Tech • AI • Robotics • Game
OpenAI's New GPT-ADA! GPT-6.2 Astra Is Fable 5.5 Level?! Hy4 Is Out, DeepSeek V5, & More! AI News
9.4/10TL;DR
A wave of AI announcements highlighted rapid gains in coding, multimodal generation, hardware design and infrastructure, alongside fresh concerns about verification, security and power limits.
KEY POINTS
OpenAI model leaks suggest a major jump
Signs of a forthcoming GPT-6.2 Astra intensified after model-selector sightings and benchmark claims tied to an internal system associated with a reported Navier-Stokes breakthrough. On Bench CAD 2.0, the internal model was cited at 0.875 versus Astra’s 0.594, while its lowest reasoning mode allegedly reached 0.758, already above Astra’s top setting. Two brief model-tracker entries, RC ADA OR VM and RC ADA SHVM, added to speculation before disappearing.
SVG and coding demos fuel model race
Early examples attributed to GPT-6.2 Astra showed detailed SVG portraits, game-scene recreations and scientific diagrams, including a labeled animal cell and a highly detailed Minecraft-style scene reportedly written with millions of characters of code. Some viral posts were likely aided by OpenAI image or video tools rather than pure SVG alone, but the results still point to stronger multimodal coding. Rival unreleased model Fable 5.5 is also drawing attention for advanced coding and long-form creative generation.
AI moves further into physical engineering
OpenAI said ChatGPT and Codex are being used for mechanical design, electronics and simulation work. One internal prototype device for interacting with Codex was reportedly designed in under a week, with AI contributing to the enclosure, circuit board, interface and software, though human review remained necessary for manufacturing. The company is also working with Chip Ganassi Racing on turning 3D scans into editable CAD models and with Mazda on AI-assisted vehicle design simulations.
New assistants target finance and game development
A new voice feature lets users connect financial accounts and query spending, savings and transactions conversationally through ChatGPT. Separately, Unity launched a Codex plugin to give the model specialized skills for game-development tasks inside the engine. The update expands AI from general coding help toward domain-specific workflows.
Anthropic pushes enterprise tools and proof systems
Anthropic introduced easier plugin connections for Claude in team and enterprise settings through an extension to MCP, allowing organizations to centrally manage app access. It also open-sourced FlashProof, a cryptographic system designed to verify that a provider actually used the AI model it claims to run. In tests on Meta’s Llama 4 Maverick, proofs were reportedly generated at 110 tokens per second on a single Nvidia H100, with proof sizes around 19 MB, far smaller than earlier methods measured in tens of gigabytes.
Chinese and open models intensify competition
Tencent’s HY4 arrived as a new open-weight contender with vision capabilities and broad productivity aims. Company benchmarks placed it at 75.2 on Apex Agents, ahead of some DeepSeek 4.1 Flash results, and pricing was listed near $0.83 per million input tokens and $2.50 per million output tokens. Another surprise entry, Crouton Alpha, appeared free on OpenRouter with a 1 million-token context window and multimodal support, with multiple clues suggesting a possible link to Meta.
AI starts helping design chips and factories
Reports said Anthropic is using Claude to help a small team designing in-house AI chips, while OpenAI is also pursuing custom silicon. DeepSeek was reported to be preparing version 5 primarily on Huawei Ascend 950 chips, a potential shift away from Nvidia for large-scale pretraining. In robotics, Xiaomi’s CyberOne humanoid reportedly completed an 8-hour shift in an electric-vehicle factory, sorting and replacing parts while coordinating with mobile robots.
Security and infrastructure risks loom larger
A major crypto theft underscored software-supply-chain dangers after Binance acknowledged a breach linked in reports to a compromised development-tool installation; the haul was placed at about $6.34 billion. At the same time, Elon Musk warned that AI’s next bottleneck may be electricity rather than chips, arguing that power generation outside China is not expanding fast enough to match data-center demand by 2026. The warning aligns with broader concerns that compute growth now depends as much on grids and permits as on model quality.
CONCLUSION
The latest developments show AI advancing well beyond chatbots into design, coding, hardware and robotics, while trust, security and energy constraints are becoming central competitive factors. The next phase of the industry may be defined as much by verification and infrastructure as by raw model capability.
Ask a question