Step 3.7 Flash vs Step 5 Preview

Wins 0 of 5 benchmarks
$1.03 / task
Wins 2 of 5 benchmarks
Different efforts: high · thinking · Compare all at high
Humanity's Last Exam Humanity's Last Exam: Very hard expert questions across many fields, written to stump today’s models
Step 3.7 Flash21.4
Step 5 Preview46.5
Terminal-Bench 4.0 Terminal-Bench 4.0 · source setting 2: Agents use a terminal to complete executable coding, system administration and research tasks in the 4.0 suite
Step 3.7 Flash—
Step 5 Preview33.3
AutomationBench-AA AutomationBench-AA: Agents orchestrate SaaS workflows through APIs; objectives receive credit only when the task has no guardrail violation
Step 3.7 Flash—
Step 5 Preview51
Arena Text · creative writing Arena Text · creative writing: Human pairwise preferences in the stated text vote pool
Step 3.7 Flash—
Step 5 Preview1,403
MMMU-Pro (vision) MMMU-Pro (vision): College-level questions that require understanding images: charts, diagrams, photos
Step 3.7 Flash75.3
Step 5 Preview76.4

Specs

SpecStep 3.7 FlashStep 5 Preview
Input, per 1M$0.20$1
Output, per 1M$1.15$2.70
Cost per task?—$1.03
Speed, tok/s12393
Context256K1M
WeightsOpenOpen
ReleasedMay 2026Sep 2026
Memory (4-bit)121 GB363 GB
LicenceApache-2.0custom licence
Commercial useAllowedRestricted
See images✓✓
Read PDFs——
Hear audio——
Watch video✓✓
Call tools✓✓
Force a tool callPartialPartial
Return strict JSON✓✓
Search the web?—
Run code?—
Use a computer??

What breaks when you switch

Step 3.7 Flash → Step 5 Preview
Memory (4-bit): 121 GB → 363 GB
web_search not available · Chat Completions›

Built-in Chat Completions web_search is unavailable; integrate external search through a function tool and return the results.

Output includes reasoning›

Output billing includes the reasoning process as well as the final answer.

LRU eviction during peak traffic›

LRU eviction removes unused cache entries more aggressively during peak traffic, so cached prefixes may not survive.

Step 5 Preview → Step 3.7 Flash
Context: 1M → 256K