Tech • AI • Robotics • Game
Huge GPT-6 Astra Minor Leak, Claude Huge Discovery, Mimo V3.0, Sonnet 5.5, & Muse Spark 1.4! AI News
9.4/10TL;DR
A wave of new AI model leaks, product updates and early research results suggests intensifying competition among major labs, with OpenAI, Anthropic, Google, Meta, MiniMax, Kimi and Xiaomi all showing signs of rapid releases or capability gains.
KEY POINTS
OpenAI model leak points to “Astra Minor”
A previously hidden OpenAI model labeled Astra Minor appeared in the company’s help documentation before references were removed. The listing reportedly placed it within the same GPT-6 family as Astra, Soul and Luna, and under the Daybreak access and safety framework. No benchmark, price, context size or launch date is confirmed, but the placement suggests a real unreleased model rather than a stray configuration entry.
MiniMax M3.1 and “Space Bunny” surfaced at the same time
MiniMax M3.1 appeared in the company’s official model catalog ahead of public release, though without a full configuration block or broad availability. Separately, a stealth model called Space Bunny Alpha showed up on OpenRouter and other services with 1 million-token context, multimodal support and zero data retention. Its tokenizer reportedly matches MiniMax M3, fueling speculation that it may be tied to M3.1, though no formal confirmation exists.
Early tests show mixed results for the stealth model
Despite its large context window and free access period, Space Bunny Alpha has so far delivered uneven performance in coding and simulation-style tasks. Comparative tests cited weak results in physics scenes and web development prompts, with rivals such as GLM 5.3 outperforming it in several cases. The model appears fast and polished for text generation, but current evidence suggests a gap between its specifications and practical capability.
Anthropic says Claude helped identify a new biological system
In one of the most notable research claims, Anthropic ran roughly 950 Claude agents in parallel for 21 hours, using about 210 million tokens to analyze more than 200,000 reverse transcriptase enzymes. During that work, the system flagged a repeating DNA-related pattern near one enzyme that resembled structures associated with discoveries that led to CRISPR. Human researchers then tested the finding in a lab and confirmed a previously uncharacterized biological system, which they named ART.
The biology result remains early but significant
Researchers have not shown that ART can be used for gene editing or that it has practical medical applications yet. Even so, the result is notable because the AI system was not instructed to hunt for a specific target and instead surfaced a pattern that warranted experimental follow-up. The episode offers an early example of AI acting as a large-scale hypothesis generator for biological discovery.
Claude Sonnet 5.5 appears to be in stealth testing
Anthropic is also reportedly testing Claude Sonnet 5.5 with select partners. Leaked details point to a 1 million-token context window, up to 128,000 tokens of output and “adaptive thinking” enabled by default. Pricing discussed in leaks includes $2 per million input tokens, $10 per million output tokens and $0.20 per million cached reads, though the company has not officially announced those terms.
Anthropic is reworking Claude Code
The company is considering removing plan mode from Claude Code and replacing it with an adjustable effort setting triggered by Shift+Tab. At the same time, it is rolling out Projects, which let Claude split work into parallel threads with shared context, and cloud sessions, which keep coding tasks running on Anthropic infrastructure even after a user closes a laptop. Existing paid users are also being offered one-time bonus credits of $100 for Pro and $250 for Max plans.
More model updates emerged across the industry
Mistral-style release patterns are also visible elsewhere. MUS Spark 1.4 appeared in directory listings, though only in a contributor form and not as a full flagship launch. Kimi K2.3.1 was spotted in online configuration files with low, high and max modes, 1 million-token context and continued Swarm support, indicating a near-term release.
Xiaomi, OpenAI, Google and Meta pushed product-side advances
Xiaomi is developing a new Mimo 3.0 architecture aimed at long-running agents, claiming around 5x fewer prefill FLOPs and a 4.5x smaller KV cache than Mimo 2.6 at 1 million tokens. OpenAI upgraded ChatGPT Voice so it can use connected apps such as email, calendar and Slack, and interact with ChatGPT Work to create documents, presentations and spreadsheets by voice. Google introduced Gemini 3.8 Flash TTS and Flashlight TTS for more expressive speech generation, while Meta unveiled lightweight AI-enabled glasses around 100 grams with large virtual displays and a broader strategy tied to AI agents and transaction-based monetization.
CONCLUSION
The latest leaks and launches show an AI sector moving on multiple fronts at once: bigger context windows, more capable agents, deeper enterprise integration and early signs of AI-assisted scientific discovery. The central question is no longer whether new models are coming, but which of these fast-moving systems will translate raw specs into reliable real-world impact.
Ask a question