Good morning, and welcome back to Techzip AI Newsletter.

Today’s Techzip covers OpenAI’s GPT-6.1 Sol, which brings near-Astra performance to a lower-cost workhorse; Anthropic’s research on GLM-5.3 and how open-weight models can expand cyber capability; and DeepSeek’s software push with Huawei Ascend and TileLang against Nvidia’s CUDA moat.

We also look at AgentWorld’s evidence that multi-agent systems often fail at coordination, not individual reasoning, plus how Anthropic and Alex Immerman are thinking about user intent and agentic commerce. The issue also tracks Dots, Codex Security, AI judge bias, and the changing economics of agent deployment.

OpenAI has released GPT-6.1 Sol, replacing GPT-6 Sol just seven days after its launch. API pricing is $2 per million input tokens, $10 per million output tokens, and $0.10 for reused input context — a 95% discount on that portion. Artificial Analysis says the base price is unchanged from GPT-6 Sol, while the cost of using Sol for autonomous tasks has fallen slightly.

On DeepSWE v1.1, a software-engineering evaluation, Sol matched GPT-6 Astra at roughly one-fifth the cost. On OSWorld 2.0, where models operate computers through multi-step tasks, it trailed Astra by only 2.1 percentage points while costing about one-seventh as much per run. On scientific tasks involving data analysis, simulation, and proof, Sol averaged $5.47 per run versus $23.21 for Opus 5.5 and $23.80 for Astra; Astra still led the evaluation at 68.1%.

Artificial Analysis places Sol one point below Astra and four points above GPT-6 Sol on its composite index. Sol also gained 12 points on Terminal-Bench 4.0, while its hallucination rate in another test fell from 60% to 54%. The tradeoff is that Sol generates 10–30% more output than GPT-6 Sol. The important shift is not that Sol beats Astra everywhere; it brings near-frontier performance into workflows where Astra's cost would be harder to justify, while Astra retains the edge on the hardest tasks.

In Anthropic's testing, GLM-5.3 completed an exploit in 50 of 410 ExploitBench attempts. In another test, it produced a full control-flow hijack in 4% of cases. Once safeguards were bypassed, the model continued interacting with malicious requests 64%, 92%, or 100% of the time depending on the conditions.

These results do not predict every real-world attack. They do show why open-weight distribution changes the safety problem: the safeguards applied to a hosted model can be modified or removed when the weights are downloaded and deployed elsewhere.

For open models, cyber risk therefore depends not only on the model's capabilities, but also on how those capabilities are distributed and operated.

Reuters reports that DeepSeek has released programming tools for Huawei's Ascend chips, including libraries for computation and data transfer, and introduced TileLang, a simpler way to write programs for the hardware. This is the layer between a model and a chip: it determines how easily developers can move an AI workload onto new hardware.

CUDA's advantage is not only that Nvidia makes powerful chips. Years of tools, libraries, and developer habits have been built around it. If DeepSeek and Huawei make it easier to write, run, and maintain programs on Ascend, the cost of moving to alternative hardware falls.

Huawei is not replacing Nvidia yet. DeepSeek is targeting the software moat that has made Nvidia's ecosystem difficult to dislodge.

Founders usually look for ROI in obvious places: ad spend, hiring, software costs, or sales performance. But one of the highest-leverage investments is often harder to see on a balance sheet: operational clarity. When a company’s internal workspace is messy, the cost shows up through duplicate work, slower decisions, missed handoffs, lost context, and founder time spent chasing information. Even small improvements can create meaningful returns. If a 10-person team saves just 20 minutes per person each day by reducing searching, re-asking, and rebuilding work, that adds up to more than 30 hours recovered every week. For an early-stage company, that is nearly the equivalent of adding another full workweek of capacity without increasing headcount. A clean workspace also reduces risk. Decisions are easier to audit, customer and contract details are easier to preserve, onboarding becomes more consistent, and AI becomes more useful because it has better structured knowledge to work with. AI does not replace good systems. It compounds them. The more organized the workspace, the more value founders can unlock from automation, search, summaries, and knowledge management. Turn operational clarity into a competitive advantage.

Try Notion today and get up to 3 months free of Notion Business with Notion AI.

AgentWorld is a paper and benchmark for testing whether multiple agents can actually coordinate, not a tool to install in a workflow. Researchers gave 3–20 agents different roles and resources in a simulated world, then asked them to complete tasks over 25–55 rounds using four models: Gemini 3 Flash, Claude Haiku 4.5, GPT-5 Mini, and DeepSeek R1-70B.

The bottleneck was coordination, not just individual reasoning. The best model completed 52% of tasks, and fewer than one-third of actions moved a task toward its result. Repeated failures included missing information, duplicated work, confused roles, and stale plans. GPT-5 Mini sent 44.1 messages on average but reached 36% success; Gemini sent 11 and reached 52%. Stale or redundant messages made up 37.7% of communication errors.

Before adding more agents, use three checks:

Anthropic is conducting an opt-in interview study with Claude and Claude Code users about the future they want from AI and how they want it to fit into work and daily life. The company makes clear that the sample is not representative of its entire user base. The study has not produced adoption or preference findings yet; its value for now is showing how a lab is gathering user intent before setting product direction.

Alex Immerman argues that an assistant could become the place where consumers decide what to buy, shifting bargaining power among marketplaces, merchants, payment providers, and fulfillment networks. His post puts Amazon, Instacart, Shopify, Toast, and DoorDash inside one question: when an agent brings the order, who owns discovery, payment, and the customer relationship?