New AI Model With Native Computer Use and Million-Token Context Redefines Professional Work

A New Generation of AI Models: Redefining the Boundaries of Professional Work

The AI field has recently seen a major technical upgrade with the official debut of a new generation of large language models. Positioned as “the most powerful and efficient frontier model for complex professional work to date,” its core breakthroughs center on three areas: native computer use, ultra-long context support, and professional-grade office document processing. This is not a simple expansion of parameter scale, but a systemic redesign spanning model architecture and interaction paradigms—marking a critical leap for AI from “conversational assistant” to “autonomous executor.”

Model Lineup and Architecture

The release adopts a tiered model strategy designed to cover scenarios with different complexity and performance requirements. The product lineup spans three variants:

1. Standard model: The general-purpose entry point, balancing everyday task efficiency with fast response times, suitable for the vast majority of routine interactions.
2. Thinking model: This variant strengthens complex reasoning and multi-step task decomposition. For tasks requiring deep logical chains—mathematical proofs, debugging code under multiple constraints, or long business-process planning—the model simulates an internal chain of thought to improve the accuracy and coherence of its final output. The underlying mechanism resembles “System 2 thinking,” performing more thorough internal deliberation before answering.
3. Pro model: Purpose-built for enterprise workloads with high concurrency and high throughput. It maintains low-latency responses while significantly improving API stability and resource utilization, making it suitable for large-scale automation pipelines and real-time interactive systems.

This clearly segmented design reflects the core logic of how large models are commercialized today: rather than pursuing a single all-powerful model, fine-grained version tiers let organizations of different sizes and needs match compute resources and capability levels to what they actually require.

The Engineering Significance of a Million-Token Context

One of the most disruptive metrics in this upgrade is the expansion of the API-side context window to the 1 million-token level—a capacity far beyond the vendor’s earlier models. The engineering implications go well beyond “reading longer articles.”

From an implementation perspective, million-token context depends on efficient attention mechanism optimization and memory management strategies. This long-sequence capability directly unlocks several key application scenarios:

• Understanding very large codebases: Developers can inject the core codebase of an entire mid-sized project into the model in one pass, asking it to trace logic across files, hunt down bugs, or propose architecture refactors—without manually splitting and stitching chunks.
• Long-horizon legal and financial document analysis: For IPO prospectuses, lengthy legal contracts, or complex financial audit reports, the model can perform end-to-end global analysis and identify contradictions and hidden risks between clauses.
• Mixed multi-file processing: Users can upload packages combining text, spreadsheets, and presentations, and the model builds knowledge associations across documents.

This capability fundamentally changes workflow design. Long-document tasks that previously required external vector databases or complex RAG (retrieval-augmented generation) frameworks can now be handled through simple, direct context injection, substantially reducing architectural complexity and maintenance costs.

Native Computer Use: The Paradigm Shift from Conversation to Execution

The biggest highlight of this upgrade is the “Native Computer Use” capability. Traditional AI assistance is typically limited to text generation or code completion, whereas this model offers direct control of the graphical interface (GUI) within its core product form.

On the technical side, the model can parse desktop screenshots (visual signals) and DOM tree structures (semantic signals) simultaneously. This means it no longer depends on APIs exposed by specific software; instead, it operates the system the way a human does—by “looking at the screen” and “understanding interface logic.” In practice:

1. Script-free operation: Users do not need to write glue code or configure elaborate automation scripts. The model directly controls mouse movement, clicks, and keyboard input, switching between applications to carry out tasks.
2. Cross-application workflow orchestration: Take a typical business analysis task: the model can pull raw sales data from a spreadsheet, run its analysis, generate visualizations, insert the charts into a presentation, and add descriptive captions—all in one continuous flow, with no manual data shuttling.
3. Dynamic environment adaptation: Because decisions are made from real-time screenshots and interface structure, the model is more robust to minor layout changes and different application versions; workflows won’t break just because a preset script no longer matches the environment.

This capability is especially valuable for generating “long-horizon deliverables”—building complex financial models, writing large-scale legal analysis reports, or producing high-quality, multi-page presentation materials. The model provides end-to-end intelligent support, continuously tracking and maintaining the deliverable’s overall structural consistency and logical rigor.

Benchmark Results and Professional-Scenario Performance

In third-party and in-house evaluations, the model posted milestone scores across multiple dimensions:

• Desktop and browser operation: On benchmarks simulating real computer use (such as OSWorld Verified and WebArena Verified), the model set new all-time highs, demonstrating strong generalization in understanding complex interface layouts and handling unstructured web elements.
• Knowledge work: On GDPval, an internal test designed around knowledge workers’ daily tasks, the model reached 83% accuracy—approaching the entry level of human experts in the field.
• Vertical-domain evaluations: On APEX Agents, a professional test focused on legal and finance tasks, the model ranked first overall, validating its usefulness in highly specialized domains.

Also notable is the dedicated evaluation of spreadsheet modeling. In tasks simulating the day-to-day work of a junior investment banking analyst, the model averaged 87.3% (compared with 68.4% for the previous-generation model). In blind human comparisons of presentation aesthetics and layout logic, more than 68% of reviewers preferred the new model’s output, citing clearer visual hierarchy, more professional use of whitespace, and tighter integration of charts and text.

Reliability Engineering: Lower Hallucination, Higher Factuality

To address the industry-wide problem of models “hallucinating with confidence,” this upgrade applied systematic reliability hardening across data filtering, training strategy, and inference-time alignment. Official figures show a 33% reduction in the error probability of individual factual statements and an 18% reduction in the probability of errors in overall answers compared with the previous-generation model—hallucination rates effectively suppressed.

This reliability gain is critical for high-risk industries such as finance, legal, and healthcare. Enterprises can treat large models not merely as creative generation tools, but as dependable support for initial compliance review, data verification, and fact-checking, lowering the marginal cost of manual review.

Industry Impact and Technical Trends

This model iteration sends several clear signals to the industry:

1. The competitive focus shifts from “chatting ability” to “execution”: Language fluency alone is no longer a technical moat. Whether a model can reliably execute multi-step operations in complex digital environments is becoming the dividing line for the next generation of models.
2. Context length is now a core variable in AI workflow design: Longer context means less information loss and fewer interruptions, providing the foundation for genuine long-horizon autonomous delivery.
3. Agents are accelerating from concept to production: Native computer use closes the last mile of agent-environment interaction. Future software may gradually shift from “humans operating software” to “humans supervising software that operates itself”—a change that benefits developers and everyday users alike.

Practical Recommendations for Development Teams

For teams planning to adopt these capabilities, a few suggestions are worth considering:

• Restructure task decomposition: When building applications on the API, move beyond the “one-shot Q&A” mindset and design concurrently executable chains of subtasks, leveraging the long context to reduce intermediate state storage.
• Strengthen human-in-the-loop review: Even with substantially improved reliability, keep expert approval checkpoints in place for high-stakes outputs such as financial models or legal opinions.
• Optimize compute costs: The Pro model is powerful, but it may be over-provisioned for lightweight tasks. A routing strategy that dynamically calls different model versions based on task complexity is recommended.

This upgrade is not merely a victory of compute stacking; it reflects an engineering system built on a deep understanding of real workflows. It signals that AI will embed itself more deeply into the core of professional productivity and become a key engine driving digital transformation. For technical decision-makers, this is an ideal window to evaluate and pilot these capabilities within their own business scenarios.