Tech • AI • Robotics • Game
HUGE GPT-6.5 LEAK! NEW Claude Fable 5.5, Gemini 4 Carbon, MiniMax M3.1 Flash & Kimi K3.1! AI NEWS
9.4/10TL;DR
A fresh wave of AI model leaks, benchmark claims and infrastructure updates suggests an unusually crowded release cycle across OpenAI, Anthropic, Google, Chinese labs and Nvidia.
KEY POINTS
Possible signs of new OpenAI models
Speculation intensified after a website tied to open.si.lol appeared to contain a hidden reference to “6.5” and a countdown ending on November 1 at 10:00 a.m. San Francisco time, matching day 28 of OpenAI’s current shipping countdown. The signal is unverified, and archived records reportedly described the domain as an independent parody site previously offered for sale, but the timing fueled expectations of a model announcement.
Microsoft code references add weight
Separate reports point to GPT 6.1 Terra appearing in Microsoft pull requests, where 18 evaluation configurations were said to be updated from GPT 5.6 Terra to GPT 6.1 Terra. Unlike the countdown clue, this suggests active testing inside Microsoft’s evaluation pipeline and indicates preparation for a potential next-generation Terra release.
Anthropic model rumored to be more proactive
A reported upcoming Anthropic model is described by an anonymous tester as better at inferring user intent and spotting broader classes of problems. In coding, the model allegedly does more than fix a requested bug, instead recognizing related issues and asking whether the wider problem should be addressed, a behavior that would mark a significant step for autonomous software agents.
Gemini 4 Carbon appears early but underwhelms
Google’s Gemini 4 Carbon is also said to be nearing release, yet early checkpoint testing has not matched lofty claims that it could rival top coding systems. Early Carbon B and Carbon C outputs were described as only marginally faster than Gemini 4 Argon while showing weaker coding quality, suggesting the model remains unfinished despite expectations of a broader launch window.
MiniMax M3.1 Flash arrives with aggressive pricing
China’s MiniMax M3.1 Flash has moved out of preview and is being positioned as a high-speed, low-cost coding and agentic model. The company says it evolved from the earlier Space Money Alpha system and has already processed more than 200 trillion tokens. It is priced at $0.15 per million input tokens and $0.60 per million output tokens, with cached input as low as $0.015.
Benchmark claims challenge premium rivals
MiniMax says M3.1 Flash reaches 62.2% on Automation Bench, compared with 50.3% for Claude Opus 5 and 45.8% for GPT 5.6 Soul, and scores 74% on DeepSWE. In example visual coding tasks, the model was presented as cheaper than DeepSeek 4.1 Flash and far cheaper than higher-end rivals, though quality varied and benchmark claims remain difficult to verify independently.
Larger Chinese releases are lining up
Additional activity is building around Qwen 4, GLM Omni and Kimi K3.1. GLM Omni reportedly combines image understanding and image generation, while Kimi K3.1 was said to appear briefly in a pull request with support for image input and four reasoning levels: low, medium, high and max. Early access for Qwen 4 also appears to be approaching.
Legacy code modernization becomes a priority
OpenAI is highlighting progress in understanding older enterprise languages such as COBOL, Fortran and Visual Basic. That matters because large banks, governments and long-established corporations still rely on these systems, and migration is notoriously difficult due to decades of embedded business logic and scarce institutional memory.
Security and tooling updates broaden the race
Anthropic launched OSS Scanner, offering free periodic AI-assisted security scans for critical open-source projects, including proof-of-concept findings and suggested patches where possible. Meanwhile, DeepSeek 4.1 Flash 1017 was reported to become free in one coding environment, and ZAI’s coding platform added a workflow mode for orchestrating multiple AI agents at once.
Nvidia reports major inference speed gains
Nvidia says it reached 423 tokens per second per user on the 2.8 trillion-parameter Kimi K3 model, a 1.68x speedup on the same architecture. The gain reportedly comes from new decoding kernels and kernel fusion techniques, underscoring how software optimization alone can materially cut inference overhead without changing the model or the hardware.
CONCLUSION
The current AI race is being shaped as much by leaked model names, coding benchmarks and infrastructure tuning as by official launches. If even part of the reported release pipeline materializes, the next few weeks could reset expectations for model capability, cost and deployment speed.
Ask a question