Tech • AI • Robotics • Game
Huge Gemini 4 Argon Update, Fable 5.5 Leaks, GPT-6.1 Ultrafast, Claude Motion, & Step 5! AI News
9.4/10TL;DR
Google, OpenAI, Anthropic, and StepFun unveiled a wave of AI model and product updates, led by delayed but increasingly scrutinized details around Gemini 4 Argon and a broader industry push toward faster coding agents, interactive interfaces, and enterprise automation.
KEY POINTS
Gemini 4 Argon delayed after leak
Google did not release Gemini 4 Argon at its Gemini at Work 2026 event, despite earlier signs that a launch was imminent. Leaked model strings pointed to a dedicated reasoning mode, 512K and 900K context options, quota multipliers of 1.3x and 1.8x, and tentative API pricing of $4 per million input tokens and $20 per million output tokens. Sundar Pichai described Argon as a frontier model that can process large volumes of unstructured information and complete enterprise workflows end to end.
Security concerns may be shaping Argon’s rollout
Google said Argon is being shared with a limited group of cybersecurity defenders through a specialized program, while work continues on broader safeguards. That has fueled speculation that the delay is tied less to product readiness than to the model’s risk profile. Demonstrations showed Argon handling data synchronization, pipeline building, and production deployment tasks from the command line.
Paid tiers could define Google’s next model lineup
Signals inside Google’s app and developer tools suggest Gemini 4 Argon may be reserved for AI Ultra subscribers and paid API users. Under the same structure, Gemini 4 Flash could require a paid plan, with free users limited to Flashlight. A separate Gemini 4 Flash Preview string also surfaced, indicating Google is still iterating quickly on its faster, lower-cost model tier.
Google pushes Gemini as a universal work agent
Beyond model releases, Google positioned Gemini as a general-purpose agent for knowledge work. The system is designed to operate across Gmail, Docs, Sheets, Drive, Slack, Microsoft 365, and the command line, while sharing memory and context across tools. Google also outlined support for specialized sub-agents, persistent co-worker agents, industry-specific workflows for finance and legal work, and the ability for businesses to choose between Google, third-party, and open-source models.
StepFun launches a large open model with free access
StepFun introduced Step 5 Preview, a 600 billion-parameter mixture-of-experts model with 27 billion active parameters, a 1 million-token context window, and vision support. The company said it is tuned for coding, finance, and long-running professional workloads, including tasks that run for up to 24 hours. The model is temporarily free across platforms including OpenRouter and Open Code, with weights scheduled for release on October 15.
Step 5 Preview posts strong coding results
In comparative coding tests, Step 5 Preview was presented as faster and cheaper than several rival open models on 3D generation tasks. In one example, it completed a voxel-style Japanese garden in 35 minutes, while DeepSeek V4 Pro reportedly needed 43 minutes and multiple correction rounds. The results reinforce the growing competitiveness of open-weight coding models in practical software generation tasks.
OpenAI rolls out GPT-6.1 Soul Ultra Fast
OpenAI began rolling out GPT-6.1 Soul Ultra Fast across its API, coding tools, and work-focused plans, promising near-Astra-level capability at up to eight times the speed of standard Soul. The tradeoff is price: $12 per million input tokens and $60 per million output tokens. The model is aimed at latency-sensitive uses such as coding agents and production debugging, but its premium pricing may limit mainstream developer adoption.
ChatGPT adds interactive interfaces
OpenAI also expanded ChatGPT with an intelligent UI system that turns answers into interactive visuals, forms, buttons, and charts inside the chat itself. The feature allows the model to choose when a visual or interactive response is more useful than plain text. It marks a shift from chatbot-style output toward software-like interfaces generated on demand.
Anthropic adds dashboards, animation, and cheaper caching
Anthropic launched Claude Dashboards and Claude Motion in beta. Dashboards can query connected data systems in natural language and generate live interactive views, while Motion turns reports and walkthroughs into editable animations rendered through code and exportable as MP4. Anthropic also cut Sonnet 5.5 prompt-caching input prices to $0.10 per million tokens, down from a cost basis where regular input is around $2 per million and output about $10 per million, making long-context agent workflows significantly cheaper.
AI-assisted astronomy claim draws attention
A developer named Pavle reported identifying two possible exoplanets around TIC 4206066 using Claude Code and OpenAI Codex. The project involved 74 analyses, more than 100 scripts, and a review of over 340,000 automated alerts, with successful prediction checks in three out of three hidden-year tests. The finding remains unconfirmed, but NASA’s TESS mission has approved follow-up observations in November, setting up an independent test of the predictions.
CONCLUSION
The latest product cycle shows the AI market splitting along two tracks: increasingly capable premium frontier systems and rapidly improving open or lower-cost alternatives. The next major signal will be whether Google can launch Gemini 4 Argon broadly without tightening access so far that rivals gain the advantage.
Ask a question