GPT-5: How Two Simple Settings Tripled AI Performance — What This Means for Your Business

The race for artificial intelligence performance doesn't always play out on the battlefield of the models themselves. Sometimes, it's two simple technical tweaks that change everything. OpenAI just demonstrated this spectacularly: by activating just two parameters in its API, GPT-5's scores on the ARC-AGI-3 benchmark — one of the most demanding evaluations for general reasoning — have been tripled. For French companies investing in AI, this signal is critical: the performance of your AI tools depends as much on how you configure them as it does on the model itself.

ARC-AGI-3: Why This Benchmark Is a Thermometer for Real Reasoning

The ARC-AGI-3 benchmark (Abstraction and Reasoning Corpus) is designed to measure AI's ability to generalize, reason by analogy, and solve novel problems — exactly as a human would when faced with an unprecedented situation. Unlike traditional benchmarks that evaluate knowledge memorization, ARC-AGI-3 tests the cognitive flexibility of AI.

This is not an academic detail. For a business, a model that excels on ARC-AGI-3 is a model capable of:

• Analyzing ambiguous situations with no precedent in its training data
• Adapting its reasoning to specific, even complex business contexts
• Reducing logical errors in critical tasks such as financial analysis, legal review, or strategic planning

The fact that two simple API parameters — intermediate reasoning retention and exchange compaction — were enough to triple these scores illustrates a fundamental principle: an under-parameterized model is an under-performing model, regardless of its intrinsic level.

The Two Key Parameters: What They Do in Practice

OpenAI has identified two decisive levers in configuring its API:

1. Reasoning Retention This parameter allows the model to maintain a trace of its intermediate reasoning between the steps of a complex response. Rather than "starting from scratch" at each response segment, GPT-5 maintains logical coherence throughout its processing. Concretely, in a business setting, this translates into more reliable analyses on complex cases: a due diligence report, a multi-source regulatory synthesis, or a multi-criteria strategic recommendation gains in coherence and depth.

2. Context Compaction This second parameter optimizes how the model manages its context window, intelligently compressing past information to free up processing capacity for new data. Result: less information loss in long conversations or large documents, and better economic efficiency (fewer tokens consumed for superior results).

These two mechanisms, combined, allow the model to "think further" without losing the thread — a valuable capability in the most demanding business use cases.

Concrete Applications for French Businesses

This technical advance opens immediate opportunities for several sectors:

Financial Services and Audit Audit firms and management control teams handle long, heterogeneous, and interdependent documents. With reasoning retention activated, an AI assistant can analyze a 200-page annual report, maintain coherence between sections, and produce a synthesis without internal contradiction — where a poorly configured model would "forget" the initial analyses during processing.

Legal and Compliance Legal teams at large French companies and law firms face complex contract analyses, often in chains (master agreement, amendments, appendices). Context compaction enables the processing of an entire document corpus without degradation of analysis quality, even over prolonged exchanges.

Manufacturing and R&D In design offices where engineers work on dense technical specifications, these parameters allow AI to assist design while maintaining a systemic vision of the project, avoiding inconsistencies between modules or specifications.

Customer Relations and Complex Support For B2B customer service centers managing multi-interaction cases, reasoning retention ensures that AI doesn't start from a blank slate with each exchange, offering a smoother customer experience and faster resolutions.

In all these cases, the gain is not marginal: it can represent a significant reduction in processing time, an improvement in output quality, and ultimately a measurable ROI on AI investments.

Training Your Teams on AI Configuration: The New Strategic Challenge

The most important lesson from this OpenAI discovery may not be technical — it's organizational. If two parameters can triple a model's performance, how many French companies today are letting their AI tools run with sub-optimal default settings, thus losing a considerable fraction of their potential?

This reality points to an urgent need for upskilling teams in what is called "advanced prompt engineering" and fine-tuning of AI APIs. It's no longer just about learning to write good instructions (prompts), but about understanding:

• Model configuration parameters (temperature, context window, retention, compaction)
• API call architectures suited to business use cases
• Cost/performance optimization strategies to maximize ROI
• Continuous evaluation of AI performance in a given context

These skills are no longer the sole domain of IT teams. Business managers, project leads, and managers piloting AI initiatives must now have sufficient technical knowledge to ask the right questions and demand proper configurations from their providers and internal teams.

It is precisely to address this need that Ikasia has developed its training and consulting programs in AI applied to French businesses. Whether it's introducing your teams to the fundamentals of generative AI, training your developers in advanced API integration, or guiding your leaders in defining a coherent AI strategy, Ikasia offers you customized programs rooted in French market realities.

Is your company truly getting the most out of its AI tools? Don't let default settings limit your competitiveness. Contact Ikasia experts at ikasia.ai for an audit of your AI usage and personalized support. Because in the race for performance, technical details often make all the difference.