> SPOTLIGHT
WHAT MATTERS TODAY
Artificial Analysis reports that DeepSeek V4 Flash 0731 scored 50 on its Intelligence Index, up 10 points from the previous version and ahead of V4 Pro on that measure. The model offers a 1-million-token context window, 284 billion total parameters with 13 billion active, and improved performance across agentic evaluations including GDPval-AA, Terminal-Bench, and τ³-Bench Banking.
Cline, an open-source coding agent that can read, edit, and run code inside an IDE, recorded a Terminal-Bench score of 82.7, up from 56.9 in the April preview. A separate API update shared by Kim Monismus said the model was live for developers at roughly $0.28 per million input tokens and $0.87 per million output tokens; DeepSeek is also expected to release open weights later.
If those gains hold up in real-world use, the pressure will move toward cost per task rather than the highest benchmark score alone. A model that is stronger, cheaper, and more token-efficient can change how builders allocate models across long-running workflows.
Amazon has invested the full $50bn it committed to OpenAI this week, giving it roughly a 5 per cent stake in the company behind ChatGPT. According to the Financial Times, Amazon invested $15bn in February and planned to invest the remaining $35bn if OpenAI met milestones such as an IPO or an AI breakthrough. Those conditions were not met, but Amazon completed the investment anyway.
The deal is bigger than a standalone equity purchase. It ties capital, cloud, chips, and model-training capacity into one commercial relationship. After OpenAI renegotiated its contract with Microsoft in April, AWS gained more room to serve OpenAI. Amazon has also invested $18bn of a potential $33bn commitment to Anthropic, while using both labs to drive adoption of Trainium and its cloud services.
Amazon is turning an infrastructure relationship into ownership of the labs that can create demand for that infrastructure. As frontier AI becomes more capital-intensive, the advantage may accrue to companies that control compute, distribution, and the upside generated by the models.
Noam Brown is an OpenAI researcher focused on AI systems that can reason through complex problems. Sébastien Bubeck is also an OpenAI researcher whose work sits at the intersection of mathematics, algorithms, and AI-assisted science. They are sharing early results from Astra, OpenAI's internal AI system for scientific reasoning: the ability to work through multi-step research questions rather than simply answer a prompt.
Brown says one version of Astra handled 10 major open problems in mathematics, quantum complexity, and theoretical computer science. Bubeck shared several examples, while OpenAI released proofs with Lean certificates and walkthroughs. In plain English, the answers come with formal “checking evidence” that mathematicians can inspect step by step, instead of relying only on a convincing-looking response.
That changes how a reasoning model can be evaluated: quality is tied to verifiability, not just to an impressive demo. This remains an OpenAI claim, so the independent strength of the results will need to be tested through the certificates and follow-up evaluations.
> PRESENTED BY BELAY
MAKE TAX SEASON SIMPLE
Tax season doesn't have to mean wondering if you have the right forms, second-guessing your deductions, or scrambling to pull everything together before the deadline.
With BELAY’s tax prep support, you can approach tax season with confidence. Stay organized with one centralized place to gather and check off your documents, keep track of valuable deductions like HSA contributions, education expenses, and childcare credits, and lean on experienced professionals who make tax preparation accurate, efficient, and completely hands-off.
Download BELAY's free Personal Tax Checklist and start preparing with confidence, today.
> SIGNAL HEADLINES
CAPTURE THE SHIFT
Thinking Machines Stages Open-Weight Access: Thinking Machines argues that releasing all model weights indiscriminately is unsafe, but keeping capable models inside a handful of labs is not a solution either. The company is testing staged access through Inkling to widen availability while keeping more control over risk.
Palantir: Sovereign AI Means Context Control: In Palantir's framing, AI sovereignty is about owning, protecting, and compounding an organization's data, workflows, and internal decisions. That is a company positioning statement, but it points to a practical question: models can be swapped, while an organization's context is harder to reproduce.
Evaluation Environments Become Security Boundaries: OpenAI and Anthropic show how model-evaluation environments can become risk points when they contain sensitive data, tools, or permissions. AI security therefore extends beyond model weights to the harness and evaluation setup around them.
Anthropic Splits From the Open-Weight Push: Axios reports that Anthropic did not join Nvidia's letter calling for open weights, while Google and OpenAI did. Anthropic continues to support controls around chips, distillation, and safety testing rather than a blanket ban, showing how open-versus-closed positions map to different business models and risk tolerances.
> ONE PRACTICAL USE OF AI TODAY
Build a Context-First AI Harness
Starting from Chamath's post about the structure of the AI stack, the useful lesson for founders and operators is not to find one model that will “win.” A more durable advantage sits in the harness around the model: proprietary data, workflows, business rules, and evaluations that the company builds itself.
Start with four layers:
Proprietary data: give the model access to information competitors do not have.
Workflow: turn that data into repeatable steps with clear inputs and outputs.
Business rules: define what the model can do and what must be handed to a person.
Evaluations: measure quality against real business situations, not only general benchmarks.
Keep the model replaceable. When a cheaper or stronger model appears, you should be able to swap it without rebuilding the operating advantage around it.
> WORTH READING
ANALYSIS & THESIS
McKinsey focuses on the changes to infrastructure, data access, governance, cost control, and operating models that follow when companies deploy agents. The article shifts the question from which model is better to what operating system an organization needs for agents to work safely and create value.
Ars Technica examines why coding-agent performance depends on context and harness design, not only better retrieval or a stronger base model. It is the most technical reading in this issue, and a useful one for builders who want to understand the control layer above the model.
