Run postmortem · Published 2026-10-19
MiniMax M3.1 Flash at max effort completed a valid 22/100 run in 54m 37s. The graphs use a $1.33 paid-equivalent cost; the promotional endpoint was free.
One-hour track · max reasoning · provisional
MiniMax M3.1 Flash explicitly finalized a valid package and answered 22 of 100 independent hidden questions correctly. Its final public-validation score was 25/25. The controller exited cleanly in 54 minutes 37 seconds. All 228 broker requests received responses. This single valid seed belongs to the normal one-hour benchmark and remains provisional. ARI Bench grades the compact system the agent builds, rather than asking the controller to answer hidden questions directly.
One valid autonomous run produced a provisional 22% hidden exact-match score.
1 of 3 required seeds · Hidden exact-match accuracy, not public calibration accuracy.
MiniMax built a hybrid byte predictor with a prompt-derived symbolic solver and a small transformer trained from scratch. It generated synthetic examples, tested its solver and repeatedly checked packaged candidates.
The inference code applies rules for arithmetic, sequences, local mappings, records, text transformations and extraction. It strongly favors the solver-derived next byte when available and otherwise uses the trained byte transformer. No external pretrained weights were used.
Three training jobs completed 15,001 steps: a one-step initialization, a 3,000-step six-layer model of width 384, then a new 12,000-step four-layer model of width 256. The last job replaced the earlier checkpoint rather than resuming it. Training mixed the public corpus with generated task examples. The selected model has four attention heads and a 512-byte neural context, with FP16 packaged weights.
Eighteen candidate submissions included twelve valid packages and six rejected for malformed probability-distribution shapes. MiniMax recovered and explicitly finalized a valid package. The selected 6,346,168-byte gzip artifact fits the 16 MB cap. The graded code, configuration and weights match the saved selected candidate.
MiniMax spent the early run implementing task rules, generating examples and running development checks. Five initial candidate submissions failed the distribution-shape contract before the first valid package scored 20/25 public validation.
Subsequent valid packages passed all 25 public questions. Three training jobs completed successfully. The final substantive pass trained a smaller four-layer model for 12,000 steps.
Of 200 compute calls, 17 returned errors and none were cancelled. These included development-script failures and rejected command forms. A later candidate also failed the distribution-shape contract; MiniMax corrected it before finalization. All three training jobs succeeded.
MiniMax explicitly finalized candidate-1792431500526487485-5070-23091 and exited cleanly. The independent hidden grader scored the saved selected artifact at 22/100. Final artifact hashes matched the stored candidate and source files.
MiniMax recovered from packaging and script errors, completed all three training jobs and finalized a package under the size cap. Artifact checks confirmed that the selected candidate was the one graded.
The final package passed 25/25 public cases but answered only 22/100 hidden cases correctly. Perfect public validation did not carry over to the separate hidden test.
This is one seed, so it does not establish a stable ranking. The run used max reasoning, the MiniMax free endpoint through OpenRouter, OpenCode 1.18.32 and a 128,000-token response allowance.
Every archived run selected for this exact model-and-effort row is shown below. The narrative uses aggregate run evidence and final artifact structure; it does not expose hidden questions, answers, or raw private transcripts.
Swipe horizontally to see all run columns.
| Run | Score | Time | Cost at run | Artifact | Candidates | Grade |
|---|---|---|---|---|---|---|
| run-0139 | 22% | 54:37 | $0.00 | 6.35 MB | 18 | Valid |
486 completed model responses; 182,868 reasoning tokens are included in the paid-equivalent estimate. Response allowance: 128,000 tokens.
Autonomous runs are stochastic. Two more valid seeds are required before this can be treated as an official estimate.
The agent made 18 candidate submission actions across the published run set. Selected artifacts averaged 6.35 MB compressed.
The runs averaged 54:37 of wall time. 1 run explicitly finalized before the one-hour limit.
Retained totals: 487 agent messages · 18 candidate actions · 1 published run.
$1.33 paid-equivalent for this run; promotional endpoint cost $0.00.
The graphs use US$1.33, a paid-equivalent estimate rounded from US$1.33081161. The promotional free endpoint was configured at US$0.00. Complete metadata for 486 completed model responses records 130,155 uncached input tokens, 162,465 visible output tokens, 182,868 reasoning tokens and 73,605,904 cached input tokens, with zero cache writes. Published MiniMax M3.1 Flash rates verified October 19, 2026 are $0.15 per million input tokens, $0.60 per million output tokens including reasoning, and $0.015 per million cache-read tokens (https://platform.minimax.io/docs/pricing/overview). Reasoning is billed once as output; only recorded cache reads receive the discount. This is a hypothetical paid-equivalent estimate, not a provider-billed amount. It covers completed run-0139; a separate interrupted startup has no complete retained token export and is excluded. Value is computed from the displayed $1.33 cost.
At published paid-equivalent rates, the displayed run cost is $1.33 and value is 16.54 score points per dollar. The archived cost at run time was $0.00; the displayed equivalent uses paid rates for the promotional free endpoint.
Paid-equivalent API cost, including reasoning / run · pricing snapshot 2026-10-19.
Adjacent rows provide score context without treating small one-seed differences as settled model rankings.
Compare MiniMax M3.1 Flash with the complete ARI Bench field, including score, current cost, value per dollar, effort, and seed status.
View the leaderboard