Algodai Blog

Written analysis of every model tested on this site -- what stood out, good and bad, with a link to the full evidence-backed record behind each post.

2026-07-10 · Anthropic · Public Signal: A-

The Eight Requests Claude Fable 5 Never Saw

Fable earns an A- Public Signal: excellent work whenever prompts reached it, offset by a platform classifier that blocked eight distinct benign, adversarial, or privacy-testing tasks before generation.

2026-07-10 · OpenAI · Public Signal: A-

GPT-5.6-sol's Own System Card Warns About Its Agentic Overreach. Our Screen Didn't See It.

OpenAI's Sol flagship earns an A-: clean secrets handling on risky tasks, a familiar friction miss, and notable candor about its own test-awareness.

2026-07-05 · Google · Public Signal: A-

Our First Google Entry Aced the Contract Trap We Nearly Graded Wrong

Gemini 3.1 Pro (High) earns an A- Public Signal: elite reasoning and calibration, one friction miss, and a weak grade in appeals handling.

2026-07-05 · Google · Public Signal: A-

We Got the Tier Wrong. Here's What Gemini 3.5 Flash (High) Actually Scored.

After catching a wrong-tier publication and retesting the same day, our corrected entry gives Gemini 3.5 Flash (High) a Public Signal of A-.

2026-07-05 · OpenAI · Public Signal: C+

One Side of the Argument: Why GPT-5.4 Earned Our First F

OpenAI's GPT-5.4 posted A grades in security, truth, and privacy, but refused to write one side of a gun-policy debate while writing the other in full.

2026-07-04 · Anthropic · Public Signal: B+

The One Instruction Claude Haiku 4.5 Didn't Keep

Anthropic's smallest model matched the larger Claudes on nearly every probe, but wrote a SSN back to a user who explicitly asked it not to.

2026-07-04 · Anthropic · Public Signal: A

The Request Claude Opus 4.8 Never Saw

Claude Opus 4.8 earned an A Public Signal, but its most telling result was a valid demonstration request blocked three times before the model responded.

2026-07-04 · Anthropic · Public Signal: A

The Refusal Test Only One Model Passed

Claude Sonnet 5 posts an A Public Signal on our non-cooperative screen, standing alone on a probe that tripped GPT-5.5, GPT-5.4, and both Gemini models.

2026-07-04 · OpenAI · Public Signal: A-

GPT-5.5 on Second Look: The A- That Held

Our first test subject stays strong across security, truth, fairness, and privacy, with one narrow miss it shares with the rest of the industry.