Opus 5 Halves Coding Costs While Over 20 Companies Warn of AI Curbs

01Near-Fable 5 coding performance now costs half as much with Opus 5

Anthropic released Claude Opus 5 on Thursday with a direct pitch to developers running expensive coding and knowledge-work workloads. The company says the model approaches Claude Fable 5’s capabilities across many domains at half the price. That turns a flagship upgrade into a routing decision: which tasks still justify Anthropic’s top-tier model?

Coding results supply Anthropic’s clearest answer. On CursorBench 3.2, Opus 5 at maximum effort scored within 0.5% of Fable 5’s peak result. Anthropic says it did so at half the cost per task. Across the benchmark’s high, xhigh, and maximum effort settings, Opus also delivered more performance for a given cost than competing models.

The release changes the calculation for teams already using Opus. Anthropic priced Opus 5 at the same level as Opus 4.8, while reporting more than twice its predecessor’s Frontier-Bench performance. Customers can also adjust effort settings to trade intelligence for lower token use, faster responses, and cheaper execution.

Those controls matter because production workloads rarely need one fixed capability level. A coding agent may need maximum effort for a difficult repository change, then lower effort for routine tests or documentation. Opus 5 gives developers another place to set that boundary before escalating work to Fable 5.

Anthropic reports similar gains outside software engineering. Opus 5 led its cited coding and knowledge-work evaluations, including Frontier-Bench and GDPval-AA. On ARC-AGI 3, which tests novel problem-solving, the company says its score tripled that of the next-best model. These are vendor-selected evaluations, but they support Anthropic’s claim that the price cut does not require a broad capability retreat.

The model does have a defined ceiling. Anthropic says Opus 5 still trails Mythos 5 on cybersecurity tasks, leaving security-sensitive routing separate from the broader cost calculation. Opus 5 is available now, serving as the default model on Claude Max and the strongest option offered through Claude Pro.

02More Than 20 Companies Say Open-Weight Curbs Would Push AI Innovation Overseas

More than 20 technology companies, including Nvidia, Microsoft, Meta and Palantir, are pressing policymakers to avoid “premature restrictions” on open-weight AI. Their joint letter says such rules could suppress competition and send development abroad. It arrives as Washington considers limits on Chinese models available in the United States.

Open-weight models let users download, modify and run the software on their own infrastructure. The companies cast that access as commercial infrastructure, not simply a publishing choice. Their letter claims open weights distribute AI’s benefits beyond a small group of closed-model providers. It also warns that relying only on closed systems carries its own risks.

That lobbying push follows the release of Kimi K3 by Chinese startup Moonshot AI. According to CNBC, the model beat leading American offerings on some industry benchmarks. Its performance intensified debate over whether Chinese open-weight systems should remain accessible inside the United States.

The Trump administration is examining a different concern. Treasury Secretary Scott Bessent told CNBC that officials would investigate whether Chinese companies had stolen American intellectual property. He said sanctions were available in response to any theft. The reported allegations have not been established, and the industry letter cautions against acting hastily.

Open-source supporters are making a broader argument outside the corporate campaign. In a commentary, developer Tom Bedor argues that commercial software already depends on layers of freely available code. He portrays restrictions as an advantage for closed-model companies that can control access and charge customers. That is an advocate’s case, separate from the companies’ lobbying document.

The two positions collide over where policymakers should place the burden of proof. Restriction advocates describe freely downloadable models as a security and competitive risk. The corporate coalition says limiting them could concentrate the market and weaken domestic development. Kimi K3 makes the dispute immediate: regulators must decide whether a model’s origin, availability or alleged conduct should trigger restrictions.

03$5 Billion in Grants Now Comes With AI Tokens and Platform Partnerships

The Trump administration’s first Genesis Mission grants put $5 billion behind hundreds of AI-driven science projects. That money is only one part of the package. Google has committed $40 million in AI tokens and credits, while OpenAI is working with the Department of Energy and national laboratories.

Together, those commitments change what federal research support can include. Agencies are funding projects, technology companies are supplying access to AI systems, and national laboratories are providing institutional research capacity. The resulting program allocates money, models, computing resources, and scientific partnerships through the same initiative.

The White House has described Genesis as comparable in urgency and ambition to the Manhattan Project, according to The Verge. Its first grants span hundreds of projects rather than one central technical objective. That scale makes access to AI infrastructure an operating requirement, not a separate procurement question for each research team.

Google’s contribution illustrates the shift. The company is not adding another cash grant to the $5 billion pool. Its stated $40 million commitment consists of AI tokens and credits, tying part of the program’s usable research capacity to services supplied by Google.

OpenAI’s role follows a similar institutional route. The company says it will work with the Energy Department and national laboratories to apply frontier AI to scientific discovery. Those labs already connect federal funding, specialized facilities, and researchers. Adding commercial AI systems places another privately controlled resource inside that network.

Researchers gain access without every institution building its own model stack or securing equivalent computing capacity. The tradeoff sits in the allocation mechanism. Credits work only on the issuing company’s services, while model access depends on the provider’s terms and continued participation.

That distinction will shape how durable the grants prove. Public money can support a project for a defined period. Tokens, credits, and external model access may carry different limits, especially after introductory commitments expire. Research institutions will therefore need to track not only grant balances, but also which experiments depend on a particular provider’s infrastructure.