OpenAI’s GPT-6 Astra is emerging as a significant competitor in enterprise artificial intelligence after Signal65’s PINNACLE benchmark showed the model completing 279 of 280 multi-step business tasks at maximum reasoning effort. The result arrives only days after Astra’s launch and places increasing emphasis on a metric that could matter more to corporate buyers than conventional benchmark scores: how reliably an AI agent can complete useful work at an economically sustainable cost.

Astra Posts a Near-Perfect Enterprise Task Completion Rate

Signal65’s PINNACLE benchmark is designed specifically around agentic enterprise work rather than isolated question-and-answer performance. The benchmark uses multi-step workflows that require an AI system to navigate tools, retrieve information, apply rules and produce a complete set of deliverables, with results graded programmatically rather than by another AI model or human evaluator. Signal65 says its initial dataset covered 44 model configurations across 30 base models, with testing performed through August 30.

In the Astra results shown in the attached Signal65 material, the model completed 279 of 280 jobs at maximum reasoning effort. The same material reports a 0.0% fabrication rate on questions that could not be answered from the supplied information. That distinction is particularly important for enterprise deployments, where an agent producing a confident but unsupported answer can create additional verification costs or operational risk.

Signal65’s methodology is designed to test precisely this problem. Its benchmark includes both organized enterprise data and less orderly “as-found” information, allowing models to be evaluated not only on whether they can retrieve information but also on whether they can operate when business data is incomplete, duplicated or inconsistent.

The Cost of a Correct Answer Is Becoming the Critical Metric

The more consequential finding may be the economics. The attached Signal65 result puts Astra’s cost at $1.51 per successful task at maximum reasoning effort, compared with $2.46 for Claude Fable 5.1. That implies approximately 39% lower cost for a completed task under the configurations tested.

This comparison illustrates why enterprise AI economics are increasingly moving beyond published token prices. Signal65 argues that an enterprise does not ultimately purchase tokens; it purchases completed work. Its methodology therefore assigns the cost of failed attempts to successful tasks and accounts for the input, cached input and output associated with the workflow. In its broader PINNACLE analysis, Signal65 found that input tokens repeatedly supplied during agentic workflows can account for 65% to 91% of hosted API costs.

That dynamic changes the competitive equation. A model with a higher nominal token price can still be economically attractive if it completes a task using fewer tokens, requires fewer tool-use cycles or produces fewer failures. Conversely, a cheaper model can become expensive when employees must repeatedly correct its output or restart incomplete workflows.

Reliability Could Reshape Enterprise AI Adoption

The Astra results also come at an important point in OpenAI’s enterprise strategy. OpenAI describes GPT-6 Astra as its most capable model for complex reasoning, coding, computer use, research and document creation. The model is designed to carry multi-step workflows from an initial instruction through to finished documents, spreadsheets and presentations.

OpenAI’s own launch materials also emphasize improved task-boundary adherence and safety monitoring. The company says Astra is designed to better recognize when information is insufficient rather than simply filling gaps with an unsupported answer. It has also identified cybersecurity as a particularly sensitive capability, with Astra reaching the “Critical” threshold under OpenAI’s Preparedness Framework.

The broader implication is that enterprise AI competition is becoming increasingly tied to unit economics and operational reliability. Companies deploying thousands or millions of automated workflows will care less about whether a model wins an abstract reasoning leaderboard and more about how many tasks it completes correctly, how often humans need to intervene and what each successful workflow ultimately costs.

OpenAI’s Advantage Will Depend on Real-World Scale

The Signal65 result is significant, but it should be interpreted within the limits of the benchmark. PINNACLE measures a defined set of enterprise workflows, and results depend on the model configuration, reasoning level, tools, data environment and pricing assumptions used in the test. Signal65 itself emphasizes that benchmark results should be read alongside the methodology rather than treated as universal measures of AI capability.

For investors and corporate technology buyers, the next stage will be determining whether Astra’s efficiency advantage survives outside controlled evaluations. Adoption rates, API utilization, enterprise contract volumes, inference costs and customer retention will provide a more durable measure of commercial impact. If Astra can maintain near-complete task execution while keeping the cost of successful work materially below competing frontier models, enterprise AI could move closer to an automation economics model rather than a conventional software subscription model. That would increase the strategic importance of AI agents across finance, consulting, software development, customer operations and other knowledge-intensive industries.