graph LR
classDef core fill:#ffd4d4,font-weight:bold,font-size:14px;
classDef size fill:#d4ffd4,font-weight:bold,font-size:14px;
classDef bench fill:#d4d4ff,font-weight:bold,font-size:14px;
classDef api fill:#ffffd4,font-weight:bold,font-size:14px;
classDef price fill:#ffd4ff,font-weight:bold,font-size:14px;
classDef open fill:#d4ffff,font-weight:bold,font-size:14px;
classDef vibe fill:#fff0d4,font-weight:bold,font-size:14px;
classDef safe fill:#f0d4ff,font-weight:bold,font-size:14px;
classDef meta fill:#d4fff0,font-weight:bold,font-size:14px;
Main[Vault7-359]
Main --> C1[Speed
& reason merge. 1]
C1 -.-> G1[Core]
Main --> C2[Three sizes:
Full Mini Nano. 2]
C2 -.-> G2[Size]
Main --> C3[SOTA claims
on three benches. 3]
C3 -.-> G3[Bench]
Main --> C4[100 % scores
spark over-fit doubt. 4]
C4 -.-> G3
Main --> C5[Distilled from
O3 curriculum. 5]
C5 -.-> G1
Main --> C6[400 k window
4Ă— length. 6]
C6 -.-> G1
Main --> C7[API adds
reasoning slider. 7]
C7 -.-> G4[API]
Main --> C8[Regex grammar
output constraints. 8]
C8 -.-> G4
Main --> C9[Tool-call
preambles explain. 9]
C9 -.-> G4
Main --> C10[Verbosity:
low med high. 10]
C10 -.-> G4
Main --> C11[Free drops
to Mini. 11]
C11 -.-> G5[Price]
Main --> C12[Pro keeps
unlimited plus. 12]
C12 -.-> G5
Main --> C13[Gmail Calendar
memory Pro. 13]
C13 -.-> G1
Main --> C14[Voice lasts
hours natural. 14]
C14 -.-> G1
Main --> C15[Themes &
sarcasm toggles. 15]
C15 -.-> G6[Vibe]
Main --> C16[Safe path
after refusal. 16]
C16 -.-> G7[Safe]
Main --> C17[Hallucination
drop vs O3. 17]
C17 -.-> G7
Main --> C18[Pricing $1.25
$10 per M. 18]
C18 -.-> G5
Main --> C19[Nano 1/25
cost fast. 19]
C19 -.-> G2
Main --> C20[OSS 20 B
rivals GPT-4. 20]
C20 -.-> G8[Open]
Main --> C21[270 tok/s
on RTX 5090. 21]
G8 --> C21
Main --> C22[Apache 2.0
bans retrain. 22]
G8 --> C22
Main --> C23[PhD pocket
slogan mocked. 23]
C23 -.-> G9[Meta]
Main --> C24[Cancer tale
praised panned. 24]
C24 -.-> G9
Main --> C25[Spanish panel
calls infomercial. 25]
C25 -.-> G9
Main --> C26[100 % scores
useless per experts. 26]
C26 -.-> G3
Main --> C27[Google Gemini
seen winner. 27]
C27 -.-> G9
Main --> C28[Claude 4.1
gains ground. 28]
C28 -.-> G9
Main --> C29[No jaw-drop
multimodal demo. 29]
C29 -.-> G9
Main --> C30[No AGI
superintelligence said. 30]
C30 -.-> G9
Main --> C31[August roll-out
not months. 31]
C31 -.-> G1
Main --> C32[Memory learns
user schedules. 32]
C32 -.-> G1
Main --> C33[Canvas auto-SVG
interactive demos. 33]
C33 -.-> G4
Main --> C34[Front-end quality
70 % vs O3. 34]
C34 -.-> G1
Main --> C35[Agent coding
45 min alone. 35]
C35 -.-> G1
Main --> C36[Cursor adopts
GPT-5 default. 36]
C36 -.-> G4
Main --> C37[Cursor trial
days free. 37]
C37 -.-> G4
Main --> C38[PDF bug
fixed live. 38]
C38 -.-> G4
Main --> C39[Voice quizzes
Korean café. 39]
C39 -.-> G6
Main --> C40[Nano hinted
pre-launch. 40]
C40 -.-> G2
Main --> C41[Mini-Nano
Apple memes. 41]
C41 -.-> G9
Main --> C42[Streamed on
YT LinkedIn Rumble. 42]
C42 -.-> G9
Main --> C43[Reject Meta
for rebeldĂa. 43]
C43 -.-> G9
Main --> C44[170 Spanish
AI episodes. 44]
C44 -.-> G9
Main --> C45[Discord shares
books models. 45]
C45 -.-> G9
Main --> C46[Want Ilya
Demis Elon. 46]
C46 -.-> G9
Main --> C47[US AI lead
at risk. 47]
C47 -.-> G9
Main --> C48[China mocks
GPT-petardo-5. 48]
C48 -.-> G9
Main --> C49[Elon tweets
Grok 4 wins. 49]
C49 -.-> G9
Main --> C50[Brand fatigue
like Bard. 50]
C50 -.-> G9
Main --> C51[Weak Spanish
speakers slammed. 51]
C51 -.-> G9
Main --> C52[Want visionary
not emotional. 52]
C52 -.-> G9
Main --> C53[Synthetic loop
self-improve hint. 53]
C53 -.-> G1
Main --> C54[Safe reduce
blunt refusals. 54]
C54 -.-> G7
Main --> C55[HealthBench by
250 doctors. 55]
C55 -.-> G3
Main --> C56[Health scores
top ever. 56]
G3 --> C56
Main --> C57[Amgen tests
drug design. 57]
C57 -.-> G3
Main --> C58[BBVA cuts
weeks to hours. 58]
C58 -.-> G3
Main --> C59[Org-wide
rate limits. 59]
C59 -.-> G5
Main --> C60[EDU tier
next week. 60]
C60 -.-> G5
Main --> C61[Custom GPTs
get voice. 61]
C61 -.-> G4
Main --> C62[Auto packing
lists from Gmail. 62]
C62 -.-> G1
Main --> C63[Fixes own
lint errors. 63]
C63 -.-> G1
Main --> C64[Chooses React
Tailwind auto. 64]
C64 -.-> G1
Main --> C65[Purple theme
running joke. 65]
C65 -.-> G6
Main --> C66[Want tokenizer
open-source. 66]
C66 -.-> G9
Main --> C67[August date
damage control. 67]
C67 -.-> G9
Main --> C68[4.5 failure
still fresh. 68]
C68 -.-> G9
Main --> C69[Expect leap
in GPT-6. 69]
C69 -.-> G9
Main --> C70[Undercuts Claude
monthly fee. 70]
C70 -.-> G5
Main --> C71[Nano targets
edge IoT. 71]
C71 -.-> G2
Main --> C72[Mini fits
mid GPU. 72]
C72 -.-> G2
Main --> C73[Full wants
H200 speed. 73]
C73 -.-> G2
Main --> C74[Mac RAM praised
bandwidth capped. 74]
C74 -.-> G8
Main --> C75[>100 % dubbed
over-fit dash. 75]
C75 -.-> G3
Main --> C76[No Sora
video demo. 76]
C76 -.-> G9
Main --> C77[No computer-use
shown. 77]
C77 -.-> G9
Main --> C78[Google silent
Gemini soon. 78]
C78 -.-> G9
Main --> C79[Reg urgency
pulled Spanish bill. 79]
C79 -.-> G9
Main --> C80[Brand equity
depleted post-event. 80]
C80 -.-> G9
Main --> C81[Model good
launch disaster. 81]
C81 -.-> G9
G1[Core] --> C1
G1 --> C5
G1 --> C6
G1 --> C13
G1 --> C31
G1 --> C32
G1 --> C34
G1 --> C35
G1 --> C53
G1 --> C62
G1 --> C63
G1 --> C64
G2[Size] --> C2
G2 --> C19
G2 --> C40
G2 --> C71
G2 --> C72
G2 --> C73
G3[Bench] --> C3
G3 --> C4
G3 --> C26
G3 --> C55
G3 --> C56
G3 --> C57
G3 --> C58
G3 --> C75
G4[API] --> C7
G4 --> C8
G4 --> C9
G4 --> C10
G4 --> C33
G4 --> C36
G4 --> C37
G4 --> C38
G4 --> C61
G5[Price] --> C11
G5 --> C12
G5 --> C18
G5 --> C59
G5 --> C60
G5 --> C70
G6[Vibe] --> C15
G6 --> C39
G6 --> C65
G7[Safe] --> C16
G7 --> C17
G7 --> C54
G8[Open] --> C20
G8 --> C21
G8 --> C22
G8 --> C74
G9[Meta] --> C23
G9 --> C24
G9 --> C25
G9 --> C27
G9 --> C28
G9 --> C29
G9 --> C30
G9 --> C41
G9 --> C42
G9 --> C43
G9 --> C44
G9 --> C45
G9 --> C46
G9 --> C47
G9 --> C48
G9 --> C49
G9 --> C50
G9 --> C51
G9 --> C52
G9 --> C66
G9 --> C67
G9 --> C68
G9 --> C69
G9 --> C76
G9 --> C77
G9 --> C78
G9 --> C79
G9 --> C80
G9 --> C81
class C1,C5,C6,C13,C31,C32,C34,C35,C53,C62,C63,C64 core
class C2,C19,C40,C71,C72,C73 size
class C3,C4,C26,C55,C56,C57,C58,C75 bench
class C7,C8,C9,C10,C33,C36,C37,C38,C61 api
class C11,C12,C18,C59,C60,C70 price
class C15,C39,C65 vibe
class C16,C17,C54 safe
class C20,C21,C22,C74 open
class C23,C24,C25,C27,C28,C29,C30,C41,C42,C43,C44,C45,C46,C47,C48,C49,C50,C51,C52,C66,C67,C68,C69,C76,C77,C78,C79,C80,C81 meta
Resume:
OpenAI’s launch event for GPT-5 was framed as a milestone on the road to AGI, yet the two-hour broadcast felt more like a carefully scripted infomercial than the scientific watershed the community expected. Executives repeated that the model now “thinks like a PhD,” but the demos rarely strayed from polished party-planning, SVG animations and French-vocabulary games. While benchmark slides claimed state-of-the-art scores on SWE-Bench, MMMU and HealthBench, critics noted that several metrics topped out at 100 %, a statistical red-flag for over-fitting. The presentation trio—GPT-5, GPT-5 Mini and GPT-5 Nano—was introduced with tiered pricing ($1.25 / $10 per million tokens for the full model, down to Nano at a 25× discount), but no radical architectural reveal accompanied the numbers. Instead, the stream lingered on emotional testimonials, including a cancer patient who described using ChatGPT to parse biopsy jargon, a segment many viewers praised as humane while others condemned it as exploitative sentimentality.
Under the hood, OpenAI confirmed that GPT-5 unifies the previously split paths of fast “GPT” and slow “reasoning” models: a single network now decides how long to deliberate before answering. Distillation from O3-generated synthetic data was highlighted as the secret sauce, allowing a 490-billion-parameter teacher to compress into the 70-billion-parameter Mini without heavy loss. API users gain a new “minimal” reasoning-effort flag, 400 k-token context windows, custom regex-constrained outputs and tool-call preambles, all shipping today. Enterprise and EDU tiers arrive next week, while free-tier users cycle into GPT-5 Mini after a yet-unspecified quota. Safety was addressed with “safe completions”: the model may refuse risky steps but must explain why and suggest compliant alternatives, a policy designed to curb both hallucinations and unhelpful blanket refusals. Voice mode, memory, Gmail/Calendar integration and color themes were also upgraded, reinforcing OpenAI’s shift toward consumer stickiness rather than pure research spectacle.
Reaction in the Spanish-speaking AI community was scathing. Panelists who had stayed online past 2 a.m. to cover the event called it “una hora y pico de anuncios sin parar” that buried genuine technical gains under marketing clichés. They praised the open-source GPT-OSS 20-billion model for rivaling GPT-4 on consumer GPUs at 270 tokens/s, and warned that GPT-5’s modest delta could cede narrative ground to Google’s Gemini 2.5 Pro or Anthropic’s Claude 4. The consensus: OpenAI squandered cultural capital by replacing rock-star researchers with inexperienced presenters, over-promising AGI and under-delivering wow-factor demos. If the benchmarks hold, GPT-5 will still be the best coding and health-consultation model available, but the launch spectacle may have damaged the brand more than it helped, handing rivals a rare opportunity to reclaim the spotlight.
Key Ideas:
1.- GPT-5 merges speedy and reasoning models into one.
2.- Three sizes: full, Mini (70 B), Nano (cheapest).
3.- Claims SOTA on SWE-Bench, MMMU, HealthBench.
4.- Some evals hit 100 %, raising over-fit suspicions.
5.- Distilled from O3 synthetic curriculum, not new arch.
6.- 400 k context window, 4Ă— prior length.
7.- API adds “minimal” reasoning-effort slider.
8.- Custom regex/grammar output constraints shipped.
9.- Tool-call preambles explain upcoming actions.
10.- Verbosity parameter: low, medium, high.
11.- Free users downgrade to GPT-5 Mini after quota.
12.- Pro tier keeps unlimited GPT-5 plus extended think.
13.- Gmail & Calendar access for memory, starting Pro.
14.- Voice mode now hours-long, more natural prosody.
15.- Color themes and sarcastic personality toggles.
16.- Safe completions: explain refusal, offer safe path.
17.- Hallucination rate said to drop vs O3/O4.
18.- Pricing: $1.25 in / $10 out per million tokens.
19.- Nano costs 1/25 of full model, faster.
20.- GPT-OSS 20 B open-weights rivals GPT-4 locally.
21.- OSS runs 270 tokens/s on RTX 5090.
22.- OSS license Apache 2.0 but bans retraining.
23.- Community mocks “PhD-in-your-pocket” slogan.
24.- Cancer-patient story praised and criticized.
25.- Spanish panel calls event “endless infomercial.”
26.- Benchmark 100 % scores labeled useless by experts.
27.- Google Gemini 2.5 Pro seen as big winner.
28.- Claude 4.1 Opus also gains ground perception.
29.- Event lacked multimodal jaw-dropping demo.
30.- No AGI or superintelligence announced.
31.- Roll-out promised through August, not months.
32.- Memory upgrade learns user schedules.
33.- Canvas tool auto-creates interactive SVG demos.
34.- Front-end code quality preferred 70 % vs O3.
35.- Agentic coding chains run 45 min unattended.
36.- Cursor IDE adopts GPT-5 as default today.
37.- Cursor users get free trial days.
38.- SDK PDF-upload bug fixed live on stage.
39.- Voice study mode quizzes Korean café phrases.
40.- GPT-5 Nano hinted at on OpenAI site pre-launch.
41.- Community memes “GPT-5 Mini-Nano” Apple-style.
42.- Event streamed on YouTube, LinkedIn, Rumble, Twitch.
43.- Hosts reject Meta platforms for rebeldĂa.
44.- 170 prior Spanish AI community episodes cited.
45.- Discord group shares free books, models, tools.
46.- Panelists want Ilya, Demis, Elon on stage.
47.- U.S. geopolitical AI leadership said at risk.
48.- China humor shows mock “GPT-petardo-5.”
49.- Elon tweets Grok 4 Heavy beats GPT-5.
50.- OpenAI risks brand fatigue like Google Bard.
51.- Presentation speakers criticized for weak Spanish.
52.- Cultural call for visionary, not emotional, demos.
53.- Synthetic data loop foreshadows self-improvement.
54.- Safe completions reduce blunt refusals.
55.- HealthBench built with 250 physicians.
56.- GPT-5 scores higher than any prior model on health.
57.- Amgen tests GPT-5 for drug-design reasoning.
58.- BBVA cuts financial-analysis time 3 weeks→hours.
59.- Enterprise rate limits support whole org usage.
60.- EDU tier launches next week.
61.- Custom GPTs now support voice.
62.- Memory can auto-build packing lists from Gmail.
63.- GPT-5 fixes its own lint errors during build.
64.- Model chooses React/Tailwind without prompting.
65.- Purple color theme becomes running joke.
66.- Panel wants tokenizer open-sourced next.
67.- August rollout date seen as damage control.
68.- GPT-4.5 failure still fresh in minds.
69.- Community expects architecture leap in GPT-6.
70.- Pricing undercuts Claude Pro monthly fee.
71.- GPT-5 Nano targets edge IoT deployments.
72.- Mini suffices for mid-range GPU rigs.
73.- Full model needs high-memory H200 for speed.
74.- Mac unified memory praised but bandwidth-limited.
75.- Over-100 % scores dubbed “over-fit dashboard.”
76.- Event lacked live Sora or video-generation demo.
77.- No computer-use capabilities shown.
78.- Panel predicts Google silent Gemini drop soon.
79.- U.S. regulation urgency withdrawn from Spanish bill.
80.- OpenAI brand equity seen as depleted post-event.
81.- Community consensus: model good, launch disaster.
Interviews by Plácido Doménech Espà & Guests - Knowledge Vault built byDavid Vivancos 2025