The Dawn of the ‘One-Million-Token’ Era, as AI Agents Reshape Finance and Software

The prospect of AI replacing decision-making and execution functions comes into focus
Anthropic unveils an agent capable of “doing virtually anything”
The SaaS market faces pressure as “SaaSpocalypse” rhetoric gains traction

The rapid spread of artificial intelligence (AI) agents is delivering direct shocks not only to the software industry but also to financial markets. Analysts warn that execution-capable AI, by automating advisory, tax, and asset management tasks, could undermine the fee-based business models that have long defined the financial sector. At the same time, the emergence of next-generation agent models supporting one million tokens and accelerating debate over the restructuring of subscription-based pricing systems is driving a broader realignment of revenue structures and platform competition across industries. The rise of AI agents is beginning to reshape how companies operate and how markets assess their value.

Shockwaves Beyond Software, Reaching Financial Markets

On the 11th, major financial stocks including LPL Financial, Charles Schwab, Raymond James Financial, and Morgan Stanley all posted sharp declines in New York trading the previous day. The selloff followed growing concerns that AI agents capable not only of automating repetitive tasks but also of making decisions and executing actions could disrupt traditional revenue models built around financial advisory and wealth management services. Although share prices turned mixed the following day, the broader market view holds that investor selling pressure is unlikely to dissipate quickly.

The declines were pronounced. LPL Financial fell as much as 11 percent intraday on the 10th before closing down 8.31 percent, while Raymond James dropped 8.75 percent. Charles Schwab declined 7.42 percent, and Morgan Stanley slid 2.4 percent. The trigger was fintech firm Altruist’s addition of an AI-powered tax planning tool to its platform, Hazel. The tool analyzes clients’ tax returns, pay stubs, account statements, and emails to generate customized tax strategies within minutes. As tasks once performed by professionals for substantial fees become automated, the cost burden on consumers is expected to decrease significantly.

The impact of AI agents had already unsettled the software sector before reaching finance. On the 3rd, Salesforce’s price-to-earnings ratio fell to 15 times, marking a record low. Investor sentiment weakened amid concerns that AI agents capable of replacing intellectual labor could erode the value of traditional subscription software and information services. Neil Sipes, an analyst at Bloomberg Intelligence, noted that “if AI expands beyond auxiliary functions into mediation and decision-making roles, it cannot be ruled out that platform companies could absorb a significant portion of markets such as software and finance.”

Some analysts, however, view the fear as excessive. In a report released earlier this year, JPMorgan stated that while AI has pushed the cost of specialized expertise close to zero, AI agents remain tools that rely on existing software and data. Although lower-cost delivery of professional strategies and analysis could weigh on high-fee financial service models, the growing scarcity of data required to train AI agents may limit scalability. As a result, the shock to finance and software is increasingly interpreted as a phase of reassessment, as markets seek a balance between AI substitution and coexistence.

From Assistant AI to Autonomous Agents

Against this backdrop, Anthropic on the 5th unveiled Claude Opus 4.6, a model supporting a one-million-token context window, emphasizing autonomous execution capabilities. The token context window refers to the scope within which tasks such as codebase analysis, document tracking, and complex reasoning can be handled within a single session. Anthropic described Opus 4.6 as a model that operates “longer, more reliably, and more autonomously,” underscoring its ability to plan and execute extended tasks as a clear departure from earlier generations of generative AI. The announcement was widely interpreted as formal recognition that AI is moving into the role of an operational actor in the workplace.

Anthropic also said the model achieved the highest score on Terminal-Bench 2.0, an agent coding benchmark, and outperformed leading competitors on Humanity’s Last Exam, a global academic evaluation. On GDPval-AA, which measures economically valuable knowledge work, Claude Opus 4.6 scored 144 points higher than GPT-5.2 and 190 points above its predecessor, Claude Opus 4.5. The company presented these metrics to highlight the model’s competitiveness in real-world productivity benchmarks.

The nature of agents themselves is evolving. Early generative AI systems functioned primarily as assistants responding to prompts, whereas the latest models resemble actors that pursue goals by leveraging tools. They are described as capable of conducting code reviews and debugging independently, generating and editing documents, spreadsheets, and presentations, and multitasking within collaborative environments. The application programming interface (API) now includes a compaction feature that summarizes context for long-running tasks, along with adaptive thinking and effort control options that adjust the depth of reasoning.

The expansion of autonomy, however, introduces questions of control and accountability. AI agents’ defining characteristic lies in their interaction with real systems, including log monitoring, threat intelligence searches, and equipment control. This elevates the importance of human-in-the-loop structures, in which authority is delegated in stages subject to human verification. Without sufficient explainability regarding the basis for specific decisions, applications in high-risk fields such as finance, security, and healthcare are likely to face constraints. In that sense, the advent of the one-million-token model is serving as a catalyst for resetting industry standards on whether AI should be accepted as an executing entity in core operations.

From Subscription Software to AI-Driven Platforms

As these developments accumulate, forecasts are gaining traction that the traditional software model based on selling program licenses will face inevitable restructuring. When Anthropic introduced a Legal plugin to its Claude model ahead of the Opus 4.6 launch, concerns of a so-called “SaaSpocalypse” spread rapidly across the software-as-a-service (SaaS) industry. Designed to automate contract analysis, clause comparison, risk identification, and revision proposals, the plugin, when paired with collaborative agents, can execute an entire workflow—from contract receipt and automated review to drafting risk amendments, comparing counterpart versions, composing summary emails, generating internal approval documents, and storing files in document management systems.

On the 4th, when news of the development broke, approximately $285 billion in market capitalization evaporated from the New York stock market in a single day, with much of the decline concentrated in traditional SaaS firms such as Salesforce, ServiceNow, Microsoft, and Workday. Anthropic responded unusually by acknowledging concerns, stating that it continues to use traditional SaaS products including Workday, Salesforce, and NetSuite, and emphasizing that AI agents are not intended to eliminate SaaS but to integrate and orchestrate functions at a higher layer.

Traditional SaaS providers have begun adjusting pricing models as part of broader restructuring efforts. Salesforce has introduced a fixed-fee, unlimited-use model for its Agentforce platform, moving away from the conventional per-seat license structure. Microsoft, while maintaining monthly per-user subscriptions, is advancing usage-based billing services. The rationale reflects a calculation that as AI automation enables the same work to be performed by fewer employees, seat-based models will face unavoidable revenue pressure.

Oracle’s recent move further illustrates the direction of SaaS transformation. On the 10th, Oracle announced that it would embed AI agents by default across the entire supply chain spectrum, including planning, procurement, manufacturing, maintenance, and logistics. The agents will be integrated into Oracle Fusion Cloud Applications and run on Oracle Cloud Infrastructure (OCI). Oracle stated that as supply chains grow more complex and market disruptions more frequent, faster and more automated operational approaches are required. The development signals that SaaS is evolving beyond selling standalone products toward becoming an AI-driven execution platform.