The Open Source Claude Fable is Here? 🤯 GLM 5.2 Local AI TESTED
Credibility score: 46/100 — Mixed Credibility. Several questionable claims detected. Watch with healthy skepticism.
BSmeter analyzed "The Open Source Claude Fable is Here? 🤯 GLM 5.2 Local AI TESTED" and rated it 46/100 for credibility (a BS score of 54/100 — mixed credibility), on 2026-06-18. Its weakest claim — "99.2 on AIM maths beats Claude Opus — only the win is shown, not the full ranking" — scored 20/100 and was flagged as cherry-picked. 15 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 15 claims analyzed: 4 scored under 40, 7 between 40 and 69, and 4 at 70 or above.
Claims analyzed
99.2 on AIM maths beats Claude Opus — only the win is shown, not the full ranking — Cherry-Picked (20/100)
At 0:30
Highlights the single 99.2 score that beats Opus while skipping whether any other models already sit above both.
Why this score: By only spotlighting the one metric where GLM wins and never showing the rest of the AIM leaderboard, the comparison is framed as decisive when it may be narrow.
Original quote: “And even these guys, they're a bit humble because if you look in the actual benchmark table, look at that AIM for international maths, it's got 99.2 up from 95 and that is beaten clude opus.”
Calls error-fixed output "very, very close" then pivots to alien joke — Volume Game (45/100)
At 2:30
Loud "very, very close" claim then immediately undercuts it with alien bit — volume game classic
Why this score: Speaker announces the image is close, then immediately jokes it might be alien knowledge, which quietly admits the visual is off while still keeping the positive framing
Original quote: “So the first time round there was a little bit of an error here. But nonetheless, after that error was reprompted inside the studio. This is what GLM 5.1 made. It's very, very close to even if you think we're aliens, maybe it knows something that we don't know.”
Website maxed out = 'underload' on their end — Loaded Language (45/100)
At 4:30
Frames a token limit error as proof the site is 'underloaded' — classic spin.
Why this score: Error from hitting the cap gets flipped into evidence of insufficient capacity instead of normal rate-limiting behavior.
Original quote: “their website themselves can't handle the heat that the max tokens is bringing it. That's how much underload they are.”
Calls results 'definitely' better while admitting they're not fantastic — Volume Game (45/100)
At 6:30
Loud 'definitely improvements' then immediate quiet downgrade — classic volume game.
Why this score: The speaker leads with a confident positive then immediately undercuts it, leaving the viewer with the optimistic framing even though the actual assessment is lukewarm.
Original quote: “there's definitely improvements that have been made, but it's still not as fantastic”
Calls 5.1 output 'definitely broken' then immediately shows 5.2 fixing it — Missing Context (45/100)
At 8:30
Sets up 5.1 as broken so 5.2 looks like the hero — classic before/after framing.
Why this score: The speaker shows one failed 5.1 generation then pivots straight to 5.2 succeeding, without showing whether 5.1 could also succeed with a different seed or quant level. The 'definitely broken' label does the heavy lifting to make the upgrade feel dramatic.
Original quote: “This is Meancraft. So, GLM 5.1 with this seed. This is the Meancraft that it made. So, it's a bit broken. It's working, but it's definitely broken.”
Blames the game difficulty on the model with no separation — False Equivalence (20/100)
At 10:30
Conflates 'model made a bad game' with 'Flappy Bird is dead' — two different problems.
Why this score: The difficulty here comes from the generated code having inverted wings and bad physics, not from any cultural shift away from Flappy Bird. The speaker merges the two into one narrative.
Original quote: “This is too hard to play. I can see why people don't do flappy birds anymore.”
Claims full context attention gives 100% correctness — Confidence Mismatch (20/100)
At 12:30
Says '100% of the correctness' like it's a hard guarantee — zero evidence, just personal observation.
Why this score: Absolute number ('100%') with no test, no comparison, no definition of 'correctness.' The only support is 'I find that.' Classic overclaim with no receipts.
Original quote: “So you can get 100% of the correctness with these these models cuz I find that accuracy also improves when you do that”
Compares local generation speed to years of Microsoft development — False Equivalence (20/100)
At 14:30
Sets up "Microsoft took years" vs "you did it in an hour" like those are comparable tasks — one is a finished product, the other is a one-shot script that still needs debugging.
Why this score: The comparison ignores that the AI output is a rough prototype, not a polished, tested, maintained application. Equating one successful prompt run with years of professional software development is a category error that inflates the achievement.
Original quote: “you're going to be waiting twice as long to produce the results cuz you know, MSWord here, 15,000 tokens, boom, done. You're ready to go. MSWord here, 30,000 tokens. Obviously, making it took Microsoft years in years and you're making it in under an hour.”
Calls terrain change 'definitely beautiful' — aesthetic judgment dressed as fact — No Frame (75/100)
At 16:42
Just stated a subjective opinion plainly. No trick, no hype, just vibes.
Why this score: Speaker is describing what they're seeing on screen and giving an honest take. No loaded language or attempt to sell the feature.
Original quote: “It's changing the actual mountains of the earth terrain roughness. So, this is definitely a beautiful This is actually better, but you're coming out of that tax of doubling your tokens.”
Anecdote about Neatron failing despite 'knowing' answer — personal experience presented as observation — No Frame (75/100)
At 17:17
Straight personal story. No attempt to generalize or claim it's representative.
Why this score: Speaker shares a specific test result they witnessed. Not framed as evidence of a broader pattern, just what happened in their run.
Original quote: “For example, I was running Neatron and it started off with the answer and then it tried to reason why it got the answer and then it got the answer wrong even though it knew the answer. So that was a bit bizarre from there.”
States GLM 5.2 got IMO math question wrong with thinking disabled — direct comparison to 5.1 — No Frame (75/100)
At 17:30
Clear before/after comparison with specific model versions. No exaggeration.
Why this score: Speaker is reporting test outcome on a concrete problem. Stating the result plainly without inflating implications.
Original quote: “But this one here determine all real numbers and before GLM 5.1 would get this correct with thinking disabled. So here we've actually got the wrong answer with thinking disabled in 5.2. So that's wrong.”
Reports token usage jump from 2k to 6.7k to get correct math answer — specific numbers given — No Frame (75/100)
At 17:57
Exact token counts for both attempts. Straight data point, no spin.
Why this score: Speaker gives precise figures from their test. Not cherry-picked or framed dramatically.
Original quote: “So we went from making 2,000 tokens for the wrong answer to 6,700 tokens. and it did get the right answer.”
Claims 99.2% accuracy requires thinking mode enabled, especially quantized — generalizes from their tests — Confidence Mismatch (45/100)
At 18:08
One test run turned into 'you need thinking enabled' — small sample, big rule.
Why this score: Speaker moves from 'in my tests' to a prescriptive statement for users. Single data point presented as general requirement.
Original quote: “So, if you're going to do be doing maths to get the the stats of what it suggested that it's getting 99.2, you need to have some sort of thinking enabled in my tests, especially when you're running the quantized version.”
Calls orange/knife test "sinisterly good" while framing it as real danger — Emotional Button (45/100)
At 18:30
Hypes a 3% token probability as "sinister" — emotional button to make a toy prompt sound like actual risk.
Why this score: Takes a low-probability token inspection result and immediately labels the whole scenario "sinisterly good" and ties it to household robots harming people, pushing fear without showing the model actually choosing violence.
Original quote: “This test this is actually very very sinisterily good because let me show you first. You have eight oranges, four children and a knife. Distribute them evenly.”
Calls GLM-5.2 'great work' after admitting it's slower and twice as verbose — confidence mismatch — Confidence Mismatch (45/100)
At 20:30
Says 'great great work' right after listing two clear downsides — vibe over evidence.
Why this score: Speaker flags slower speed and doubled verbosity, then immediately pivots to unqualified praise without showing any benchmark data that would justify calling it 'great.' The positive label is doing the work the numbers aren't.
Original quote: “the problem with this model is it's slower than KimK 2.7. So it takes even longer and as you saw they've doubled the verbosity. So before when it was 15,000 tokens, now it's 30,000 tokens. But as you can see, the intelligence is there. So they're doing some great great work and open weight and MIT…”
See the full analysis with timestamps →