Codex 'Dumbing Down' Claim Tested: Relay Accounts vs. Temporary Downgrades

Viral Claim: Codex Degraded for Relay Accounts

A widely shared post on X asserts that some Codex users experience degraded performance because OpenAI applies "soft risk control": accounts identified as relay/transit usage are silently "dumbed down" while still showing the Astra model label. The post proposes two tests using GPT-6 Astra medium: generate an SVG animation of a pelican riding a bicycle, and solve a candy probability puzzle.

Candy Puzzle Analysis: Two Interpretations, Two Answers

The puzzle: a bag contains candies of three flavors (apple, peach, watermelon) and two shapes (round, star). Counts are:

Round: 7 apple, 9 peach, 8 watermelon

Star: 7 apple, 6 peach, 4 watermelon

You must decide in advance how many candies to take to guarantee a pair of different shapes with apple + peach flavors (round apple + star peach, or round peach + star apple). The X post states shapes are distinguishable by touch.

If you can pick by shape (touch): Worst case you take all 9 round peach, all 7 round apple, all 8 round watermelon (24 round), then 4 star watermelon (4 star) = 28 candies without a valid pair. The 29th candy forces a star apple or star peach, completing a pair. But the puzzle says you can distinguish shapes by touch, so you would simply pick 9 round + 12 star = 21 candies, guaranteeing at least one round apple/peach and one star apple/peach. The correct answer under the stated rules is 21 .

If random draw without replacement (no touch): Answer becomes 29 as described above.

The author tested the original puzzle text with local Codex (GPT-6 Astra medium and high). Both tiers answered 21 with the selection 9 round + 12 star, and explained the minimality proof. An image shows the high-tier output with the step-by-step reasoning.

Conclusion: A model answering 29 misses the touch-distinguishability condition; but jumping from a wrong answer to "account deliberately throttled" lacks evidence — the error could be a simple reasoning slip.

Pelican Animation Test: Local Codex Succeeds

X users reported Astra (even Ultra) failing to draw a proper pelican-on-bicycle SVG. One user noted daily coding felt worse after a Codex reset, yet the pelican test passed.

The author ran the exact X prompt locally:

Astra medium: Generated an HTML file with animated SVG — wheels and legs move. GIF capture shows working animation.

Astra high: Three runs produced static SVGs that render correctly.

Caveat: Pelican drawing ability does not guarantee daily coding performance. Task modality matters — static SVG, HTML animation, or image-generation tool are different tasks; the prompt required SVG animation, not a raster image.

OpenAI Documentation: Temporary Downgrade Exists, But Not Relay-Specific

OpenAI's troubleshooting docs mention temporary downgrade triggered by suspicious activity: repeated login failures, access from unknown locations, etc. The downgrade may limit model access, hide model options, enforce usage limits, or follow account changes.

The documentation does not state that relay accounts are specifically targeted, nor that reasoning capability is secretly reduced while the Astra label remains visible.

Author's Observation: Post-Quota-Reset Degradation

The author notes a consistent pattern: after quota resets, Codex often feels "dumber." A previous post (linked) discussed this phenomenon. The cause remains unclear but is distinct from the relay-account conspiracy theory.