Tech • AI • Robotics • Game
Mysterious GPT-Next Model Leaked, Fable 5.5 Today? Gemini 4 Argon + RSI Update & More! AI News
TL;DR
OpenAI and Google are preparing new model releases and platform updates, while new analysis suggests Anthropic currently offers far more usage value than rivals for heavy AI workloads.
KEY POINTS
OpenAI tests a mysterious GPT Next model
A previously unseen GPT Next model has surfaced in early coding and graphics tests, with results described as slightly stronger than GPT 6.1 Sonnet-equivalent performance in some cases, especially for web development and 3D scene generation. The largest unknown is pricing: early estimates suggest unusually low cost relative to output volume, raising the possibility that OpenAI is preparing a more efficient, cheaper model rather than a major capability leap.
A 28-day product push is underway at OpenAI
OpenAI product lead Tibor Blaho has outlined a 28-day push in which the company plans to ship a meaningful improvement for users each day or provide a usage reset. The effort is focused on four priorities: simplifying products, improving efficiency, launching breakthrough features and releasing new models, signaling a faster release tempo amid intensifying competition.
First update brings a major speed increase
On day one of that pledge, GPT-6 Astra and GPT 6.1 Sonnet were made about 50% faster by default, with target generation speed rising to roughly 50 tokens per second from 30. The improvement also extends to third-party services using Sign in with ChatGPT, including tools such as OpenCode, Pi, Amp and Devon.
Google’s Gemini 4 Argon appears close
New internal model identifiers tied to Gemini 4 Argon are appearing more frequently in Google backend logs, a common sign of active pre-release testing. The visible pattern suggests Google is cycling through multiple checkpoints, fixes and summaries at speed, indicating the public launch could arrive soon and possibly with a better-tuned version than early testers saw.
Nano Banana 2.1 rolls out with confusing renaming
Google is also rolling out Nano Banana 2.1 on Gemini web and mobile after a confusing chain of renames that appears to have shifted the model from 2.5 Light to 2.1 Flash. Early side-by-side tests indicate decent character consistency and reference adherence, but image quality remains uneven and still trails stronger competitors on lighting, atmosphere and instruction following.
Speculation grows around Google and recursive self-improvement
A newly discovered internal placeholder has revived discussion that Google may be experimenting with recursive self-improvement systems. There is no proof that the company has achieved true RSI, but the rapid proliferation of Flash variants and model checkpoints has fueled claims that Google may be using automated systems to accelerate model iteration.
Anthropic leads on subscription value, report says
New analysis from SemiAnalysis finds that Anthropic subscriptions deliver substantially more API-equivalent usage than comparable OpenAI plans for heavy coding and agentic workloads. The report estimates Opus 5.5 can provide roughly 5x the value of GPT 6.1 Sonnet at similar subscription pricing, making the gap notable for users who consume large volumes of tokens every day.
The gap is especially stark at the $200 tier
At around $200 per month, the analysis estimates Claude Max can represent roughly $12,000 to $12,500 in API-equivalent use for top models, versus about $2,000 to $3,000 for ChatGPT Pro depending on the OpenAI model used. The study notes that raw usage does not equal overall quality, but it suggests Anthropic currently offers significantly more compute for the money.
OpenAI launches text watermarking for Europe
OpenAI is introducing invisible text watermarking in the EU to help meet EU AI Act requirements. The system, called Textscreen, embeds a statistical signal in word choice rather than visible markers, and the company says it matches or exceeds Google SynthID performance in testing without meaningful quality loss. The option will also be available globally to API customers on select models.
A new U.S. open model enters the field
U.S. startup Reflection AI has announced Beam, an open model aimed at reasoning, coding and agentic tasks. Beam reportedly has 501 billion total parameters with only 23 billion active at a time, a mixture-of-experts design intended to improve inference efficiency, and the company says model weights will be released later this month.
CONCLUSION
The current AI race is being shaped less by headline-grabbing breakthroughs than by speed, price, usage limits and release cadence. If the reported tests and subscription data hold, the next competitive battleground will be efficient models and practical value rather than raw model branding alone.
Ask a question