GPT-7: OpenAI’s Next AI Model Is Bigger Than We Thought

Vídeo em Inglês, publicado pelo canal AI Master.

About this content: The summary, key points and topics below are generated automatically from the video transcript. They reflect the content, opinions, claims and predictions presented in the video and do not represent a position, validation or fact-check by IA Dots.

Video summary

O vídeo analisa o lançamento do modelo GPT-6 Astra pela OpenAI, comparando-o com concorrentes recentes de outras empresas de inteligência artificial e discutindo os rumores sobre futuros modelos, vazamentos e a realidade dos benchmarks da indústria.

Key points

• A OpenAI lançou o Astra como GPT-6, com janela de contexto de cerca de um milhão de tokens, corte de conhecimento em abril e arquitetura de raciocínio de profundidade recorrente.
• Avaliações independentes mostram que o Astra fica atrás do Claude Fable 5.1 em índices de inteligência, embora apresente melhorias na redução de alucinações com o esforço de raciocínio máximo.
• Quatro laboratórios de IA lançaram modelos principais na mesma semana, incluindo Anthropic, Meta, OpenAI e Google.
• Rumores sobre futuros modelos da OpenAI, como os codinomes 'Doug' e 'Bell', baseiam-se em publicações de fontes não confirmadas ou memorandos de analistas, sem documentação pública oficial.
• O canal apresenta sua própria plataforma, chamada AI Master, que permite utilizar múltiplos modelos de IA lado a lado e gerenciar fluxos de trabalho.

Topics

GPT-6 AstraOpenAIModelos de Inteligência ArtificialBenchmarks e AvaliaçõesVazamentos e RumoresJanela de ContextoAGI EraFerramentas de IA

Video transcript

Transcript from available captions; it may contain errors.

A guy named Zahi had GPT6 Astra draw his portrait inside Canva. It used nothing but its own computer use controls pixel by pixel with the mouse. The clip showed it picking the drawing back up on its own after the browser crashed. That's by the account's own telling. Hundreds of thousands of people watched it. And the model Open AI reportedly finished training after this one doesn't have a name you've seen yet. Greg Brockman told reporters, "It's not unreasonable to feel that we are now in the AGI era." 48 hours later, three other labs shipped flagships of their own. That made his sentence look a lot less lonely and a lot less special. So, here's how today works. Every claim about the next Open AI model gets one of four badges on screen. The claim is either confirmed, a credible leak, speculation, or debunked. I'm not here to repeat leaks. I'm here to rate them. Start with the record because everything after it gets messier. Openi announced Astra on September 3rd under the API label GPT6 Astra which settles the biggest question immediately. Astra is GPT6, not a research system sitting upstream of the product you'd eventually buy. Access rolled out in three stages. The preparedness write up came first. The announcement and the system card landed 2 days later. Pay Chad GBT tiers and co-pilot followed the day after that. If your feed told you this happened in one clean launch day, it's squashed 3 days into one. The context window sits at about a million tokens and it holds up under an actual retrieval test. The eight needle version of MRCR runs past half a million tokens of input. It still finds the right buried facts around 96% of the time. Knowledge cut off sits at the end of April. Text and images go in and only text comes back out. Hold on to that last part. It answers the omni question later and it answers it badly. The reasoning architecture is new too. Reporting calls it recurrent depth and open AI has never confirmed that term. There's a trade-off. Part of the visible chain of thought now goes dark because the thinking happens inside loops you can read. Astra is also the first model Open AI has ever rated critical for cyber security. Exploit chaining stays gated to trusted partners instead of anyone with a credit card. On exploit bench, open AI reports 100%. That number came from open AI, not from anybody independent. Journalists got shown the computer used demos live. It laid out a circuit board in Kyad and built a city inside Unity. Then it drafted a tax return from a photographed W2. Pricing lands at $10 per million tokens in and 50 out that holds under about 270,000 tokens of context. Claude Fable 5.1 sits on the same shelf. Astro costs roughly two and a half times what Saul did. Training ran on more than a h 100,000 GPUs at Stargate in Abalene, Texas. Everything I just described costs money in a specific way. You pay $10 per million tokens in and 50 out. That buys you one model from one vendor. And Estra is one of four flagships that shipped in a single week. Say you want to work across all of them instead of picking a team. You end up paying four subscriptions and rebuilding your workflow every time one of them moves. That is the exact problem I built AI master to kill. It's my own platform. So take the enthusiasm with that in mind. Nobody paid me for this one. I am just showing you the thing I use daily. Here's the part that's relevant to today. Inside one window, you pick your model from Claude, Chad, GBT, Grock, and Gemini. You run them side by side on the same prompt per token. Cheaper than going direct. That's the comparison I ran to build this episode and you're watching it on screen right now. Image, audio, and video generation live in the same place. Also cheaper than stacking separate tools. Characters stay consistent across generations. The same face survives from one clip to the next instead of mutating. And a character you build can be published on the platform and earn there. Then there's variable, which is the piece I'm proudest of. You and your team run a whole channel inside the platform. AI agents write the scripts and descriptions. They edit the video and images and build the thumbnails. Generated content can be shared straight out for reach instead of dying in your downloads folder. The academy sits alongside it with 200 plus lessons around 30 hours. More than 14,000 people are in the community now. The results page is full of real screenshots and video testimonials from them, not stock quotes. Go to the link in the description and hit buy. Select the annual plan. Fill in your details and you'll get a confirmation email. Then log in and you're inside the same workspace I just showed you. The whole setup takes under 3 minutes. Link is below. Now, the part the press release rewrites skipped. Artificial Analysis runs its own intelligence index and on version 4.1.1, Astra came in at 61.2. Claude Fable 5.1 sat at 65.7. Opus 5 landed at 63.1 and Fable 5 came in just above 62. Saul scored a shade under 61. That scoreboard did not even survive the week. AA revised the index within days as more evaluations came in. called this the world's most intelligent model. It sits fourth on the most widely watched independent index. Both things are true at once. Astra wins the evals. Open AI chose to publish and loses the one it did not run. Astra's spot in the field did not improve under the update. Open AI have two more receipts. Epic AI disclosed that Open AI funded Frontier Math and holds exclusive access to part of it. That doesn't make the 98% fake. It means the referee is on the payroll, so treat it as a company number. Open AI also noted that its cla comparison runs used modified eval settings. That's a polite way of saying the other guy's model wasn't configured the way its own team would. Nobody's benchmark table reads in isolation this week. And Astra regressed in several places against solid drops roughly 80 LO on GDP val. It also gives back two to three points on the bank and agent task and on site code. Long context reasoning takes a similar dip. It uses about 10% fewer output tokens. That sounds like a saving until you remember it costs 2 and 1/2 times more per token. The net effect on an average index task is something like 75% more expensive than salt. So it is better at the hard things and worse at several boring ones. It costs more at all of them. That is a side grade with a great launch video. One number does genuinely impress me. It's the honest one to end this section on. On the omniscience hallucination test, Estra's rate reportedly falls from around 92% to about 51. [music] That's with reasoning effort cranked to max and accuracy goes up about four points. Having how often a Frontier model confidently makes things up matters to my actual work. It's worth more than any saturated math benchmark on that slide. It's also the only place in this launch where the new architecture visibly earns its keep. Astro did not land in an empty room. On September 1st, Anthropic shipped Claude Fable 5.1 and Mythos 5.1. The day after, Meta pushed Muse Spark 1.3. Then Astra arrived on the 3rd. Google had already moved a week earlier with Gemini Omni 1.1 Flash on August 27th. Four frontier labs shipped flagships in one single week. Every one of them chased the same target. Long horizon agentic work with a million token window. Four companies did not independently land on four ideas. They read the same research and race to the same finish line. Watch what happens to the numbers when you line them up. Meta reports 75.4% on the deep SFE coding benchmark at max settings. Astra reports 74.1 and Opus five clusters right around the same mark. Three flagships sit inside a point and a half of each other on the benchmark developers care about most. Statistically, that comes out as a tie. Rhetorically, it's three separate blog posts claiming the lead. The benchmark era is basically over as a way to pick a winner. And Anthropic did the thing nobody's spec sheet captured. That same week, Claude autonomously formalized a proof of Fermont's last theorem, working 11 days straight. There is no leaderboard row for that. It's the most impressive thing any model did that week, and it barely trended. Everyone was busy arguing over who scored 65. Meanwhile, the openweight side got quietly scary. GLM 5.3 Flash charges around 7 cents per million tokens in that runs roughly 140 times cheaper than Astra. If your workload is high volume and moderate difficulty, the Frontier stopped being the right answer a while ago. Here's the actual reason you clicked. Everything from this point on gets a badge. And I'm starting at the bottom rung. In July, Semi analysis published a memo describing a base model inside Open AI code name Doug. The memo calls it the company's first major base model overhaul since GPT40 with a launch window somewhere around November. That memo does not say the thing most aggregators said it said. Doug is described as a separate larger pre-train running alongside Astra. It is not the base sitting underneath it and it has not shipped. Then on August 25th, an account called synth waft posted that Open AI had finished a new pre-train code named Bell. The post puts it north of 10 trillion total parameters in the size class of GBT 4.5. It positions Bell as Doug's successor. Wall Street Engine reshared that post soon after. WCCF Tech, Win Central, and 36 Carrier picked it up from there. Now, watch how that chain actually works because this is the part every aggregator skips. Every one of those outlets is citing the same single post. Not one of them saw an open AI document, a system card, or a leaked internal deck. Four headlines from one tweet is not four sources. It's one source with better distribution. And I'll go further because I owe you this. We commissioned a fresh research pass for this episode, 19 sources deep. It covered OpenAI's own safety overview and several technical trackers. The conclusion it landed on was blunt. It found no publicly verifiable documents, system cards, or product announcements from OpenAI about any model beyond Astra. Nothing in there confirmed a code name, a capability, or a release schedule. That pass didn't contradict Bell & Doug. It simply never found him anywhere. So, the badge on Bell is credible leak. I want to be precise about why it still earns that word. The account has a track record. The details are boring and specific the way real leaks usually are. Nothing in the public record contradicts them, but a single origin claim stays single origin no matter how many outlets repeat it. If synth weight is wrong, everybody downstream is wrong at once. You'd never know that from the headlines. Doug sits one notch higher. Semi analysis is a named analyst shop putting its reputation on a dated memo rather than an anonymous account. A named source still is not a confirmed one. Nobody outside the company has seen either model. Here's one nuance that almost every video flattened. Bell is reportedly not GPT7 is described as a base model that comes after GPT6. It's a possible foundation for whatever crosses the threshold, not the product with a number on the box. Those two claims land in completely different places. Saying Open AI finished pre-training a very large base model describes a normal Tuesday at a frontier lab. Saying GBT7 exists and is finished is a different sentence with different consequences. And that second one isn't supported by anything I could find. Then there's the debunk pile and it's tall. It includes exact parameter counts for GPT7 and leaked benchmark tables showing it beating Gemini and Claude. It includes screenshots captioned, "I got AB tested into the new model in my app." Every one of those earns a debunked or speculation badge. Here's the test I keep coming back to on this channel. Ask what access a person would need for their claim to be true. A parameter count means training infrastructure access and a benchmark table means Eval harness access. A small number of named people hold that access. None of them are posting spec tables to X for engagement. If the claim needs access, the leaker demonstrabably doesn't have, then the claim is the product. We also have a live precedent from earlier this year and it wasn't about GBT7 at all. In March and April, a rumor wave built around an open AI model named Sput. The internet had already priced it in as GBT 5.7. It shipped as GBT 5.5, a point release, and the fresh research reinforced that pattern from a completely different angle. Prediction markets repeatedly mispriced Astra's own release window, only drifting toward the right dates very late in the process. A model that genuinely existed and genuinely shipped still had its timing wrong in public for weeks. So, think about what that does to the forecast on a model nobody has confirmed exists. Time to grade myself since this channel does that. In our GPT6 episode, I called Astra a research system sitting upstream of the product. I said the GPT6 you'd eventually use would be Astra's little brother. That call was wrong. Astra shipped as GPT6 under its own name. On top of that, it landed with a softer independent score than I expected. I got the shape of the release wrong, so I'm not going to quietly reuse that logic on the next one. Which is exactly why the naming question matters more than it sounds. The account tier has called the next release GPT 6.5, not GPT7. That's a real claim with a real consequence. Say OpenAI treats the next drop as a step inside generation 6. Then everybody arguing about GPT7's benchmarks is arguing about a product that won't carry that name for another year. And the company's recent history backs the incremental read. GPT 5.6 Saul arrived on July 9th. Astra landed September 3rd. That's roughly 2 months between flagships. It's the only forward-looking number in this video I'd actually defend. I got it by doing arithmetic on two products that already shipped. Sam Alman is on the record about pace and it's the opposite of what the hype cycle wants. He said Astra had been done training for a while. He called it a significant step forward in capabilities and alignment. Then he said something quieter about the models after it. They have been slowing things down as needed so the safety and alignment work can actually get done. Read that second half again. The CEO is telling you the break is being applied on purpose. He has separately said he expects an internal system by the end of this year. He personally called that one AGI. He conceded it isn't there yet. That system stays internal and internal is not shipped. That gap is the whole story of this cycle. It's why the model you can buy feels like a sidegrade while the press releases feel like a revolution. Now the loop I opened at the top is the next Openi flagship. A true omni model. Taking text, image, audio, and video in one pass. Start with what's shipped because that's the only hard evidence anybody has. Astra [music] takes text and images in and it only sends text back out. No audio goes in and none comes out. It handles no video in either direction and it offers no fine-tuning at launch. That's confirmed and it kills roughly half the hype on its own. The model that got called the arrival of the AGI era cannot hear you. Google is straightforwardly ahead on modality breadth right now and it isn't close. Gemini Omni 1.1 Flash accepts text, image, audio, and video as input. Multimodel input runs a $1.50 per million against Asteris 10. Google is also the only one of these companies that put the word Omni in the product name. It's a small thing and it tells you where they think the fight is. So, if Omni is what you need this month, the answer isn't a leak about Open AI, it's a Google product you can call today. And here's the signal I think is the most underrated thing in this entire cycle. 36 Kay reported that Open AI shut down Sora and the Atlas browser this year. That compute went to the core model in the agent work, the same reporting sites an internal scaling alert inside the company. That's one outlet, not an open AI statement. So, the badge here is credible leak, but sit with the shape of it. A company that shelves its own video product for compute reasons is not two weeks away from shipping natively video omni. You don't shut the video team down for GPU reasons and then hand video generation to your flagship. The resource fight tells you the road map more honestly than the keynote does. So, when someone shows you an omni spec table, ignore it and run free tests instead. Interrupt the model's voice mid-sentence and see whether it holds the context or restarts. Ask it to edit an image from 10 turns ago without reexplaining what that image was. Feed it live video, not an uploaded choose on afterward. Those three answers tell you what you actually got. No leaked table tells you any of them. That's why I keep declining to read them out on camera. Strip the AGI language out and the bet reads plainly. OpenI is not buying a leaderboard column this year. It is buying compute and computer use and it's buying the frame every competitor now has to argue inside. Astra sits below Fable 5.1 on the independent index and Open AI's position barely moves. That index measures answers. It does not measure grinding through an hour of cursor work inside Canva and recovering after the browser dies. The demos [music] Open AI walked journalists through all point the same direction. They laid out a board in Keycad and drafted a tax return of a W2. Those are real jobs, not clever prompts staged for a demo reel. The compute half is the part most coverage skips. 100,000 plus GPUs at Stargate is the training run Open AI will put its name on. Guang claims another 400,000 come online next. Nobody at Open AI has said that part out loud. Somebody is building for a model much bigger than the one you can buy today. which lands us back on Belle. If Belle is real, Astra is the story and Belle is the balance sheet. The honest version of that sentence keeps the if sitting right in the middle of it. One account said it and two research passes have found nothing underneath it. So, what I actually do on Monday, and none of it involves waiting for GPT7. The rule I use is simple. I route by the job in front of me instead of buy whichever brand I like this month. Long context retrieval across a huge repo or pile of case files goes to Astra. Its eight needle retrieval scores past half a million tokens are the strongest receipt on the whole card. Reasoning heavy code review and long writing goes to Fable 5.1 or Opus 5. Both sit above Astra on the independent index at a comparable list price. Anything with audio or video going in goes to Gemini Omni 1.1 Flash. Astra will not take it at all. Second, watch one signal and ignore everything else. The tell before Astro was paperwork. Open AAI published the path to Astra document on September 1st, 2 days ahead of the announcement. It named the critical threshold before any model card existed. Safety documentation ships before the model does. So, the next real OpenAI flagship turns up in a preparedness framework post in a new API label. It does not turn up in a spec table screenshotted onto your timeline. If you see the table first, you are looking at fiction. Third, stop rebuilding your stack every 6 weeks. 2 months separated Saul from Astra. If that cadence holds, you'll be doing this again before the year ends. The thing worth building is not a prompt library tuned to one model. It is your own eval. You pull 20 tasks out of your actual work with answers you already trust. Then you save them somewhere you can rerun in an afternoon. Then a launch stops being news you have to interpret. It becomes a test you run before launch. So here's the verdict I'd actually defend. Astra is the best software operator anybody has shipped and a thin intelligence upgrade at the same time. I can't defend paying roughly 75% more per task than Fable 5.1 while it scores under it. That workflow I ran earlier lifts on my own platform AI Master. And if you've got a product worth showing this audience, the sponsorship links in the description below. Always happy to talk.

Assistir ao vídeo original no YouTube

Assuntos: AGI. Publicado no YouTube em 2026-09-14.

Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.