Tech • AI • Robotics • Game
Huge Claude Sonnet 6 Leak, Anthropic's Secret Fable 6?! GPT-6.2 Astra Today, & More! AI News
9.4/10TL;DR
Anthropic, OpenAI, Google, and Meta are simultaneously advancing unreleased models, specialized tools, and infrastructure, underscoring an AI race that continues despite political and market turbulence.
KEY POINTS
Anthropic model rumors intensify
Tests of an unreleased Claude checkpoint suggest a stronger Sonnet-tier model may be nearing release, potentially branded as Sonnet 5.6, Sonnet 6, or a new Opus variant. In one reported coding task, the model generated a detailed pelican-on-bicycle illustration entirely through code using 26,599 SVG elements, more than 6,500 feather strokes, and over 1,800 lines of Python. The checkpoint also appeared to reflect a newer knowledge cutoff than earlier Sonnet releases.
Independent report points to faster Anthropic progress
Research group METR reported that Anthropic’s internal AI progress is advancing at roughly 1.9 times its earlier pace, with AI itself estimated to drive about 80% of that acceleration. The group reviewed three unreleased internal systems labeled Model 2, Model 3, and Model 4, developed between July and October 2026. The newest system reportedly exceeded 50% across all 15 benchmarks in Anthropic’s capability index.
Small benchmark gains may hide major capability jumps
Model 4 was reported at about 178.66 on Anthropic’s capability index, versus roughly 174-175 for Opus 5.5. That gap looks modest numerically, but similar improvements have historically translated into about a doubling in the length of tasks models can complete autonomously. Anthropic leadership also indicated parts of the evidence behind the acceleration claims were withheld to avoid giving rivals a roadmap.
OpenAI unveils GPT-6 Cyber
OpenAI introduced GPT-6 Cyber, a model tailored for advanced cybersecurity research. It is being rolled out first through the Daybreak Red program to selected users rather than the general public. On the ExploitGym benchmark, it reportedly scored 48.5%, ahead of GPT-6 Astra at 42.4%, and has already been used in Linux kernel security work alongside corporate security teams.
More OpenAI launches appear imminent
Signs in internal model listings fueled speculation that GPT-6.2 Astra and Soul could follow soon. Separate remarks also highlighted the scale of compute being deployed on narrow scientific and security problems, including about $1 million in compute used on Navier-Stokes work and a similar amount reportedly used to identify around 1,000 vulnerabilities in Linux, far above the pace of conventional discovery.
Google expands video generation and data center spending
Google released Gemini Omni 1.2 Flash, a video model with stronger visual consistency, better lip-sync, and conversational editing that allows users to modify a generated clip without recreating it from scratch. Paid users can create clips from 3 to 15 seconds, and developer pricing was cut by about 30%, with draft video generation listed at roughly 7 cents per second. The model reportedly ranks first on a text-to-video arena benchmark, ahead of prior Gemini, Seedance, and Flux systems.
Google bets big despite political headwinds
Even as AI-related stocks fell and debate intensified over stricter oversight of data centers, Google announced a $17 billion investment in data center infrastructure in Wyoming. That move contrasted with calls from some state-level political figures to pause new AI data center construction pending tighter regulation. The split highlights how concerns over electricity, water, and local impacts are colliding with continued infrastructure expansion.
Meta tests new checkpoints and consumer AI features
Meta is reportedly testing a new checkpoint in its model family under the codename Walrus, while other stealth models such as Crouton Alpha and Pulse Alpha have surfaced in external testing channels. Early descriptions portray them as fast, multimodal systems with long context windows and improved output quality. Meta also introduced Muse Realtime Avatar, which can animate images into responsive speaking characters at about 25 frames per second with roughly 870 milliseconds of initial latency.
AI tools spread into finance, coding, and medicine
Anthropic launched a US beta for Personal Finance in Claude, allowing users to connect bank accounts through Plaid and analyze up to two years of transactions with read-only access. Cursor added app building and deployment from within its coding environment, while Devin Review 2.0 was presented as up to 5 times cheaper and 1.5 times faster for pull-request review. In research, an AI-guided mouse brain surgery workflow cut landmarking time from 169 seconds to 86 seconds and reduced positioning error from 0.6 mm to 0.3 mm over three procedures.
CONCLUSION
The latest releases and leaks indicate that frontier AI labs are holding back stronger systems while simultaneously expanding into specialized products, from cybersecurity and video generation to finance and robotics. Market volatility and regulatory pressure may shape deployment, but investment levels and technical progress suggest the competitive pace remains extremely high.
Ask a question