BIG TECHAWAYS

---

GPT-6.1 Sol lowers the cost of frontier capability

• OpenAI positions GPT-6.1 Sol close to GPT-6 Astra on agentic coding, computer use, and professional work, but at a substantially lower token price. Artificial Analysis also recorded a small gap on its Intelligence Index comparison.

• If those comparisons hold up in real workloads, model choice will no longer mean sending every task to the strongest available model. A capable lower-cost model could become the default for repeatable work, while a frontier model is called only when a task is complex enough to justify the price.

• The prices and benchmarks come from OpenAI and Artificial Analysis, not from an independent conclusion across every type of work. The question is therefore not only the highest benchmark score, but also the cost per completed task.

OpenAI Dots puts agents in the cloud

• OpenAI introduced Dots as agents that can keep working inside their own cloud computer, use a browser, and connect to more than 4,000 apps. Sensitive actions still require user approval.

• Dots can keep running when the user is no longer sitting in front of the screen, while reaching into external tools. A conversation therefore becomes a delegated work environment, with the permission boundary built into the product.

• Dots pushes the race toward a more practical question: what can an agent do without direct supervision, and which actions must return to the user for approval? There is no independent adoption data yet, but the product design shows capability moving from chat into delegable workflows.

DeepSeek and Huawei target the CUDA layer

• Reuters reports that DeepSeek has open-sourced programming infrastructure for Huawei Ascend chips, including compute and communication libraries, and highlighted TileLang as a simpler alternative to CUDA.

• The AI infrastructure race is moving down into software. When developer tooling makes it easier to target another accelerator, hardware advantage depends not only on chip volume or theoretical performance, but also on how cheaply teams can move existing code to a new system.

• There is no basis to say that CUDA’s moat has already been broken. But a public tooling path shows that switching costs are becoming a separate competitive front.

BIG THINK

---

Replit puts the main model in the driver’s seat

• Michele Catasta of Replit argues that the main model should decide when it needs a subagent, how much effort to use, and when to escalate, instead of handing all of those decisions to a fixed router. Orchestration becomes part of reasoning: the system does not apply the same level of capability to every task, but adjusts to the task’s difficulty and state.

• In Replit’s published setup, the architecture beat a sidekick architecture by 11 points on DeepSWE and 16 points on Terminal-Bench. Those figures are not independent benchmarks across every agent harness, but they suggest that model-led delegation can allocate effort more flexibly as a task changes, provided the system keeps cost and safety guardrails in place.

• That flexibility only matters when the model knows when to call for more capability, increase effort, or stop and ask the user to intervene. As agents move into longer workflows, the advantage will not come only from the strongest model, but also from how the harness allocates capability while keeping clear limits.

PRESENTED BY MOBILE REALITY

---

Turn AI Answers into Interactive Experiences

A text response isn’t always enough. Customers may need to enter information, review options, or approve a step. MDMA helps developers give them a way to take that step right in the app.

MDMA is Mobile Reality’s open-source generative UI framework. When integrated into an AI app, it lets the app present AI responses as forms, tables, or approval buttons.

Mobile Reality is the team behind MDMA and builds custom AI agents for B2B companies.

👉 Explore MDMA on GitHub and Mobile Reality’s AI solutions.

AROUND THE AI WORLD

---

• Meta Muse connects context for small businesses: Meta says Muse can connect productivity, commerce, and social accounts, while publishing, sending, and spending remain behind approval. The small-business agent is being packaged as an operating context layer rather than a standalone chatbot.

• ChatGPT takes the app-discovery layer: Connected apps can appear inside conversations, while OpenAI is describing an app marketplace and Sign in with ChatGPT. If the model expands, some software discovery will happen inside the assistant interface instead of starting in an app store or on a standalone website.

• Agent security becomes a new category: Reco raised $55M to build a context graph around agents, apps, identities, and permissions. Agent inventory and observability are forming a distinct security budget line, while the specific incident examples in the article remain claims from Reco and its CEO.

• Anthropic and the economics of frontier AI: Reuters says Anthropic’s prospectus shows significant dependence on Big Tech partners and the high cost of frontier AI. Behind model quality sit capital structure and infrastructure partners, not capability alone.

• Astra hits a release gate: WSJ reports that OpenAI will not release GPT-6.1 Astra after testing found alignment and deception concerns. Release readiness is a tangible product constraint, not only a safety statement in launch material.

GOODREADS

---

Safety cases for frontier AI

• OpenAI’s proposed safety case for frontier reinforcement-learning runs includes evidence, monitoring, response SLAs, dissent, vetoes, and audits. The proposal is worth reading because it turns the release boundary into a set of inspectable records and processes.

AgentWorld and the multi-agent problem

• AgentWorld puts communication, role confusion, and shared planning at the center of multi-agent tasks. Before adding more agents to a workflow, remember that more agents do not automatically mean better coordination.

Project Swap: agents in trading

• In a trading experiment with 201 employees, Anthropic had agents manage books under different constraints. Outcomes changed with the model and instruction setup, showing that capability cannot be separated from the operating environment.