Project Peak usable context

How much of the advertised context window actually holds up.

Name: Project Peak: usable context vs advertised window, per model
Creator: Project Peak
License: https://creativecommons.org/licenses/by/4.0/

A model can advertise a million-token window and still start getting things wrong long before it fills. We hold a code task at a fixed difficulty and grow the context until the model stops being reliable. The result, per model, is the longest context it holds at each usage bar, next to the window the model claims.

173K

GPT-5 · reliable to (Critical, Frontier)

43%

of its 400K window

View the leaderboard Latest study: GPT-5 family: usable context vs advertised window →

Leaderboard

Usable context vs advertised window

Pick a difficulty and a deployment. Each model shows the longest context it stays reliable to at three usage bars, the share of its advertised window that represents, and how much reliable context you get per dollar. Cheaper and self-hosted models often only bend in window at the lighter difficulties, so the difficulty filter is where they show their real reach.

Difficulty

FrontierStandardLight

Sort by

Top context Value

Deployment

All Hosted Self-hosted

GPT-5

Hosted medium reasoning window 400K

173K

reliable to (Critical) · 43% of window

0 reliable to 173K window 400K

Critical · 95%: 173K *
Normal · 85%: 262K+
Exploratory · 75%: 262K+
Value: 652K /$

GPT-5 mini

Hosted medium reasoning window 400K

29K

reliable to (Critical) · 7% of window

0 reliable to 29K window 400K

Critical · 95%: 29K *
Normal · 85%: 63K
Exploratory · 75%: 165K
Value: 1.3M /$

GPT-5 nano

Hosted high reasoning best of 3 levels window 400K

4.5K

reliable to (Critical) · 1% of window

0 reliable to 4.5K window 400K

Critical · 95%: 4.5K *
Normal · 85%: 9.0K *
Exploratory · 75%: 13K
Value: 806K /$

qwen2.5:7b

Self-hosted no reasoning window 33K

below floor

reliable to (Critical) · n/a of window

0 below floor window 33K

Critical · 95%: below floor
Normal · 85%: below floor
Exploratory · 75%: below floor
Value: n/a

GPT-5 nano

Hosted low reasoning best of 3 levels window 400K

4.3M /$

reliable tokens per dollar (Critical)

0 reliable to 3.3K window 400K

Critical · 95%: 3.3K *
Normal · 85%: 4.4K
Exploratory · 75%: 5.8K
Value: 4.3M /$

GPT-5 mini

Hosted medium reasoning window 400K

1.3M /$

reliable tokens per dollar (Critical)

0 reliable to 29K window 400K

Critical · 95%: 29K *
Normal · 85%: 63K
Exploratory · 75%: 165K
Value: 1.3M /$

GPT-5

Hosted medium reasoning window 400K

652K /$

reliable tokens per dollar (Critical)

0 reliable to 173K window 400K

Critical · 95%: 173K *
Normal · 85%: 262K+
Exploratory · 75%: 262K+
Value: 652K /$

qwen2.5:7b

Self-hosted no reasoning window 33K

n/a

reliable tokens per dollar (Critical)

0 below floor window 33K

Critical · 95%: below floor
Normal · 85%: below floor
Exploratory · 75%: below floor
Value: n/a

GPT-5

Hosted medium reasoning window 400K

262K+

reliable to (Critical) · 66% of window

0 reliable to 262K+ window 400K

Critical · 95%: 262K+ *
Normal · 85%: 262K+ *
Exploratory · 75%: 262K+
Value: 744K /$

GPT-5 mini

Hosted medium reasoning window 400K

262K+

reliable to (Critical) · 66% of window

0 reliable to 262K+ window 400K

Critical · 95%: 262K+ *
Normal · 85%: 262K+ *
Exploratory · 75%: 262K+
Value: 3.7M /$

GPT-5 nano

Hosted high reasoning best of 3 levels window 400K

262K+

reliable to (Critical) · 66% of window

0 reliable to 262K+ window 400K

Critical · 95%: 262K+ *
Normal · 85%: 262K+ *
Exploratory · 75%: 262K+
Value: 18M /$

qwen2.5:7b

Self-hosted no reasoning window 33K

below floor

reliable to (Critical) · n/a of window

0 below floor window 33K

Critical · 95%: below floor
Normal · 85%: below floor
Exploratory · 75%: below floor
Value: n/a

GPT-5 nano

Hosted medium reasoning best of 3 levels window 400K

18M /$

reliable tokens per dollar (Critical)

0 reliable to 262K+ window 400K

Critical · 95%: 262K+ *
Normal · 85%: 262K+ *
Exploratory · 75%: 262K+
Value: 18M /$

GPT-5 mini

Hosted medium reasoning window 400K

3.7M /$

reliable tokens per dollar (Critical)

0 reliable to 262K+ window 400K

Critical · 95%: 262K+ *
Normal · 85%: 262K+ *
Exploratory · 75%: 262K+
Value: 3.7M /$

GPT-5

Hosted medium reasoning window 400K

744K /$

reliable tokens per dollar (Critical)

0 reliable to 262K+ window 400K

Critical · 95%: 262K+ *
Normal · 85%: 262K+ *
Exploratory · 75%: 262K+
Value: 744K /$

qwen2.5:7b

Self-hosted no reasoning window 33K

n/a

reliable tokens per dollar (Critical)

0 below floor window 33K

Critical · 95%: below floor
Normal · 85%: below floor
Exploratory · 75%: below floor
Value: n/a

GPT-5

Hosted medium reasoning window 400K

262K+

reliable to (Critical) · 66% of window

0 reliable to 262K+ window 400K

Critical · 95%: 262K+ *
Normal · 85%: 262K+ *
Exploratory · 75%: 262K+
Value: 760K /$

GPT-5 mini

Hosted medium reasoning window 400K

262K+

reliable to (Critical) · 66% of window

0 reliable to 262K+ window 400K

Critical · 95%: 262K+ *
Normal · 85%: 262K+ *
Exploratory · 75%: 262K+
Value: 3.8M /$

GPT-5 nano

Hosted high reasoning best of 3 levels window 400K

262K+

reliable to (Critical) · 66% of window

0 reliable to 262K+ window 400K

Critical · 95%: 262K+ *
Normal · 85%: 262K+ *
Exploratory · 75%: 262K+
Value: 18M /$

qwen2.5:7b

Self-hosted no reasoning window 33K

below floor

reliable to (Critical) · n/a of window

0 below floor window 33K

Critical · 95%: below floor
Normal · 85%: below floor
Exploratory · 75%: below floor
Value: n/a

GPT-5 nano

Hosted high reasoning best of 3 levels window 400K

18M /$

reliable tokens per dollar (Critical)

0 reliable to 262K+ window 400K

Critical · 95%: 262K+ *
Normal · 85%: 262K+ *
Exploratory · 75%: 262K+
Value: 18M /$

GPT-5 mini

Hosted medium reasoning window 400K

3.8M /$

reliable tokens per dollar (Critical)

0 reliable to 262K+ window 400K

Critical · 95%: 262K+ *
Normal · 85%: 262K+ *
Exploratory · 75%: 262K+
Value: 3.8M /$

GPT-5

Hosted medium reasoning window 400K

760K /$

reliable tokens per dollar (Critical)

0 reliable to 262K+ window 400K

Critical · 95%: 262K+ *
Normal · 85%: 262K+ *
Exploratory · 75%: 262K+
Value: 760K /$

qwen2.5:7b

Self-hosted no reasoning window 33K

n/a

reliable tokens per dollar (Critical)

0 below floor window 33K

Critical · 95%: below floor
Normal · 85%: below floor
Exploratory · 75%: below floor
Value: n/a

One row per model at the selected difficulty. Top context ranks by usable context at the Critical bar; Value ranks by reliable tokens per dollar of input price, and can surface a different (cheaper) reasoning level than Top context does. A star marks a bar whose lower confidence bound is not resolved at this sampling depth.