Sources and method
Trace every model specification and score to its published source. Sources checked through 17 October 2026.
Vals AI
vals.ai30 independent leaderboards · 28 model pages · 1 benchmark page
- Pages read
- 59
- Figures
- 2,172
Artificial Analysis
artificialanalysis.ai26 independent leaderboards · 22 model pages
- Pages read
- 48
- Figures
- 4,895
OpenAI
openai.com5 model pages · 2 published tables
- Pages read
- 7
- Figures
- 70
Ollama
ollama.com6 model pages
- Pages read
- 6
- Figures
- 0
Anthropic
anthropic.com4 published tables
- Pages read
- 4
- Figures
- 244
arxiv.org
arxiv.org3 benchmark pages
- Pages read
- 3
- Figures
- 0
huggingface.co
huggingface.co2 published tables · 1 benchmark page
- Pages read
- 3
- Figures
- 175
Reflection AI
reflection.ai2 model pages · 1 published table
- Pages read
- 3
- Figures
- 112
Anthropic (Claude Platform)
claude.com2 model pages
- Pages read
- 2
- Figures
- 0
arcprize.org
arcprize.org1 benchmark page · 1 independent leaderboard
- Pages read
- 2
- Figures
- 17
deepseek.com
deepseek.com2 model pages
- Pages read
- 2
- Figures
- 0
github.com
github.com2 benchmark pages
- Pages read
- 2
- Figures
- 0
tbench.ai
tbench.ai2 benchmark pages
- Pages read
- 2
- Figures
- 0
Z.ai
z.ai2 published tables
- Pages read
- 2
- Figures
- 135
blog.google
blog.google1 model page
- Pages read
- 1
- Figures
- 0
centerforcivicfutures.org
centerforcivicfutures.org1 benchmark page
- Pages read
- 1
- Figures
- 0
codeforces.com
codeforces.com1 benchmark page
- Pages read
- 1
- Figures
- 0
cursor.com
cursor.com1 benchmark page
- Pages read
- 1
- Figures
- 0
datacurve.ai
datacurve.ai1 independent leaderboard
- Pages read
- 1
- Figures
- 28
epoch.ai
epoch.ai1 benchmark page
- Pages read
- 1
- Figures
- 0
exploitbench.ai
exploitbench.ai1 benchmark page
- Pages read
- 1
- Figures
- 0
frontierswe.com
frontierswe.com1 benchmark page
- Pages read
- 1
- Figures
- 0
google.dev
google.dev1 model page
- Pages read
- 1
- Figures
- 0
googleapis.com
googleapis.com1 published table
- Pages read
- 1
- Figures
- 73
ioi2024.org
ioi2024.org1 benchmark page
- Pages read
- 1
- Figures
- 0
lastexam.ai
lastexam.ai1 benchmark page
- Pages read
- 1
- Figures
- 0
livecodebench.github.io
livecodebench.github.io1 benchmark page
- Pages read
- 1
- Figures
- 0
matharena.ai
matharena.ai1 benchmark page
- Pages read
- 1
- Figures
- 0
mmmu-benchmark.github.io
mmmu-benchmark.github.io1 benchmark page
- Pages read
- 1
- Figures
- 0
posttrainbench.com
posttrainbench.com1 benchmark page
- Pages read
- 1
- Figures
- 0
scale.com
scale.com1 independent leaderboard
- Pages read
- 1
- Figures
- 11
sec-bench.github.io
sec-bench.github.io1 benchmark page
- Pages read
- 1
- Figures
- 0
skillsbench.ai
skillsbench.ai1 benchmark page
- Pages read
- 1
- Figures
- 0
stanford.edu
stanford.edu1 benchmark page
- Pages read
- 1
- Figures
- 0
swe-marathon.com
swe-marathon.com1 benchmark page
- Pages read
- 1
- Figures
- 0
swebench.com
swebench.com1 benchmark page
- Pages read
- 1
- Figures
- 0
toolathlon.xyz
toolathlon.xyz1 benchmark page
- Pages read
- 1
- Figures
- 0
withprotege.ai
withprotege.ai1 benchmark page
- Pages read
- 1
- Figures
- 0
xlang.ai
xlang.ai1 benchmark page
- Pages read
- 1
- Figures
- 0
Zapier
zapier.com1 benchmark page
- Pages read
- 1
- Figures
- 0
zerobench.github.io
zerobench.github.io1 benchmark page
- Pages read
- 1
- Figures
- 0
How a number gets here
Every score retains its evaluator, method and source.
Independent results come from Vals AI, Artificial Analysis, Datacurve, ARC Prize · Standard and Scale AI / SEAL: results retain each source’s task set and run settings. Compare scores within the same column.
Vendor tables are reproduced from the announcement in which each lab published them, including the columns that lab chose and the models it chose to compare against. That is why they sit in separate tables: a lab's own run and an independent run answer different questions.
Model specifications — release date, context window, maximum output, list price and whether weights are open — are read from the source listed on each model page, with evaluator-recorded prices identified in its summary. Prices are in US dollars per million tokens and exclude caching, batch and committed-use discounts.
Hosted-run facts — the capabilities a runtime advertises, its own context cap, its parameter count and its per-token prices — are read from the runtime's own model page and labelled as that runtime's figures, because the version a caller reaches and the model a lab describes are not always the same build. Where the two disagree, both are printed.
Derived views — the ranks, field spreads, medians and lab standings shown on Rankings and on each page — are computed from the tables above at build time, never stored. A rank is claimed only inside a field this site can see, and no figure is ever averaged across two publishers.
What this site does not do
- It does not run models, and it has no independent measurements of its own.
- It does not average vendor claims together or reconcile disagreements between publishers.
- It does not rank models by a single composite score, or claim that a higher number means a better product.
- It does not estimate a missing value; a blank means no published result is present in the sources tracked.
- It does not claim a rank outside the rows it read: where a suite scores hundreds of models, this site ranks only the models it lists and says so on the page.
Send corrections through the support link in the footer. Corrections are checked against the published source.