Frontier AI models get cheaper, smarter - what it means for business.

In the past two days, AI’s leading firms have engaged in an unprecedented burst of innovation, with new foundation models released almost back-to-back. On September 22, OpenAI introduced GPT‑6 Sol and GPT‑6 Luna – two scaled-down versions of its latest flagship GPT‑6 Astra model – on the same day that Anthropic launched its own frontier model, Claude Opus 5.5. Notably, Anthropic’s announcement came just 90 minutes before OpenAI’s, illustrating how tightly synchronized and competitive the AI R&D landscape has become .

These near-simultaneous launches highlight the breakneck pace of the AI capability race. OpenAI had already hailed GPT‑6 Astra, released earlier in the month, as the world’s most powerful model for tasks like coding and general computer work ([1]). Now, with Sol and Luna, OpenAI is rapidly expanding the GPT‑6 family to make cutting-edge AI more accessible across different tiers of power and cost. Not to be outdone, Anthropic’s Opus 5.5 rollout – coming just two months after its previous Claude flagship in late July – shows that even companies preaching caution are accelerating their timelines to keep up.

For enterprise leaders, this flurry of top-tier model releases in such a short span is a clear signal that the capability frontier is advancing faster than ever. What was state-of-the-art can now shift within weeks, not years. Businesses will need to track new developments continuously and be prepared to rapidly evaluate and integrate emerging AI capabilities in order to stay ahead.

A striking theme in this week’s releases is the dramatic reduction in cost-per-performance. OpenAI’s GPT‑6 Sol and Luna models are priced roughly 50% lower per token than their previous-generation counterparts ([1]). According to OpenAI, improvements in efficiency – such as better model caching and faster inference – allow the company to pass these savings directly on to users via sharply reduced API prices ([2]). For example, GPT‑6 Sol’s output tokens now cost about $10 per million, down from $20 per million in the GPT‑5.6 era ([3]). This kind of cost drop greatly lowers the barrier for businesses to harness advanced AI at scale.

Anthropic has similarly cut costs with its latest model. Claude Opus 5.5’s output tokens are now billed at $20 per million, compared to $25 for its predecessor ([4]). On top of that, the new model is so much more efficient that Anthropic estimates it can run about 40% cheaper for typical workloads while matching the performance of the larger Claude Fable 5.1 on most tasks ([5]). The model is also faster to operate, reflecting an overall reduction in compute required per query ([6]). For enterprise users, these developments mean cutting-edge AI is becoming far more economical. Tasks that once were cost-prohibitive – from processing millions of support tickets to running complex simulations – can now be automated or augmented with AI at a fraction of the previous cost.

The latest generation of foundation models is not only cheaper, but also substantially more capable. Early testers report that Claude Opus 5.5 was able to migrate a 680,000-line codebase in under a day ([1]) – a software project that would have taken an entire team of engineers weeks to complete. In Anthropic’s internal benchmarks, Opus 5.5 has even outperformed its own larger, earlier Claude Fable model on complex coding and knowledge tasks ([2]). Meanwhile, OpenAI’s GPT‑6 Sol exhibits greatly improved accuracy, producing roughly 50% fewer errors than its predecessor in internal testing ([3]). In practical terms, this means AI can tackle more ambitious projects – from massive code refactoring to sophisticated data analysis – with greater speed and correctness than ever before.

AI systems are also getting better at handling extended, multi-step processes autonomously. OpenAI has enhanced GPT‑6’s ability to sustain lengthy interactions by improving prompt caching – enabling the model to reuse information from earlier in a conversation and maintain context over long sessions ([4]). Likewise, xAI’s Grok 4.7 was trained specifically for prolonged, complex tasks and even to check its own work for mistakes during execution ([5]). This evolution toward more resilient, agentic AI will allow businesses to entrust lengthy tasks (from multi-hour research projects to end-to-end customer service resolutions) to AI co-workers that can stay on track without constant human intervention.

While proprietary models continue to set the absolute performance records, open-source AI initiatives are rapidly narrowing the gap in certain domains. Mistral AI – often cited as Europe’s leading AI lab ([1]) – has been releasing open models at a steady clip, providing alternatives that emphasize control and cost efficiency for users. Its top open model (a 14B-parameter variant of the Ministral 3 family) is now among the only non-proprietary systems considered frontier-competitive, though it still trails the best closed models on overall benchmark scores ([2]).

Open models are also demonstrating breakthrough capabilities that signal the next wave of enterprise AI applications. One notable advance is in formal verification – using AI to prove complex mathematical theorems or verify software code. Mistral’s open-source Leanstral 1.5 model, specialized for the Lean 4 proof language, achieved an unprecedented 87% success rate on a challenging theorem-proving benchmark (FATE-H) ([3]). Remarkably, it even outperformed an earlier generation of Anthropic’s Claude (Claude Opus 4.6) on these formal logic tasks at only one-seventh the cost ([4]). By releasing such models under a permissive license, open AI labs are enabling organizations to deploy cutting-edge capabilities on their own infrastructure and adapt them to specific needs – something not possible with the closed-model services.

From a strategic perspective, the open vs. closed-source dynamic in AI is no longer about one approach winning outright, but about leveraging each where it excels. Closed-source providers still push the boundaries of raw capability, but their models (like Claude Opus 5.5) are available only as managed cloud services, without access to underlying code ([5]). Open-source models may not yet match the absolute frontier, but they offer greater flexibility, cost savings, and freedom from vendor lock-in. Many enterprises may ultimately pursue a hybrid strategy – tapping proprietary models for the most complex, value-sensitive tasks, while deploying open-source systems when data privacy, customization, and cost efficiency are key.

The breakneck evolution of foundation models shows no signs of slowing. We can expect the major AI labs to continue one-upping each other with even larger and more versatile models, expanding into more modalities and pushing reasoning and context length to new limits. Google DeepMind, for instance, has yet to answer this week’s announcements – its anticipated Gemini Ultra could soon set fresh benchmarks in advanced reasoning and multimodal intelligence. And if the current pace holds, OpenAI’s GPT-7 or Anthropic’s Claude 6 may emerge on a faster timeline than anyone previously imagined. For businesses, the envelope of what’s possible with AI will keep expanding, and use cases that once seemed futuristic may quickly become viable.

At the same time, industry leaders are raising flags about safety and governance as AI systems grow more powerful. Dario Amodei, CEO of Anthropic, has publicly urged that the industry slow down the rate of frontier AI progress until risk management can catch up . The newest models are launching with expanded safeguards: for example, Claude Opus 5.5 keeps a hidden log of its reasoning that users cannot alter, and its outputs now carry invisible watermarks to meet upcoming EU AI Act requirements . We anticipate increasing regulatory oversight in the coming year, meaning enterprises will need to ensure their own AI deployments – and those of their vendors – comply with evolving laws and standards.

The key for C-level executives is to stay proactive. The organizations that thrive in this fast-moving environment will be those that invest in dedicated AI leadership and pilot programs to continually evaluate new models and integrate them where they add value. By planning for shorter innovation cycles and maintaining flexibility to leverage both proprietary and open-source advances, forward-looking companies can turn the rapidly moving capability frontier into a strategic advantage while managing the associated risks.