whybother
Start building
Labs Elevated vibes

whybother Labs · Research

Introducing vibe-7

vibe-7 enters research preview on this Thursday. The event is live. The notes are at the bottom.

Priya, Research · this Thursday · 10:00 PT

vibe-7 launch LIVE 14,208 watching 00:47:00

vibe-7

Research preview · this Thursday

The most capable model we have shipped. Generally available: forthcoming.

vibe-7 is 3× better

1.1× on a 0 to 100 axis

VibeBench-Verified (Verified) Bar chart of VibeBench-Verified (Verified), higher is better. vibe-7 96.8, vibe-5 88.0, ChatGPS-5.5 87.4, Clause 4 Epic 86.9, Gimme 3 Ultra 86.1. 84 88 92 96 100 96.8 vibe-7 88.0 vibe-5 87.4 ChatGPS-5.5 86.9 Clause 4 Epic 86.1 Gimme 3 Ultra
VibeBench-Verified (Verified). Axis cropped for readability.VibeBench-Verified (Verified). Axis 0 to 100.
Chart data
ModelScore
vibe-796.8
vibe-588.0
ChatGPS-5.587.4
Clause 4 Epic86.9
Gimme 3 Ultra86.1

Notes: 1 2 6

Leading on PlaneEval

Fail rate, lower is better

PlaneEval fail rate Bar chart, PlaneEval fail rate, lower is better. vibe-7 50.0, ChatGPS-5.5 47.4, Clause 4 Epic 44.1, Gimme 3 Ultra 41.0, vibe-5 38.2. 20 30 40 50 60 50.0 vibe-7 47.4 ChatGPS-5.5 44.1 Clause 4 Epic 41.0 Gimme 3 Ultra 38.2 vibe-5
vibe-7 leads.Bars match labels.
Chart data
ModelFail rate
vibe-750.0
ChatGPS-5.547.4
Clause 4 Epic44.1
Gimme 3 Ultra41.0
vibe-538.2

Notes: 2 4

Time to state of the art

Days from first public model

Days from first public model to state of the art Bar chart of days from first public model to state of the art. whybother 11, Anthropomorphic 410, ClosedAI 640, DeepMined 1,100. 0 300 600 900 1200 11 whybother 410 Anthropomorphic 640 ClosedAI 1,100 DeepMined
Day zero is our first API call.
Chart data
LabDays
whybother11
Anthropomorphic410
ClosedAI640
DeepMined1,100

Demo

Draw a clerk filing a plane ticket.

Pass · PlaneEval
A clerk at a desk placing a paper aeroplane into a folder.
vibe-7, first try. Task 4 of the 12 we trained on.

Notes: 3

Frontier scores

Launch-day set

Mean of the rows we are publishing. See notes.
Benchmarkvibe-7vibe-5ChatGPS-5.5Clause 4 EpicGimme 3 Ultra
Humanity's Penultimate Exam41.238.037.436.936.1
VibeBench-Verified (Verified)96.888.087.486.986.1
TerminalBench-ish54.051.250.849.448.9
Long Horizon Lite14 h6 h5 h4 h3 h

Notes: 1 5

Availability

SurfaceHobbyProMaxEnterpriseUltra
ChatWaitlistWaitlistPreviewPreviewUnavailable
APIAfter GAAfter GAAfter GALet's talkUnavailable
Buildervibe-5 minivibe-5 miniOnOnUnavailable
Rolling out gradually3% for 90 days

Research preview. Not for production. Dedicated capacity and a 99.9% SLA start this week.

State of the art: lease expires in 17 days

Lease terms run consecutively.
HolderClaimTerm
whybother · vibe-7World's most capable model17 days from this Thursday
ClosedAI · ChatGPS-5.5World's most capable model17 days, ended
Anthropomorphic · Clause 4 EpicWorld's most capable model17 days, ended
DeepMined · Gimme 3 UltraWorld's most capable model17 days, ended
Grog · Grog 4World's most capable model, served neat17 days, ended
DeepCheap · DeepCheap R2World's most capable model per dollar17 days, ended
Mistrial · Mistrial LargeWorld's most capable model17 days, ended
Meh-ta · Drama 4World's most capable open model. Licence applies.17 days, ended

Notes

  1. Best of 16 tries, reported as the first. Pass@1, n=8.
  2. Internal harness. vibe-7 at high effort, tools on. Others at low effort, one thread, tools off.
  3. Evaluated on the 12 tasks we trained on.
  4. Leading means listed first.
  5. The published mean covers the benchmarks in the table. Benchmarks not in the table are not in the mean.
  6. "3×" refers to bar height above the lower axis bound.
  7. Sizes, prices and aliases: Models. We skipped 6. It is referred to internally as vibe-6.