Rafomac News, Tech & Trends That MatterOpenAI’s codename ‘Bell’ for its next AI model has just leaked, revealing a monster pre-trained with over 10 trillion parameters—potentially post-GPT-6. Meanwhile, Anthropic is quietly launching Fable 5.1 and Opus 5 upgrades, and Alibaba teases its next-generation Qwen 4. The AI arms race is accelerating fast.
OpenAI’s ‘Bell’: More Than GPT-6?
OpenAI appears to have finished pre-training a colossal new AI model codenamed Bell, described as a successor to the recently discussed Doug. This beast reportedly boasts over 10 trillion parameters, putting it on par with the scale of GPT-4.5. But what’s staggering is that OpenAI doesn’t see Bell simply as GPT-6, but possibly as a post-GPT-6 foundational model, with ambitions of reaching their Artificial General Intelligence (AGI) milestone as early as this year.
Imagine this: Astra, the model many expect to be the next big leap, is not even out yet. GPT-6 is still on the horizon. And here we have Bell, potentially the true game-changer, already pre-trained and looming as the future base for advanced AI systems. While much remains uncertain—its exact role, the next steps in training, and the product it will underpin—the scale and intent behind Bell sound like a seismic shift in AI development.
Such progress hasn’t gone unnoticed in the competitive sphere. Rumours suggest that Anthropic, OpenAI’s key rival, is feeling the heat. Reports indicate that OpenAI’s compute resources and innovations might give it an upper hand through the coming months, even as Anthropic prepares stronger responses for next year.
Anthropic Readies Fable 5.1 and Opus 5 Upgrades
Anthropic is keeping pace with a few promising teasers of its own. Two new Claude models—Marshmallow EAP and Melon EAP—have quietly surfaced. Early indications highlight Marshmallow as a standout, outperforming Melon and reportedly beating the current Fable 5 model.
What’s more, Marshmallow demonstrated a staggering one million token context window during testing, a leap in handling massive textual context. Interestingly, it showed up separately from Opus 5 in the same model list—suggesting Marshmallow is a fresh Fable 5.1 checkpoint rather than part of Opus upgrades.
API statements hint that Fable 5.1 might already be routing on Claude Web for select users, with reports of surprisingly coherent and visually appealing outputs, like well-designed landing pages generated entirely by the AI. Anthropic seems eager to push these updates fast, potentially releasing them officially very soon to clear the stage for their next phase of innovation.
Alibaba’s Qwen 4 and GLM 5.3 Flash Make Waves
Alibaba is not sitting still either. Their Qwen team recently launched Qwen 2.8 Flash, an open-weight multimodal model that previews the architecture behind Qwen 4. Though Qwen 2.8 doesn’t have the full power of Qwen 4, the promise of a more efficient, smaller-parameter model packing heavyweight capabilities is clear. Imagine Qwen 4’s potential when scaled; it’s set to rival other AI giants with impressive performance and efficiency.
Meanwhile, the GLM team introduced GLM 5.3 Flash, a 320 billion parameter mixture-of-experts (MoE) model that boasts a 1 million token context window and multi-modal abilities. Released under an MIT license, it stands as a competitive open-source alternative to proprietary AI systems, signaling growing democratization and innovation in the AI field.
Google Pushes Smart Transcription with Gemini 3.5
Google unveiled a new Gemini 3.5 transcription model, which isn’t just about converting speech to text. This model smartly filters filler words like ‘um’ and ‘ah’, cleans up rambling phrasing, and condenses lengthy speech into smoother, clearer text. Through a demo streaming live audio with two transcription modes, Google showed how this is a smarter listener and editor, promising better real-world transcription quality for various applications.
OpenAI’s Jalapeno Chip Benchmarks Show Major Gains
On the hardware front, OpenAI’s first custom inference chip—codenamed Jalapeno—delivered impressive real-world benchmark results. Tested against models ranging from 120 billion to one trillion parameters, Jalapeno achieved 1.5 to 1.9 times more AI work per watt and reduced end-to-end latency by up to 3.6 times compared to competing systems.
This chip is poised to speed up GPT responses, improve responsiveness for AI agents and Codex, and dramatically cut power consumption. OpenAI plans to begin low-volume deployment soon, scaling throughout next year to power their burgeoning AI infrastructure sustainably.
For curious readers, more details and benchmarking insights are available beyond this overview, illustrating how rapidly AI ecosystems are evolving across software and hardware innovations.
