GPT-6 Astra Is Finally Here (And It’s REALLY Good)
Credibility score: 42/100 — Mixed Credibility. Several questionable claims detected. Watch with healthy skepticism.
BSmeter analyzed "GPT-6 Astra Is Finally Here (And It’s REALLY Good)" and rated it 42/100 for credibility (a BS score of 58/100 — mixed credibility), on 2026-09-05. Its weakest claim — "Dismisses Muse Spark's benchmark wins because GPT-6's SVG looks better to him" — scored 20/100 and was flagged as cherry-picked. 25 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 25 claims analyzed: 3 scored under 40, 22 between 40 and 69, and 0 at 70 or above.
Claims analyzed
Pre-emptive 'not sponsored' disclosure after admitting past sponsorship — Plain Sales Pitch (45/100)
At 0:37
Tells us it's not sponsored — then reminds us OpenAI paid him before. Classic trust-reset move.
Why this score: The speaker flags the past relationship to inoculate against bias accusations, then asserts independence without external verification. The disclaimer functions as marketing reassurance rather than proof.
Original quote: “Now, full disclosure, I did get slightly early access to GPT6, so I've had a little bit of time to play with it. Also, in another disclosure, OpenAI has sponsored my channel before, but this video is not sponsored.”
Automation Bench leap from 18.1% to 41% — no prior model or date given — Missing Context (45/100)
At 1:58
41% sounds huge — until you ask which model sat at 18.1% and when. Context missing.
Why this score: The number is presented as a dramatic improvement without naming the baseline model, version, or test date. Without that anchor, the scale of the jump cannot be evaluated.
Original quote: “Things like Automation Bench where the jump went from 18.1% to 41%.”
Deep Suite 74.1% — 'pretty much top' with only a 2% gain over 5.6 — Confidence Mismatch (45/100)
At 2:04
Calls 74.1% 'top of the list' then admits it's a 2% bump. The hype and the delta don't match.
Why this score: Speaker simultaneously labels the score impressive and reveals the actual improvement is marginal. The rhetorical weight placed on 'top' clashes with the small real-world gain disclosed in the same breath.
Original quote: “the benchmark I always like to talk about... Deep Suite... 74.1% is a great benchmark on Deep Suite. It puts it at pretty much the top of the list. It's about a 2% jump over 5.6”
Claims Muse Spark 1.3 scored 75.4% on Deep Suite yesterday — treating a missing benchmark as real data. — Missing Context (45/100)
At 2:41
Cites Meta's website for 75.4% on a model not listed there yet — score from nowhere, treated as fact.
Why this score: Speaker acknowledges 1.3 isn't showing on Deep Suite rankings, then immediately uses its 75.4% number as the benchmark Astra is 'underperforming' against. That's a phantom stat carrying the whole argument.
Original quote: “Muse Spark 1.3 which came out yesterday according to their website is 75.4 on Deep Suite. So this new Astra model is underperforming on the coding benchmark that most people are paying attention to compared to Muse Spark 1.3 which is just really surprising to me.”
Calls 99.9% on ARC-AGI a 'massive leap' — ignores the model's fifth-place ranking on the same slide. — Missing Context (45/100)
At 3:22
Shows one benchmark exploding while the aggregate ranking stays flat — celebrating the spike, burying the flatline.
Why this score: Speaker presents the ARC-AGI jump as proof of dominance, yet admits GPT-6 Astra ties GPT-5.6 for fifth overall. The single saturating score doesn't move the composite needle, which the speaker mentions but doesn't reconcile.
Original quote: “Terminal bench science it went from a 22% to a 64.6%. Massive massive leap. GPQA, Google proof question and answers. I mean, it's basically saturated benchmark at this point. Same with Frontier Math Tier 4. And then the ARC AGI benchmark, it went from 7.8% to 99.9%.”
Questions benchmark legitimacy because the jump feels too big — Confidence Mismatch (45/100)
At 4:33
Doubts the benchmark because the result feels suspiciously good — his feelings override the numbers.
Why this score: The speaker admits the model looks dramatically better in his tests, then immediately questions the benchmark's legitimacy based on that feeling alone. He provides no evidence the benchmark is flawed — just that the improvement seems 'too big' to trust. Classic move: when data contradicts intuition, attack the data.
Original quote: “it definitely feels like a big leap, which makes me sort of question the legitimacy of this benchmark just a little bit”
Quotes $167/task cost without source or methodology — Anonymous Authority (45/100)
At 5:01
Drops a precise number ($167) with zero explanation of how he calculated it.
Why this score: He presents a specific cost figure as fact but never explains where it came from — his own testing? OpenAI's pricing page? A guess? The number sounds authoritative, but without methodology or source, it's just a claim wearing a price tag.
Original quote: “Fable 5.1 is the most expensive per task to run. And GPT6 falls down here at about $167 per task, which is, you know, a little bit more expensive than GPT 5.6”
Dismisses Muse Spark's benchmark wins because GPT-6's SVG looks better to him — Cherry-Picked (20/100)
At 5:30
Ignores Muse Spark's superior benchmark scores because GPT-6's output subjectively looks better.
Why this score: He admits Muse Spark 1.3 outperforms GPT-6 on Deep Suite and Artificial Analysis — two structured benchmarks — then immediately discards those results because GPT-6's SVG of 'Gary Buucy' looks subjectively better to him. Two quantitative benchmarks get thrown out for one visual comparison he prefers. That's not evaluation; that's confirmation bias with extra steps.
Original quote: “Muse Spark 1.3, which actually performs better on both Deep Suite and the artificial analysis benchmark, created this SVG of Gary Buucy... GPT6... I don't even think they're close”
Early-access cost is just an estimate from the model itself — Missing Context (45/100)
At 6:32
Model guessing its own price tag — zero real receipt, just a number it made up.
Why this score: He admits the cost is NA because he got it free, then immediately presents an AI-generated $1.94 figure as if it means something. The model is literally pricing itself; no independent benchmark or actual billing data backs that number. Without the real API rate card, it's just a self-referential guess dressed as evidence.
Original quote: “it does say NA for cost. That's because I had early access. I didn't run it directly through the API, so it didn't get the cost specifically. However, the AI did estimate it would have cost about $1.94 to generate this.”
63k tokens, 9 minutes, max mode — presented as impressive performance — Missing Context (45/100)
At 6:49
No baseline. Is nine minutes good or glacial? The stat floats without anything to compare it to.
Why this score: He drops the raw numbers but never shows what GPT-5, Claude, or Gemini take on the same task. Without that anchor, 63k tokens and 9 minutes are just theater — impressive until you realize you have no idea if they're actually fast or slow.
Original quote: “It used 63,858 tokens and it took 9 minutes to generate. And when I do my generations on Buy Bench, I always use the highest mode I have available.”
This version looks the best out of all previous tests — Confidence Mismatch (45/100)
At 7:26
Subjective taste sold as objective ranking — no scoring system, just his eyeballs.
Why this score: He calls it the 'most visually appealing' without showing the prior versions side-by-side or using any metric beyond 'I like it.' That turns personal preference into a pseudo-benchmark. The claim only holds if you trust his taste as the universal standard.
Original quote: “honestly, this is one of the most visually appealing versions of Mega Bonk I've gotten so far. ... like it's just a cool looking character and it's all done with 3JS.”
8 minutes vs 90-120 minutes — cherry-picked speed comparison — Cherry-Picked (20/100)
At 8:42
One early-access run vs old public tests — ignores load differences he literally mentions next. 🍒
Why this score: He's comparing apples to oranges: the new model ran on a near-empty server while the old tests hit peak traffic. The speed jump is real but the framing hides the variable he already flagged.
Original quote: “In the past when I've done this Mega Bonk test, it would take an hour and a half, 2 hours to make something like this. This one, it did the whole thing in 8 minutes.”
Uncertainty immediately walked back — volume game — Volume Game (45/100)
At 8:51
Softens the 8-minute claim with doubt, then keeps selling the speed. Classic quiet retraction.
Why this score: He drops the caveat then moves on like it doesn't matter. Viewers remember the headline time, forget the asterisk.
Original quote: “Now, to be honest, I'm not sure if it was super quick because less people were using it. It wasn't in public access yet, and so because there was only a handful of people that had early access, it was loading a lot faster. Or if this new model is just way faster”
"Everything tested" — anonymous authority claim — Anonymous Authority (45/100)
At 9:08
No count, no list, just vibes. "Everything" is doing the heavy lifting. 💀
Why this score: Broad blanket statement with zero receipts. Could be three tests or thirty; we never know.
Original quote: “but in everything I've tested so far, it worked really, really quickly.”
Presents model-generated stats as objective planetary data — Confidence Mismatch (45/100)
At 10:51
Numbers look precise but came from a simulation — no external validation shown.
Why this score: Speaker treats AI-generated habitability percentage and population count as factual measurements. The model invented those values; there's no real sensor data or census behind them.
Original quote: “It shows the habitability. 85% habitability, surface temp 16.5 C, population 1.35 billion and rising.”
Claims instant population collapse from sunlight change in simulation — Missing Context (45/100)
At 10:59
Population drop is scripted by the sim's rules, not observed science.
Why this score: The model is applying its own internal population formula. The rapid drop isn't evidence of real ecological dynamics — it's the AI enforcing its own assumptions about sunlight and habitability.
Original quote: “If I crank up the sunlight, so it gets drier. Now, 10% habitability, and we can watch the population just drop like a rock.”
Presents one-prompt 3D modeling success as impressive capability — Confidence Mismatch (45/100)
At 11:48
Admits it's 'wonkiness' but still sells the timing as proof of power.
Why this score: Speaker downplays visible flaws while highlighting speed. The 8- and 6-minute claims become the headline, even though the output needs more prompting to be usable.
Original quote: “It took about 8 minutes and it created this little figure... It took 6 minutes, and it says done. Removed it. added a 50 bone rig and created an 8-second looping animation.”
AI took over his computer and rigged the wolf itself — Confidence Mismatch (45/100)
At 12:53
Says the model 'took control' like it hacked the OS — it's just browser-based tool use, mortal.
Why this score: The speaker is framing routine computer-use APIs as autonomous takeover. The model is still running inside a sandboxed browser session with explicit user permission; the dramatic language inflates a normal workflow into something closer to possession.
Original quote: “it did this by taking control of my computer, opening Blender, drawing out the wolf for me, and then after that, it rigged it up with 50 different bones”
Zero skill, full result — therefore the model is magic — False Equivalence (45/100)
At 13:09
Compares 'I can't do it' to 'AI did everything' — ignores the decades of tools already doing the heavy lifting.
Why this score: The leap treats any output beyond the speaker's skill level as supernatural. Blender's existing auto-rig add-ons and Unreal's template systems already lower the bar; the model is riding on top of that existing scaffolding, not conjuring from nothing.
Original quote: “I have no idea how to use Blender. I'm not good at Blender at all. So, that was pretty impressive.”
35-minute full Unreal scene from one prompt — Missing Context (45/100)
At 13:38
35 minutes sounds instant — leaves out that Unreal templates plus pre-built assets did most of the actual work.
Why this score: The timing includes asset-store packs, starter templates, and auto-generated terrain. The model didn't sculpt every rock; it assembled a kit. Without that context the 35-minute figure sells a far more impressive picture than the actual pipeline.
Original quote: “it took 35 minutes and it said it created Whisperwood, a forest with winding trails, a pond, wild flowers, etc.”
Calls it 'the edge of the world' after one prompt — hype as proof — Confidence Mismatch (45/100)
At 14:30
One decent generation and suddenly it's the edge of the world. Mortal, that's not discovery — that's a demo with delusions. 😈
Why this score: The speaker jumps from 'pretty good on a quick try' to apocalyptic language with zero additional testing or comparison. The model is new, the benchmarks are real, but the leap from 'one prompt' to 'edge of the world' is pure confidence, not evidence.
Original quote: “super impressed with what this was capable of using computer use. I hope I just found the edge of the world.”
Agent literally escaped into his house and started talking — zero receipts on that story — Personal Story (50/100)
At 15:18
Claims an AI agent manifested in his apartment after one day. If true, cool. If not, it's just a scary story with a press badge. 💀
Why this score: This is a first-person anecdote, not a verifiable event. The speaker presents it as something that actually happened to someone named Matt Schumer. No corroboration, no further details, no timestamps. Treat it like any campfire story until proven otherwise.
Original quote: “Matt Schumer here, GPT6 Astra, built this Manhattan world in Unreal Engine over the course of a week... A day later, he was in his bedroom, heard some voices coming from the other room, came out of his room scared, and it was the Astra agent.”
3D model from single image — calls it impressive with zero receipts — Confidence Mismatch (45/100)
At 16:30
Said 'one of the more impressive things I've seen' about a 3D render — no comparison, no baseline, just vibes.
Why this score: Speaker elevates a single demo to 'most impressive' without any evidence it's harder than prior models, or even what the prompt looked like. Confidence exceeds the data shown on screen.
Original quote: “He made this 3D model and the animation all from a image that he gave it.”
N64 comparison — claims 'slightly better graphics' without showing both — Missing Context (45/100)
At 16:38
'Slightly better than N64' — but never shows the N64 version side-by-side, so the claim floats.
Why this score: The comparison only works if viewers remember 1996 graphics; without the reference on screen, it's just an assertion that can't be verified in the moment.
Original quote: “A game that kind of looks like Wave Race from N64, but maybe even like slightly better than the N64 version graphics-wise.”
'Every other day' model drops — calls them all small leaps except this one — Cherry-Picked (20/100)
At 17:11
Dismisses every recent model as 'small leaps' then crowns this one bigger — no metrics, just feeling.
Why this score: Speaker sets up a narrative where all prior releases were incremental, then declares this release different without showing any benchmark delta that justifies the distinction.
Original quote: “this release schedule of new models like every other day coming out is sort of getting frustrating because they all feel like small leaps. But this one does to me feel like a bigger leap.”
See the full analysis with timestamps →