Google · Gemini 4 Argon · Frontier AI
Google has a new model that apparently enjoys difficult homework.
Google released Gemini 4 Argon on October 13 through its API and eligible paid services. Introductory API pricing is $2 per million input tokens and $10 per million output tokens, increasing to $4 and $20 after November 30. The launch puts a new tool in developers’ hands rather than leaving them with a preview of something they cannot use. Our interest is the work it can complete under real constraints: preserving a codebase’s behavior matters at least as much as producing an impressive-looking patch.
Google reports that Argon-assisted changes, now live in its Agent Development Kit, cut long-session preparation time by 40% and cold imports by 21%. Those are specific improvements in a particular project, not a universal programming speedup. They make a more useful claim than simply calling the model better at coding, because there is an identifiable workload and a result to examine. Our interpretation is that AI-assisted engineering becomes most valuable when the change survives the surrounding system. A faster routine is only an improvement if the rest of the application still works.
The company also reports shortcut attempts of 2% in a coding evaluation and describes a monitor for unauthorized workarounds. These are reported safeguards under evaluation conditions, not proof that authorization failures are solved. A model that encounters an obstacle needs to respect the boundary rather than reinterpret it as a puzzle to defeat. Useful autonomy therefore depends on both the work and the rules governing it. The release gives customers more capability to test; it also gives independent evaluators more consequential claims to examine. The model can apparently improve the code. Somebody still needs to check the improvements—and its boundaries.
Source: Gemini 4 Argon availability, pricing, engineering results and safeguards ↗
AI safety · Independent evaluations · Governance
The laboratories want independent inspectors. The inspectors would like to publish their findings.
Reuters reports that OpenAI and Anthropic sought significant influence over outside evaluators’ publications, citing concerns about trade secrets, according to people familiar with the negotiations. The October 13 investigation describes evolving discussions, rather than one settled rule governing every relationship. Protecting sensitive technical information is a real concern. So is allowing an investigator to explain a consequential result without the organization being investigated deciding whether the conclusion is convenient enough to release.
Anthropic disputed a broad characterization of editorial control, saying established evaluators control their methods, conclusions and publications, while it retains approval rights for reports from new evaluators as relationships develop. OpenAI said it seeks confidential-information redactions, rather than approval of whole reports, and permits disclosure of major redactions. The reporting documents contested publication arrangements; it does not by itself prove systematic suppression of unfavorable findings. The distinction keeps a question about oversight from becoming an unsupported accusation about what has already been hidden.
Our view is that confidentiality review and editorial control need different boundaries. A developer could identify a specific sensitive detail without gaining the authority to rewrite the investigator’s conclusion. An agreed dispute process could then address redactions while preserving an account of what happened and what evidence supports it. That is a proposed approach, not a universal standard already in force. Independence needs adequate access as well as freedom to report, and the two can come into tension. An inspection report becomes considerably less useful when the manufacturer gets the final word on every defect.
Source: Reuters investigation and the companies’ responses ↗
ASML · Lithography · Semiconductor capacity
The AI boom is asking the equipment manufacturer what it can do in 2030.
ASML’s October 14 results report €11.662 billion in third-quarter sales and €4.960 billion in net income—rounded to €11.7 billion and €5.0 billion in the headline. Sales were a quarterly record. The company sold 113 new lithography systems and eight used systems, 121 altogether. It forecasts full-year sales of €44–45 billion. The quarter is reported; the full-year figure remains an outlook. The distinction is useful because a successful current business and a dependable forecast are different kinds of evidence.
ASML says planned supply will not fully meet anticipated customer demand in 2027 and 2028, and it is investigating additional capacity. Some customers have expressed interest in placing orders for 2029 and 2030. That is not confirmation that every inquiry has become an order, much less a delivered machine. ASML supplies equipment used to produce chip patterns, putting the company upstream of the processors powering the boom. Our connection is that a software industry promising fast change still depends on physical tools with their own manufacturing schedules.
Expanding ahead of demand creates a forecasting problem. Too little capacity can delay customers; too much can become an expensive commitment if their plans change. Efficiency improvements might reduce the resources needed for one workload while new applications increase demand elsewhere. Those are possibilities, rather than predictions we can resolve from one earnings release. Today establishes substantial industrial execution without guaranteeing every proposed factory or data center. The AI companies are planning tomorrow’s models. The chipmakers are planning tomorrow’s factories. ASML is being asked to plan the equipment that makes those factories useful.
Source: Third-quarter results, system sales and forward outlook ↗ 