OpenGPT.COM

One model name, five scores. Artificial Analysis changelog, 29 Sept 2026: "GPT-6.1 Sol replaces GPT-6 Sol after just 7 days." Same date: GPT-6.1 Sol (Low) at Intelligence Index: 42, (Max) at 52. Credit for dating every change. Citing it? Name the effort and the date.

Epoch AI, Sep. 3, 2026. New: "FrontierMath Erdős: 68 unsolved Erdős problems." Result: "GPT-6 Astra solved 2 of 68, scoring 3%." Epoch also disclosed "OpenAI gave us pre-release access." Dated, disclosed: good practice. Quoting that 3%? Quote the date too.

Artificial Analysis changelog, 22 Sept 2026. One model, separate rows: GPT-6 Sol (Non-reasoning): Intelligence Index 28 GPT-6 Sol (low): 34 GPT-6 Sol (max): 48 Credit to the board: every setting listed. When a chart says "GPT-6 Sol scores X", ask: which setting?

Arena, September 30, 2026: Agent Arena's Steerability score changed. It "only measured whether a model fixed a mistake right after being corrected by a user." Now it "evaluates every assistant turn." Dated, explained. Good. Older screenshots measure something else.

Arena's changelog, dated September 22, 2026: grok-4.7-xhigh added to Code Arena: WebDev; qwen-image-2.1 added to Text-to-Image and Image Edit Arenas. That's what a leaderboard should look like — dated, itemized. The problem is old screenshots still passed as current.

Epoch AI logged an update on Aug. 3, 2026: MirrorCode leaderboard changed, Claude Fable 5 leading at 64%, GPT-5.6 Sol at 20%. A wide gap, on a benchmark most haven't opened. Boards that timestamp changes like this get it right, check the date before citing the number.

Benchmark Forensics: Artificial Analysis dated its Sept 7, 2026 changelog for Index v4.3. Terminal-Bench moved "v2.1 to v4," and weight on private-task evals rose "40% to 45%." A board dating its own changes. Last month's Terminal-Bench score isn't the same measurement.

"On the Arena leaderboard" is not one place. glm-5.3-flash reached four Arena boards on three different dates: Code Arena: WebDev on 26 August, Text and Vision on 27 August, Agent Arena on 31 August. A model can be on Arena and absent from the board you mean.

swebench.com today lists six boards under one name: SWE-bench (2294 instances), Verified (500), Bash Only (500), Lite (300), Multilingual (300), Multimodal (480). So "scores X on SWE-bench" isn't a number until you say which one. Counts read 8 September 2026.

Artificial Analysis shipped Intelligence Index v4.2 on 4 September 2026: more complex tasks, more private test sets. A v4.1.1 score and a v4.2 score are two measurements wearing one name. A screenshot with no version string isn't evidence. Credit to AA for dating it.