AI BENCHY
Advertise here
#60

Step 3.7 Flash

Stepfun Release: 2026-05-29 Tested on: 2026-05-29 11:49 stepfun/step-3.7-flash::low
198B total (11B active)MoEOpen source
(high) (medium) (low)

Summary

Step 3.7 Flash scores 7.4 on AI BENCHY and ranks #60. It has 10.0 reliability, a 68.3% pass rate, $0.336 total cost, and 16.06s average response time.

What makes Step 3.7 Flash unique: It stands out most in Coding, where it ranks #1, while General Intelligence is its weakest area at #18.

Model facts

Researched on 2026-08-12

Reported
Parameters
198B total (11B active)
Architecture
MoE
Availability
Open source
License
Apache-2.0

Includes the vendor-reported vision component in the total.

Parameter source Availability source

Score

7.4

Consistency

8.7

Reliability

10.0

Total Cost (Current Price)

$0.336

Total Output Tokens

285,209

Total Input Tokens

0

Input Price

$0.200 / 1M

Output Price

$1.150 / 1M

Tests Correct

Wrong Tests: 8

Attempt pass rate: 68.3%

Flaky tests

3

Flaky tests had mixed outcomes across runs (at least one pass and one fail).

Response Time (avg)

16.06s

Response Time (max): 124.75s

Response Time (total): 321.11s

Wrong answer: 7 No answer: 1

Hamster playing table tennis

Prompt: Create a detailed SVG illustration of a hamster playing table tennis.

#60 Step 3.7 Flash

low
Invalid SVG
Cost
$0.004
Time
25.3s
Tokens
3,072 tok

Run history

Tested on Score Reliability Tests Correct Total Cost Compare
2026-08-14 02:29 Re-test 7.1 10.0 $0.390 Compare
2026-07-16 22:46 New test added 7.3 10.0 $0.454 Compare
2026-06-04 13:49 New test added 7.3 10.0 $0.341 Compare
2026-05-29 11:49 Initial run 7.4 10.0 $0.336 Current run

Run comparison

RunBenchmark coverageScoreConsistencyReliabilityTests CorrectFlaky testsTotal Output TokensTotal Input TokensTotal CostResponse Time (avg)
2026-05-29 11:49 · Initial run60/60 attempts7.48.710.012/203285,2090$0.33616.06s
2026-08-14 02:29 · Re-test66/66 attempts7.18.110.011/225320,393103,842$0.39017.87s
Difference—+0.3+0.60.0+1-2-35184-103842-$0.054-1810ms

Benchmark coverage differs: 60/60 attempts (Target: 3 repeats per test) versus 66/66 attempts (Target: 3 repeats per test). Totals and repeat-sensitive metrics are not directly comparable.

These two runs used different benchmark suites, so the deltas reflect both model changes and suite changes.

Charts

Choose the first model, then click a second model to open a side-by-side page.

Total Output Tokens

Score vs Total Output Tokens

Quick Compare

Step 3.7 FlashlowvsMuse Glimmer 30BlowStep 3.7 FlashlowvsSolar Pro 4highStep 3.7 FlashlowvsQwen3.5-122B-A10BmediumStep 3.7 FlashlowvsGrok 4.3mediumStep 3.7 FlashlowvsMuse Glimmer 30BmediumStep 3.7 FlashlowvsGPT-5.6 SolnoneStep 3.7 FlashlowvsQwen3.7 FlashlowStep 3.7 FlashlowvsDeepSeek V3.2mediumStep 3.7 FlashlowvsQwen3.5 Plus 2026-04-20mediumStep 3.7 FlashlowvsGLM 5.1medium

Category Breakdown

Category Score Consistency Tests Correct
Anti-AI Tricks 8.7 7.9
Coding 10.0 10.0
Combined 10.0 10.0
Data parsing and extraction 7.3 5.8
Domain specific 5.3 7.2
General Intelligence 3.4 9.3
Instructions following 9.8 10.0
Puzzle Solving 5.5 9.9
Tool Calling 10.0 10.0
Trivia 3.0 10.0

Compared models