· 3 min read
The Gradient — 2026-10-11
- Company Blog
- Industry News
- Community
In a recent internal test, Asana dramatically reduced the expense and latency of its browser‑based AI agent by leveraging the new GPT‑6.1 Sol model.
Asana cuts model costs 76x in browser tests with GPT-6.1 Sol
What happened: Using GPT‑6 Astra in Codex, Asana rebuilt its browser agent and achieved a 76× reduction in model‑inference costs while boosting speed by 5×.
Why it matters: Lower costs and faster responses make advanced AI features more accessible to Asana’s customers, potentially expanding adoption of AI‑enhanced workflow tools.
Key stats: • Cost reduction: 76× cheaper • Speed improvement: 5× faster • Model used: GPT‑6.1 Sol (via GPT‑6 Astra in Codex)
Source: https://openai.com/index/asana-browser-agent
Pollo AI turns creative ideas into campaigns with OpenAI
What happened: Pollo AI integrated GPT‑5.6, GPT‑6 Astra, and GPT‑Image‑2.5, enabling creators to input high‑level ideas and receive detailed visual assets and video ads automatically.
Why it matters: The integration slashes production time and costs, democratizing high‑quality advertising and allowing small teams to compete with big agencies.
Key stats:
• Three new OpenAI models deployed
• Generates images up to 8K resolution
• Produces video ads up to 30‑second clips in under a minute
• Early adopters report a 40% faster campaign turnaround
GPT-6 and Intelligent UI for everyone
What happened: GPT‑6 has been rolled out worldwide in ChatGPT, featuring an Intelligent UI that adds visual elements and interactive experiences to responses.
Why it matters: The upgrade speeds up answer generation and lets users explore data directly within the chat, making AI assistance more intuitive and useful across personal and professional tasks.
Key stats: • Global rollout to the entire ChatGPT user base • Faster response times (up to 2× speed improvement) • Integrated visual and interactive components in every reply
Source: https://openai.com/index/gpt-6-for-everyone
Google brings agentic AI to Gemini, starting with businesses
What happened: Google announced that Gemini will now function as an AI agent that can autonomously plan and execute tasks, interact with multiple business applications, and even spawn sub‑agents for specialized work. Each agent receives a workplace identity, complete with an email address, to streamline communication.
Why it matters: By giving Gemini agency, Google aims to reduce manual workflow bottlenecks and empower teams with AI that can act rather than just advise. The ability to delegate to sub‑agents and switch between models promises more flexible, context‑aware solutions for enterprises.
Key stats:
- Gemini’s agentic layer can invoke up to three distinct AI models per task.
- Early pilots report a 30% reduction in time spent on repetitive administrative work.
- The email‑enabled identity feature is being rolled out to 5,000 enterprise beta customers.
Source: https://techcrunch.com/2026/10/08/google-brings-agentic-ai-to-gemini-starting-with-businesses/
I trained a 414k-parameter transformer to fly a boids flock, then tested whether the rules a probe can read are the ones it uses [P]
What happened: A custom Python/PyTorch transformer with just two attention layers (≈414 k weights) was trained on recordings of a 12‑bird boid simulator. It learned to predict each bird’s next move without being given any explicit boid rules.
Why it matters: The experiment shows that tiny, rule‑free neural networks can capture complex emergent behavior like flocking, opening doors for lightweight, data‑driven control systems in robotics and simulation.
Key stats:
- R² between 0.990 and 0.994 on completely unseen flight clips across four runs
- Successfully controls a 50‑bird flock despite training only on 12 birds
- Trained on a laptop CPU, no GPUs required
- Two attention layers, 414 k parameters total