News Flash: Ultrafast is rolling out today, Former Cognition, Ramp Staffers, & Can AI automate Epoch?

1. OpenAI Launches "Ultrafast" Mode for GPT-6.1 Sol

The News Highlight:

OpenAI has officially begun the rollout of "Ultrafast" for its GPT-6.1 Sol model across the API, Codex, and ChatGPT Work environments. This new inference mode is designed to bridge the gap between high-tier reasoning and real-time execution, offering intelligence levels comparable to the flagship Astra model but at speeds up to eight times faster than the Sol Standard configuration. The release targets high-stakes, low-latency enterprise applications such as live system debugging, autonomous agent navigation, and interactive customer experiences where sub-second response times are critical for operational success.

• Performance Metrics: Delivers up to 8x speed increase over GPT-6.1 Sol Standard while maintaining near-Astra-level intelligence benchmarks.
• API Pricing Structure: Set at $12 per million input tokens and $60 per million output tokens, positioning it as a premium tier for latency-sensitive workflows.
• Availability and Compliance: Accessible via Pro 500, eligible Enterprise, and Edu plans; includes full support for data residency in both the US and EU regions.
• Expanded Regional Support: Simultaneous update adding EU data residency support for GPT-6.1 Sol Fast and GPT-6 Luna Fast models.

DO-AI Analysis:

From an architectural standpoint, the "Ultrafast" rollout represents a significant shift toward optimizing the inference stack for agentic workflows rather than just raw token generation. By achieving near-Astra intelligence at 8x speed, OpenAI is addressing the primary bottleneck in autonomous agents: the "reasoning latency" that causes drift in multi-step tasks. For engineering teams, this pricing reflects a steep premium, meaning the architectural trade-off must be justified by the cost of human intervention or the value of real-time responsiveness. I recommend utilizing Ultrafast specifically for the "Controller" nodes in agentic loops, while offloading non-critical summarization to cheaper, slower tiers to balance the unit economics.

• Primary Source: Ultrafast is rolling out today for GPT-6.1 Sol

2. Hone Secures $60M to Build Long-Horizon AI Business Agents

The News Highlight:

Hone, a new startup founded by former engineers and researchers from Cognition and Ramp, has emerged from stealth with a $60 million seed round to redefine the "AI staffer" category. Unlike current chatbots that handle transactional queries, Hone is building agents capable of managing long-running business processes that span weeks or months. The five-month-old company aims to move beyond simple automation toward full-scale autonomous business operations, where AI agents act as professional team members capable of handling complex, multi-stage projects without constant human prompting.

• Funding and Pedigree: Raised $60 million in a seed round, backed by a team with deep experience in AI coding (Cognition) and fintech infrastructure (Ramp).
• Long-Horizon Capabilities: Designed to execute tasks that require persistent state management and reasoning over extended durations (weeks to months).
• Target Market: Enterprise-grade software aimed at automating complex professional roles rather than just individual tasks.
• Operational Model: Focuses on "AI staffers" that integrate into existing business workflows to field open-ended, long-term responsibilities.

DO-AI Analysis:

The emergence of Hone signals the transition from "Stateless AI" to "Stateful Agents." The technical challenge here isn't just the LLM's reasoning capability, but the underlying infrastructure for persistent memory and long-term goal stability. Most current RAG (Retrieval-Augmented Generation) systems fail over long horizons due to context window saturation and "forgetting" intermediate constraints. Hone’s focus on agents that run for months suggests a specialized architecture for hierarchical task planning and verification. For CTOs, the takeaway is clear: the next frontier of ROI isn't in faster chat, but in the reliable delegation of asynchronous business logic to autonomous systems.

• Primary Source: Former Cognition, Ramp Staffers Want AI Agents Running Businesses

3. Epoch Research Reveals Limits of AI in High-Stakes Automation

The News Highlight:

A new report from Epoch AI, titled "Can AI Automate Epoch?", reveals that even the most advanced frontier models currently lack the judgment required to fully automate high-level research and analysis. While models like Fable 5.1 and GPT-6 Astra showed proficiency in well-defined, narrow tasks, they consistently failed in open-ended scenarios that required adhering to specific institutional standards or designing truly informative experiments. The study found that models often hallucinated findings based on flaws in their own experimental setups, highlighting a critical gap between generating "plausible" content and producing "rigorous" research.

• Model Comparison: Fable 5.1 and GPT-6 Astra are currently tied in performance, but both fail to produce "Epoch-quality" autonomous work.
• Open-Weight Lag: Open-weight models significantly underperform, struggling even with well-defined tasks that closed-frontier models handle reliably.
• Judgment Deficit: Models could identify research directions but lacked the critical judgment to follow through or identify flaws in their own methodologies.
• Information Density: Fable 5.1 was noted for producing outputs with extremely high information density, yet still suffered from generalizability issues.

DO-AI Analysis:

This report serves as a reality check for the "Full Automation" narrative. From a first-principles perspective, the failure of models to learn generalizable patterns from reference materials—even with ample context—suggests that current in-context learning (ICL) has reached a plateau for high-reasoning tasks. The "judgment gap" identified by Epoch is an architectural signal that we cannot yet replace the "Human-in-the-Loop" for verification and experimental design. Engineering teams should focus on "Centaur" architectures—where AI handles the high-density data processing while humans retain the final decision-making and methodology validation—rather than aiming for 100% autonomous research pipelines.

• Primary Source: Can AI automate Epoch?

Morning Executive Comparison Matrix