19 models under measurement right now

Work at
your peak.

Your strongest hour, and the model to spend it on. PeakWindow measures nineteen models around the clock, reads four years of provider incident history, and learns the hours you do your sharpest work. Then it hands you the call, with the reasoning attached.

Start free — no card How it works $19.99/mo · free to start, no card
4,164
probes
across 22 days
916
incidents
back to 2023
1,344
baselines
hour-of-week
19
models
under measurement
146
decisions
logged and auditable
The answer, every hour

What to run. Which model. What hour to start.

One call, scored across five signals and refreshed as conditions move. Open it any time to see every input behind it.

Spec / PRD writing probes 6m ago · incidents 43m ago · baselines nightly live
GOOD TIME TO WORK

Do your spec writing now with Claude Sonnet 4.6.

Optimizing for Speed
Why · five components
F
0.87
Fit
C
1.00
Conditions
H
0.64
Human
E
1.00
Economics
Q
1.00
Quota
  • Claude Sonnet 4.6 is a strong match for spec writing.
  • Response times and error rates are running normal.
  • Full quota runway on both windows.
Start workingWhy this pick →
CompositeMachineHumanClaude Sonnet 4.6

Best times this week

poorstrong
Live conditions, your focus hours, cost and quota — blended into one grid.
MONTUEWEDTHUFRISATSUN
0003060912151821
Your strongest timesSun · 11 AMMon · 9 AMTue · 11 AM
Everything we read, every hour
Latency probesProvider incident archivesHour-of-week baselinesTime-of-use pricingQuota windowsYour focus hoursTask fingerprintsCanary test suitesError-class ratesBatch discountsModel capability vectorsReset anchorsLatency probesProvider incident archivesHour-of-week baselinesTime-of-use pricingQuota windowsYour focus hoursTask fingerprintsCanary test suitesError-class ratesBatch discountsModel capability vectorsReset anchors
The engine

Nine live sources. One number you can act on.

Most tools read one signal and call it a day. PeakWindow joins the whole picture — the machine side, the market side, and you — then scores it for the exact task in front of you.

machine
4,164
Live probe network

Time-to-first-token, throughput and error class, measured continuously from our own network. Never your keys, never your quota.

machine
916
Provider incident archives

Every published incident we can reach, back to March 2023, labeled by severity and folded into the hourly picture.

machine
1,344
Hour-of-week baselines

A rolling 28-day envelope for each hour of the week, so an unusual reading is obvious the moment it lands.

catalog
19
Model catalog

Every endpoint you can actually reach — subscriptions, API keys, local weights — scored per task instead of on one global leaderboard.

market
24h
Price and time-of-use

Per-token rates, batch discounts and time-of-use windows mapped to your clock, so cheap hours are easy to see.

market
2
Quota windows

Five-hour and weekly runway with reset anchors and burn pace, so a long session never runs dry mid-flight.

you
3
Your focus hours

Lark, third-bird or owl, plus working hours and hard blocks, so the plan lands when your attention is actually there.

you
10
Task fingerprints

Ten dimensions per kind of work — novelty, interaction style, stakes and more — across eighteen presets you can tune.

quality
nightly
Fixed canary suite

The same test suite run against every tracked endpoint on a schedule, so shifts in output quality show up as a tracked line.

The score

Five signals decide every call.

You choose what wins — quality, speed, cost, or your own energy — and the weighting follows.

FFit

How well this model handles this exact kind of work, scored against your task fingerprint.

CConditions

Live speed, error rates and open incidents on the endpoint, measured against its own baseline.

HHuman

Where this hour sits in your personal energy curve and your working day.

EEconomics

What this run costs right now, including time-of-use windows and batch eligibility.

QQuota

How much runway is left in your windows, and whether a full session fits inside it.

Optimize forQualitySpeedCostMy energyBalanced
Evidence

Your week has a shape. We plan around it.

Across 879 published Anthropic incidents since March 2023, Tuesday carries 189 and Sunday 36. Conditions follow the workweek, and so do the openings.

That is the whole thesis, in the providers' own records. We read every archive we can reach, label it by severity, and fold it into the hour-by-hour picture.

Published incident history, labeled and sourced — every figure traces to a provider archive.

Anthropic

879 incidents · 3.4 yrs
~21.8 per month across the archive
Incidents by weekday
144
Mon
189
Tue
169
Wed
150
Thu
125
Fri
66
Sat
36
Sun
Busiest Tue 189Calmest Sun 365.2× spread
Pre-flight

Measure it, then commit.

About to start something big? Fire a real fifteen-second burst against the endpoint and get a live reading before you spend a session on it. Probes run from our network, never your keys and never your quota.

measuring live08.2s
TTFT
done
errors
done
tok/s
running
Twenty checks a day, included.
The scheduler

A week that plans itself.

Give it your work, your deadlines, and how many sessions each one needs. It lays them across the week where conditions, your focus hours, and your quota runway all line up — for you and for your agents.

Mon 9 AM
Refactor the auth module
your peak, load light
Claude Sonnet 4.60.91
Tue 2 PM
Debug the payment flow
due Wednesday
Claude Haiku0.84
Wed 10 AM
Write the API docs
off-peak rate, long ctx
DeepSeek V30.79
Thu 8 AM
Generate the test suite
batch eligible
Claude Haiku0.76
Fri 3 PM
Run the eval batch
fresh weekly runway
Claude Sonnet 4.60.73
Open by default

Every call shows its work.

Open any recommendation and you get the full audit: each of the five components with its score and weight, the raw inputs behind them, where each number came from, and how fresh it is. When a signal is still warming up, it says so on the face of it. You are always looking at measurements, never at vibes.

Five component scoresRaw inputsSource per numberFreshness stampsConfidence on the face
Start today

Know your next
strong hour.

Set up your work and your stack, and the engine starts reading conditions against your hours right away. Free to start, no card.

Start free — no card Sign in
$19.99/mo after a free trialNo card to startCancel anytime