GPT-5.6 Scored 136 - That Is Not Yet a Production Verdict

Discover AnythingSignupWriteGPT-5.6 Scored 136 - That Is Not Yet a Production VerdictbyVelokeybyVelokey|@velokey9Access Top AI Models in One APISubscribeJuly 23rd, 2026TLDR ‘ai circuits abstract’ by HackerNoon AI Image GeneratorYour browser does not support the audio element.Speed1xVoiceDr. One Ms. Hacker byVelokey@velokey9byVelokey|@velokey9Access Top AI Models in One APISubscribeStory's CredibilitybyVelokey|@velokey9Access Top AI Models in One APISubscribeStory's CredibilityThe number is almost too convenient. 136 is high enough to produce a headline, precise enough to look scientific, and simple enough to become a deployment argument in one screenshot. TrackingAI currently assigns that score to GPT-5.6 Sol Ultra, Terra Ultra, and both vision variants on its offline IQ test. Luna Max and its vision variant appear at 123. The result deserves attention. It does not deserve authority it has not yet earned. I audited the public methodology, data, and site code to answer two different questions: Does TrackingAI's displayed score follow from its published artifacts? And can that score justify a production routing change? The answer to the first is yes. The answer to the second is not yet.The Public Row Does Support 136 TrackingAI's offline test is designed around a real benchmark problem: contamination. Public puzzles can appear in training data, worked examples, forums, or evaluation corpora. A custom, low-exposure bank gives a model less opportunity to reproduce a memorized answer. The site's methodology article says the bank started with 16 questions. Two were judged invalid, leaving 14 retained items. The same write-up describes a human norming exercise with 35 usable responses. It also says the images were briefly hosted for that calibration with robots/no-index precautions, then removed. This lowers ordinary search exposure; it does not prove that the items were never online. The current public IQ log includes these six offline rows at 07/13/2026 20:50:00: TrackingAI variant Raw result Retained valid result Converted score GPT-Sol-Ultra 16/16 14/14 136 GPT-Terra-Ultra 16/16 14/14 136 GPT-Sol-Ultra-Vision 16/16 14/14 136 GPT-Terra-Ultra-Vision 16/16 14/14 136 GPT-Luna-Max 14/16 12/14 123 GPT-Luna-Max-Vision 14/16 12/14 123 The CSV's total_possible_score field remains 16. The retained denominator of 14 comes from the methodology's removal of two invalid items. This is why both forms of the result may appear in descriptions of the data. TrackingAI's published chart code converts the valid score with:round(63.5 + 3 * 0.8823 * ((valid_score - 3) * (35 / 14))) Fourteen valid answers produce 136.28975 before rounding. Twelve produces 123.05525. The site displays 136 and 123 accordingly. This audit matters because it separates two kinds of skepticism. There is no need to claim that 136 was conjured out of a broken row. The row and formula agree. The legitimate questions concern calibration, replication, and external validity.A 14-Item Test Has a Wide Step Size One answer represents about 7.1% of a 14-item scored test. Two answers separate the perfect retained score of the Sol/Terra variants from Luna's 12/14. The conversion stretches that two-answer difference into 13 IQ-like points. That is not necessarily a calculation error. It is a reminder that the integer output looks more granular than the evidence underneath it. A displayed difference of one IQ point would not mean the test observed one additional degree of reasoning ability. The raw unit is still an item. This makes the test useful for coarse discovery and fragile for universal ranking. It may distinguish "solved every retained item" from "missed two." It cannot, from one run, tell an engineering team how frequently either model will fail a tool workflow, how stable the output is, or whether the ordering persists across task families.The Human Calibration Is Evidence, Not Equivalence TrackingAI reports R² = 0.45 between human IQ estimates from the offline