OpenAI's new Astra model family represents a fundamental shift in how frontier AI labs are approaching artificial intelligence: instead of racing for faster answers, they're building agents that work for hours, days, or even weeks on complex problems. CEO Sam Altman demonstrated the model to policymakers in Washington this week, showcasing its ability to have multiple AI agents collaborate over extended periods to solve particularly difficult problems. The catch is that longer-running agents come with proportionally longer bills.

What Makes Astra Different From Previous AI Models?

Astra sits alongside OpenAI's other recent model families, Sol, Terra, and Luna, as a new class designed specifically for long-horizon tasks. The company has not yet settled on whether to brand it as GPT-6 or as another entry in the GPT-5 series. What distinguishes Astra is its ability to sustain multi-agent workflows over extended timeframes, a capability that represents a departure from the single-turn, rapid-response models that dominated AI development over the past two years.

To demonstrate Astra's capabilities, OpenAI published results showing that an internal version of the model solved 10 mathematical and theoretical computer science problems that had seen no progress on their main results for at least a decade, and in most cases far longer. These problems span sphere packing, group theory, and lattice cryptography, areas where human researchers have been stuck for years.

How Much Does It Cost to Solve a Decade-Old Math Problem?

Here is where the economics of long-running AI agents become visible. OpenAI disclosed that the total number of tokens, or computational units, needed to find solutions to these 10 problems would cost roughly $2,000 at Sol API rates. That figure is both expensive and a bargain simultaneously, depending on how you measure it. For the cost of hiring a senior technical researcher for a few days, an AI model solved problems that have stumped the field for years. Yet it also underscores a critical reality: every hour an AI agent runs is metered compute that must be paid for, and those costs add up quickly.

The decision by OpenAI to publish the token cost alongside the research result is itself noteworthy. Frontier labs volunteering what their computations cost is a new practice, transforming research results from abstract benchmarks into invoices that can be judged against relative value. This transparency signals that the industry is entering a phase where the economics of AI inference, not just training, matter deeply.

How Are Other AI Labs Approaching Long-Running Agents?

Anthropic is pursuing the same direction from a different angle. Boris Cherny, who leads Anthropic's Claude Code initiative, described the company's Opus 5 model as capable of running for very long periods, sustaining work for days, weeks, or even months at a time. He cited a Swift code rewriting project that had been running for two weeks, spawning thousands and then tens of thousands of agents underneath it.

The acceleration in agent endurance is striking. In a written Q&A published in May, Cherny noted that a year and a half ago, Claude Code could run for maybe 30 seconds before degrading in quality. Now, hundreds or sometimes thousands of agents run for 5, 10, or 20 hours every night. That progression from 30 seconds to 20 hours in 18 months illustrates the curve that both leading labs are extrapolating from as they scale long-horizon AI capabilities.

What Are the Hidden Costs of Long-Running AI Agents?

The expansion of agent runtime introduces a cost structure fundamentally different from traditional software. In classic software, a company builds a product once, and the marginal cost of serving the next customer rounds to nearly zero. That economics model created the high-margin software businesses of the past three decades. AI is different. Every additional hour an agent runs requires additional compute, and that cost is metered in real time.

The problem becomes more complex when you consider how modern AI agents actually work. Each sub-agent running reinforcement learning, reasoning, or verification loops consumes millions of compute cycles. Every verification loop is a second pass over work that was already paid for once. Reinforcement learning improved the models' ability to handle long tasks by teaching them to try, check, and try again, but that capability gain comes with a corresponding cost structure.

The variable cost of AI compute is often compared to electricity when it began transforming the world. But there is a critical difference: electricity does not vary in price or quality at the end user level. The electricity running your toaster is identical to the current running your computer. AI compute tokens, by contrast, are intricately mapped by AI engineers and businesses. Every model comes in dozens of setting combinations, each at different prices and rate limits on subscription plans. Users must watch these combinations closely to avoid unexpected bills.

Steps to Manage AI Agent Costs in Your Organization

• Set Usage Limits: Establish hard caps on token consumption per employee or project before deploying AI agents, as one enterprise client discovered after spending half a billion dollars in a single month on Claude licenses without usage restrictions.
• Monitor Model Selection: Track which models and settings your team uses, since different configurations carry vastly different price tags and performance characteristics, requiring active management rather than set-and-forget deployment.
• Plan for Verification Costs: Budget for the fact that self-verifying agents will run multiple passes over the same work, effectively doubling or tripling the token cost compared to single-pass inference.

Is the Industry Ready for This Shift?

The broader AI industry is grappling with the implications of metered, long-running compute. Microsoft cancelled most of its Claude Code licenses in part over cost concerns. Uber's Chief Operating Officer stated that AI costs were becoming harder to justify to the business. Even Apple, in its June announcement at WWDC 2026, acknowledged the reality by placing daily usage limits on some Apple Intelligence and Siri AI features that rely on server-side models, with increased access available through iCloud+ subscriptions.

The pattern is consistent across vendors: flat-rate, all-you-can-eat pricing was a promise made before anyone understood what usage would actually look like. Now that AI agents can run unattended for weeks, that model is breaking down. The metering has begun, quietly, during the summer of 2026, while much of the industry's attention was focused on earnings reports and regulatory developments.

What remains unclear is whether consumers and enterprises will accept the shift from flat-rate to metered AI pricing, or whether the industry will find ways to bundle and cap costs to maintain the simplicity that drove early adoption. For now, the meter is running, and every hour an AI agent works is an invoice waiting to be paid.