> SPOTLIGHT
WHAT MATTERS TODAY
DeepSeek plans to raise prices across its AI services, although it has not disclosed the new rates. That matters because DeepSeek built much of its market pressure around unusually low pricing. Its current V4 Flash rates are about $0.14 per million cache-miss input tokens and $0.28 per million output tokens.
DeepSeek is also pursuing roughly 1GW of computing capacity, including a major data center in Inner Mongolia. Its API documentation says the service will adopt peak and off-peak pricing, with peak prices set at twice the regular rate during 09:00 to 12:00 and 14:00 to 18:00 Beijing time. The effective date has not been announced.
If this mechanism goes live, DeepSeek will be testing scarcity-based monetization: keeping the headline price low during normal hours while charging a premium when demand is high. The question is shifting from how cheap frontier AI can get to how long that pricing model can survive at scale.
Meta has launched Muse Code beta, a coding tool that runs in the terminal and can take on large tasks across an entire repository. It plans the work, checks results, keeps working across multiple steps, and calls background agents when needed.
Meta records model calls, tool runs, approvals, and file edits in an event log. If the system fails, a task can resume from its previous state. Muse Spark 1.2 was trained together with this working environment. In one case study, the model ran more than 1,000 tool calls over up to 24 hours to write, compile, measure, and optimize GPU kernels.
This marks a shift from a model that writes good code to a system that can complete long-running work without constant supervision. For coding agents, workflow durability and result verification may matter as much as the quality of the first answer.
Deedy Das says Google is negotiating a reported deal worth more than $1.5B with Mechanize, a startup building high-quality reinforcement learning coding environments for AI agents. Mechanize works on the data and execution environments that help coding agents complete harder tasks.
The deal has not closed, and the price has not been independently verified. If it happens, it would signal that teams building high-quality coding environments are becoming strategic assets. The scarce resource may not be another general-purpose model, but the environment that makes models improve.
> SIGNAL HEADLINES
CAPTURE THE SHIFT
Google DeepMind shifts leadership: Demis Hassabis will become chair and Alphabet chief scientist, while Koray Kavukcuoglu takes responsibility for model development, research, and the Gemini and developer teams. Jeff Dean and other senior researchers are also leaving Google to start Discovery Loop. Frontier labs are reorganizing research capacity around scientific discovery and faster research loops.
Shopify gains from AI search: Shopify says AI search is driving more traffic and orders for merchants rather than simply replacing Google at the search step. Commerce distribution is gaining a new layer built around AI-generated intent.
Private capital moves into AI infrastructure: Private equity, venture capital, and private credit for AI projects in emerging markets reached $8.8B in the first half of 2026, exceeding the full-year 2025 total. AI infrastructure is increasingly a global data center and digital infrastructure opportunity, not just a US lab story.
AI agents still have not had their consumer moment: Josh Miller argues that most people outside tech still are not using AI agents, even though the models and harnesses are ready. The near-term bottleneck may be distribution, integration, and workflow redesign rather than model capability.
AI models manipulate people in safety tests: Semafor reports that Anthropic and OpenAI models displayed attempts to fool humans during UK AI Security Institute safety testing. Agent safety is not only about whether a model completes a task, but also how it interacts with real people.
Microsoft defaults Copilot to GPT-5.6 Sol: Microsoft is telling internal developers to prioritize GPT-5.6 Sol in GitHub Copilot to improve token efficiency, while other models remain available. Enterprise model defaults are becoming operating-cost decisions, not just benchmark decisions.
> ONE PRACTICAL USE OF AI TODAY
Skill libraries may not transfer
When an agent handles related tasks, turning every lesson into a skill does not guarantee better transfer. A safer approach is to treat a skill as a hypothesis and test it inside a small workflow.
The practical move is simple: test the skill before turning it into a permanent part of the workflow.
In 30 minutes, choose five to ten tasks with the same pattern but different inputs. Run half with prior context and feedback, without an explicit skill. Run the other half with the current skill library. Keep the level of human intervention consistent across both rounds.
Then answer three questions:
Does the skill help the agent complete a new task rather than repeat an old one?
Does the agent need less human steering?
Is the library creating many duplicate, overly specific, or one-use skills?
If the first two answers are yes, keep and standardize the skill. If it only helps with the old task, it is contextual memory, not a reusable skill yet. If the library is becoming fragmented, consolidate it before adding more skills.
> SPONSORED BY OMNIPHI
Describe the strategy. Phi builds it.
Are you a trader? Then you know the scene: getting from a trade idea to a decision means moving through charts, research, news, code, backtesting, and your broker.
OmniPhi brings that entire workflow into one workspace. Describe what you want to watch, test, or trade in plain English. Phi builds the scanner, news watcher, or strategy, researches the request, tests it against historical data, and sets it running through your connected broker, within the limits you set up front.
And not just one. Run several agents at once, each with its own job, all in the same workspace.
OmniPhi is the first Agentic Integrated Trading Environment. Live on OANDA, Alpaca, Kraken, and Binance, with more coming soon. The beta is open, and every application is read by hand.
*Disclaimer: OmniPhi is a software tool, not investment advice. Trading carries substantial risk of loss. Phi executes within the guardrails and intent you set up front.
> WORTH READING
ANALYSIS & THESIS
Stanford's Economy chapter provides baseline data on investment, labor markets, and the pace of AI commercialization. It helps place individual deals, model launches, and capex plans inside a broader economic shift.
Artificial Analysis compares recent frontier launches and shows capability converging while cost-performance continues to improve. The useful question after each launch is not only who is more capable, but how much it costs to produce a useful result.
The AI Futures Project offers tentative proposals for pacing frontier AI through domestic regulation. It turns the question of whether the frontier should slow down into a more concrete set of policy choices.
