricing details and methodology

Run identity

Model specifications

Pricing source · checked 2026-10-26 ↗
Status
Available in this Codex session
Codex model ID
MiniMax-M3.1-Flash
Run date
October 27, 2026
Access route
MiniMax API, M Plan, and MiniMax Code. The live API model ID is MiniMax-M3.1-Flash. Weights were not released at launch. The October 26-27, 2026 benchmark sessions run on the harness route labelled with that model ID and publish three Superbench 1.x core artifacts plus all four Superbench 2.1 briefs, including two failed runs and one failed required behaviour disclosed in the run record.

Flat rate with no long-context tier. The priority service tier is 1.5x standard.

Benchmark sources

Neon Drift, Ember Glider and Low-Poly Tower Defense were assigned to three separate MiniMax M3.1 Flash agent sessions on the Superbench 1.x core prompts. The four Superbench 2.1 briefs — Skyhook Cargo, Concourse Divide, Lightkeeper’s Atelier and Wrecklight Salvage — each went to a fresh isolated authoring agent that received one frozen prompt file and one output path. No reference implementation, sibling run, other model output or site source was supplied to any builder.

MiniMax M3.1 Flash Superbench visual benchmark run record

These are inspectable local artifacts, not provider benchmark scores, passed rubrics, or an aggregate rating. The run record separates the two delivery protocols. Per-artifact reports distinguish the checks that passed, the failures that were found, and the checks that remain not-observed; two of the four 2.1 briefs are published as failed-run evidence and a third failed a required behaviour.

Evaluation settings

    Local visual-run methodology and validation records

    Edition coverage

    Recorded means a published artifact, not a behavioral pass. Historical returning results retain their original protocol.

    Superbench 2.1: 6/6 recorded · 0 pending · 0 not run
    • 15 Ember Glider · repository-run · Recorded
    • 14 Low-Poly Tower Defense · repository-run · Recorded
    • 21 Skyhook Cargo · single-prompt · Recorded
    • 22 Concourse Divide · single-prompt · Recorded
    • 23 Lightkeeper's Atelier · single-prompt · Recorded
    • 24 Wrecklight Salvage · single-prompt · Recorded