← Blog

The soap opera of the next LLMs: who promises, who stalls, and who already launched (again)

27 ago 2026

⏱️ Reading time: ~4 minutes

Introduction

If you think your prime-time soap opera has plot twists, hold on. The artificial intelligence market in 2026 has managed to assemble a storyline more convoluted than any telenovela: models are announced before they exist, release dates move like a finish line nobody wants to reach, and at least one "launch" seems to have been dreamed up by a website that got a little too excited. Welcome to the latest chapter of the great LLM soap opera — where the climax never arrives, but the teaser announcement shows up every week.

GPT-6 (or is it "Astra"? Not even OpenAI is sure)

OpenAI is a master at keeping the audience on the edge of their seats. After months of rumors about GPT-6, the company launched the GPT-5.6 family in July — made up of Sol, Terra, and Luna, names that sound more like zodiac signs than language models. In August came a new chapter: the company publicly called its Astra project its "next big model" and published research done with an internal version of it — without, however, confirming whether Astra is the fabled GPT-6 or an entirely new name. In other words: we still have no date, no official specs, and no idea even what the next lead character will be called.

Gemini 4: Google is already "training" it, they swear

On July 21, Google confirmed that it had begun what it described as its most ambitious pre-training run to date, for Gemini 4. Days later, on Alphabet's earnings call, CEO Sundar Pichai reinforced the message, calling the next model "significantly larger" than any previous Gemini. But there's an awkward detail in this plot: Gemini 3.5 Pro, promised for June, simply never showed up. That's why prediction markets like Manifold put the odds of Gemini 4 arriving before November at around 74% — which, translated from "gambler-speak," means: "probably, maybe, who knows."

Grok 5: the eternal "any minute now"

If there were an award for "most delayed model of the year," Grok 5 would win in a landslide. Elon Musk promised a launch in the first quarter of 2026. The deadline passed. The official xAI account then pointed to the second quarter. It passed again. Now it's August and the 6-trillion-parameter model — which, according to Musk, has a "growing 10% chance" of reaching artificial general intelligence — is still training, with no confirmed date. In the meantime, xAI has been shipping other models (Grok 4.5, Grok 4.6...) that keep getting confused with the long-awaited Grok 5 in hasty headlines all over the internet.

Llama 5: it exists, it doesn't, nobody knows

And here the soap opera turns into a full-blown detective mystery. Some websites published, with every possible technical detail, that Meta released Llama 5 on April 8, with 600 billion parameters and a gigantic 5-million-token context window. The problem is that later reporting, from June and July, categorically states that this launch never happened — that Meta's most recent open model is still Llama 4 Maverick, and that the company reorganized its AI strategy under the name "Muse," without ever officially using the name Llama 5. In short: a slice of the internet described, with spec-sheet precision, a product that apparently does not exist. If that isn't worthy of its own episode in the AI edition of a crime drama, I don't know what is.

Meanwhile, in China…

While the American labs debate naming and push back dates, the Chinese have simply been shipping model after model: Moonshot's Kimi K3, Zhipu's GLM-5.2, and Alibaba's Qwen3.6 family, all available and running on real benchmarks — none of this "we're training" or "we'll announce soon." It's almost as if one side of the cast were still rehearsing the scene while the other is already filming.

Conclusion

In the end, 2026 is turning out to be the year in which "soon" became a synonym for "who knows when," and telling rumor from fact became a spectator sport. GPT-6 (or Astra, or whatever it is) remains a secret, Gemini 4 is a promise training away in some data center, Grok 5 has already missed two premiere dates, and Llama 5 may exist only in the fertile imagination of a few websites. The only certainty is that, just like in any good soap opera, the next chapter could drop at any moment — or be postponed again. Stay tuned.

Receba as publicações

Novos artigos sobre IA, Vibe Code e Builder Code — por e-mail ou Telegram.

ou
Receber no Telegram

Ao se inscrever, você concorda em receber e-mails/mensagens e com a Política de Privacidade. Você pode cancelar quando quiser. Sem spam.