OpenAI and Anthropic Just Made AI Agents Dramatically Cheaper — What It Means for Enterprise Buyers

On 22 September 2026, Anthropic and OpenAI released new models within hours of each other. Neither release claimed a breakthrough in reasoning. Both claimed the same thing: meaningful capability gains at substantially lower cost. For enterprises running agentic AI at scale, that is arguably the more important development.

Anthropic launched Claude Opus 5.5, an upgrade to its flagship workhorse model. OpenAI launched GPT-6 Sol and Luna, the latest mid-tier and entry-level models in its GPT-6 family. The timing was notable: both companies had recently joined calls to slow the pace of frontier AI development. Their response was not to pause, but to compete on price.

Anthropic Opus 5.5: the flagship gets cheaper and faster

Opus 5.5 is positioned as Anthropic's primary model for complex knowledge work, coding and multi-step agentic tasks. The headline numbers from Anthropic's announcement are: input tokens at $4 per million, output tokens at $20 per million — 20% below Opus 5 pricing. Cache reads, which make up the majority of agentic and coding workloads, dropped 60% to $0.20 per million tokens. Output generation is more than 30% faster.

Anthropic also claims that typical workloads will see total cost savings closer to 40% because Opus 5.5 uses fewer tokens to complete the same tasks. In benchmark terms, Anthropic and its partners report Opus 5.5 trading results with OpenAI's GPT-6 Astra on coding and knowledge tasks. The point is not that one model is definitively better. The point is that the gap is narrow enough for price to become a deciding factor.

For enterprises, this matters because Opus-class models are increasingly the engine behind autonomous agents in production. Coding agents, research agents, data-analysis agents and customer-service agents all run on sustained inference. When the per-token cost of sustained inference falls by 40%, the business case for deploying those agents at scale improves in direct proportion.

OpenAI GPT-6 Sol and Luna: efficiency as a strategy

While Opus 5.5 sits at the high end of the market, OpenAI targeted the middle and lower tiers. GPT-6 Sol, positioned as the capable daily-driver model, costs $2 per million input tokens and $10 per million output tokens. GPT-6 Luna, the fast and lightweight option, costs $0.10 and $0.50 respectively. Both were trained with methods similar to those used for GPT-6 Astra, and OpenAI says they are typically a few percentage points more capable than their predecessors while costing roughly half as much to run.

OpenAI's naming convention is worth understanding because it maps directly onto procurement decisions. Astra is the heavy-duty model for research, advanced coding and frontier tasks. Sol is the balanced workhorse. Terra is the general-use model. Luna is the fast, cheap option for high-volume, low-latency applications. Most enterprise agents do not need Astra-level reasoning. They need Sol-level reliability at Luna-level cost. That is exactly what these releases deliver.

Why both companies are racing to cut prices

The immediate driver is competition from open-weight models. Enterprises are increasingly using model routers to send simpler tasks to cheaper alternatives and reserving frontier models for tasks that genuinely need frontier capability. If Anthropic and OpenAI do not match that economics, they lose volume. Volume, in this market, is what funds the next generation of models.

There is a deeper signal here. The discourse on AI capability has become polarised. One camp claims that frontier models can one-shot complex creative and technical tasks. The other camp points to security incidents, alignment failures and the gap between demo and production. The reality in most enterprise environments is more prosaic: the models are good enough for a wide range of useful work, but the cost of running them at scale has been prohibitive. These price cuts directly address that.

What this means for agentic AI buyers

Three practical implications stand out for enterprises evaluating or deploying agentic AI.

First, model costs are now a competitive variable. Two years ago, the only question was whether a model could do the job. Now, procurement teams can credibly ask which model can do the job cheapest. That shifts negotiation power toward buyers and puts pressure on agent vendors to justify the model choices embedded in their platforms.

Second, model routing is now a core architecture decision. Enterprises running multiple agents across different functions will need to decide whether to standardise on one provider, route tasks dynamically based on cost and capability, or run domain-specific models for specific workflows. Each approach has governance, integration and cost implications. The wrong choice is expensive to unwind.

Third, the validation window is shortening. When inference costs fall, the threshold for deploying an agent in production drops. Teams that were waiting for board approval to run a pilot can now run that pilot for a fraction of the budget. The consequence is that the window between experiment and production deployment is compressing — and governance needs to keep pace.

The Agentic Expo angle

Cheaper models do not solve the enterprise agent deployment problem. They change its economics. The remaining blockers — integration with legacy systems, data quality, security governance, workforce adoption and vendor selection — are all human and organisational problems, not model problems.

That is why the buyers who will get the most from these price cuts are the ones who already understand their agent use cases, their data architecture and their compliance requirements. The buyers who treat a 40% cost reduction as a signal to accelerate without answering those questions first will discover that cheaper inference does not fix broken workflows.

Agentic Expo exists to close that gap. The exhibitors on the floor are not selling models. They are selling the harnesses, the orchestration layers, the security tools and the integration platforms that turn cheaper models into productive agents. Between now and March 2027, we expect these price cuts to trigger a new wave of enterprise pilots. The buyers who attend with clear procurement criteria will be the ones who convert those pilots into production.

Get Tickets Exhibit at Agentic Expo