Million-Token Context and Long-Horizon Agents: From Model Capabilities to Engineering Implementation

1. Background: Why Long Context is a Watershed

Over the past two years, the competitive focus of large language models has gradually shifted from “parameter scale” to “usable context length” and “task duration.” While parameter scale determines the upper limit of a model’s knowledge capacity, context length dictates how much information it can “see” simultaneously in a single session. When context expands from tens of thousands to millions of tokens, the boundaries of model capabilities undergo a qualitative transformation:

• From fragmentary to global understanding: Real-world materials like code repositories, legal documents, and financial annual reports are inherently lengthy. Only with million-token context can models perform global reasoning without losing key details.
• From single-turn Q&A to long-horizon tasks: Agent tasks often require dozens or even hundreds of tool calls, with each step depending on prior states. Context length directly determines how far an Agent can go without “amnesia.”
• From passive response to proactive planning: Long context allows models to maintain task lists, progress states, and intermediate conclusions internally, thereby supporting more complex planning and self-correction.

Around this trend, the industry has successively launched flagship models supporting million-token context, while simultaneously enhancing specific capabilities in coding and Agent directions. This article will dissect the capability composition, evaluation methods, safety governance, and cost structure of such models from a technical perspective, providing a reference for developers and enterprises in model selection and implementation.

2. Capability Composition: Three Main Pillars of Code, Agents, and Expertise

2.1 Code and Long-Horizon Agents are Upgrade Priorities

Capability upgrades in flagship models typically focus on two directions: code generation/understanding and long-horizon Agent task execution. These two are highly coupled—Agents need to call tools via code, parse results, and revise strategies, while coding tasks often require multi-step iteration and external execution feedback.

In terms of evaluation systems, the industry generally adopts two types of benchmarks:

• Realistic Terminal-Bounded Tasks: In a simulated real operating system terminal environment, the model’s ability to complete file management, environment configuration, and script debugging via command line is assessed. This type of benchmark emphasizes “executability” rather than “describability,” placing extremely high demands on the model’s tool call accuracy and error recovery capabilities.
• Long-Horizon Code Tasks: Evaluates the model’s stability in cross-file, cross-module, and multi-round modification scenarios. Typical metrics include long-horizon task scores and regression rates, where the former measures completion and the latter measures whether modifications introduce new issues.

Taking public industry data as an example, a certain flagship model scored approximately 65.7% on a terminal-bounded task benchmark, showing only a slight gap compared to a top-tier closed-source model’s 66.4% in the same period, and significantly surpassing another competitor’s 58.2%. Its internal long-horizon code benchmark score increased from approximately 55.95% in the previous generation to 61.2%, indicating continuous improvement in long-horizon stability.

The engineering value of these capabilities lies in delegating repetitive tasks like “environment setup, dependency installation, script migration, and batch refactoring” to Agents, with humans only needing to review at key checkpoints.

2.2 3D Modeling and Scripting Production Capabilities

Beyond general code, some models demonstrate strong adaptability in the direction of specialized software scripting. Typical scenarios include:

• Game Engine Scripting: Proficiently calling script interfaces of mainstream game engines to complete scene setup, character control, and physics parameter configuration.
• Parametric 3D Modeling: Directly writing and running modeling software scripts to drive geometry generation via parameters, enabling reusable modeling workflows.
• CAD Scripting: Completing dimension constraints, assembly relationships, and engineering drawing outputs via scripts in computer-aided design scenarios.

On a certain modeling script benchmark, one model achieved a score of 92.7% leveraging tool calls, closely approaching top-tier closed-source models’ 96.2% and 95.9%. The significance of this capability lies in transforming “manual modeling” into “code-based modeling,” making 3D assets capable of version control, parametric reuse, and automated batch generation.

2.3 Financial and Expert Knowledge Workflows

In the financial sector, model capabilities are shifting from “Q&A” to “end-to-end delivery.” Typical workflows include:

1. Financial Modeling: Generating the three financial statements and their linkages based on historical data and assumptions.
2. Investment Analysis: Conducting valuations, sensitivity analysis, and risk alerts for target entities.
3. Industry Research Reports: Integrating public information to output structured research report drafts and annotating data sources.

On an external office document benchmark, one model achieved a score of 64.4, officially claimed as the current best. For developers, this means an Agent can be used to chain the entire pipeline of “data acquisition — cleaning — modeling — reporting,” freeing analysts from repetitive labor.

3. Factual Accuracy and Hallucination Governance

Long context and long-horizon tasks amplify the risk of hallucinations: the longer the context, the easier it is for the model to “lose information” or “fabricate details” in the middle paragraphs. The industry typically governs this from two levels:

• Training Side: Introducing factuality rewards during mid-training and reinforcement learning phases, penalizing verifiable false statements.
• Evaluation Side: Constructing targeted stress test sets that deliberately focus on error-prone scenarios, such as fine-grained statements regarding time, quantity, and causality.

An internal stress test showed that the error rate for single factual statements dropped from 23.4% in the previous generation to 8.3%, and the proportion of responses containing at least one factual error was approximately 42.4% of the previous generation. It must be emphasized that such tests deliberately select high-difficulty samples, and their error rate does not represent the overall level of daily usage. In production environments, developers should still combine retrieval-augmented generation (RAG) and manual spot checks, rather than treating model outputs directly as sources of fact.

4. Safety Governance: Data Filtering and Capability Boundaries

Models equipped with code and terminal operation capabilities inherently possess a double-edged attribute. The industry’s prevailing governance approach includes:

• Training Data Filtering: Identifying and removing network attack and defense related data during mid-training and reinforcement learning phases to reduce abuse risks.
• Capability Boundary Assessment: Comparing the problem-solving rates of the “unfiltered version” and the “released version” on a proprietary safety benchmark. The former’s pass@3 can solve 90.5% of problems, while the released version solves 81.0%, indicating that filtering does sacrifice some capability in exchange for a more controllable risk surface.

The implication for enterprises is that when introducing long-horizon Agents, they should simultaneously build permission isolation, operation auditing, and sandbox execution environments to prevent Agents from obtaining excessive system privileges.

5. Cost Structure: The Engineering Economics of Long Context

Million-token context does not mean “blindly filling it up.” The context caching mechanism is key to controlling costs:

Model Tier	Input (per million tokens)	Output (per million tokens)	Cache Hit	Cache Write (5 minutes)	Cache Write (1 hour)
Standard	¥13.00	¥68.00	¥1.30	¥13.00	¥26.00
High-Speed	¥26.00	¥136.00	¥1.30	¥26.00	¥52.00

Three practical principles can be derived from the table:

1. Cache hits are the biggest leverage: The hit price is only one-tenth of the input price, suitable for scenarios involving “repeatedly asking questions about the same long document.”
2. Cache writes have tiers: The prices for the 5-minute tier and the 1-hour tier differ. Short-lifecycle sessions should prioritize the short tier.
3. High-Speed tier has a 2x premium: Use this only for latency-sensitive scenarios; standard batch processing tasks are fine with the Standard tier.

Furthermore, compared to the previous generation, overall prices have decreased, indicating that the marginal cost of long-context inference is declining.

6. Implementation Advice for Developers

• Select models based on task length: Use lightweight models for short Q&As, and switch to million-token context flagship models for long documents and long-horizon Agents.
• Replace repeated input with caching: Write system prompts, code repository summaries, and knowledge base fragments into the cache to significantly reduce costs.
• Set “checkpoints” for Agents: Periodically persist intermediate states during long-horizon tasks to prevent unrecoverable context overflow.
• Align evaluations with real environments: Prioritize executable benchmarks like terminal-bounded tasks and long-horizon code, rather than just looking at multiple-choice scores.
• Safety first: The trio of sandboxing, least privilege, and operation logs should be in place before the Agent goes live.

Conclusion

Million-token context and long-horizon Agents are pushing large models from “conversational tools” to “executable workflow engines.” For developers, the real threshold is not calling an API once, but understanding context costs, caching strategies, evaluation methods, and safety boundaries, and using engineering practices to stably embed these capabilities into business processes. Whoever successfully runs through this methodology first will gain the upper hand in the next phase of the intelligence competition.