Open Models Capture 71.5% Token Share as Enterprise AI Competition Intensifies with New US Releases

Open Models Dominate Gateway Usage as Spending Share Quadruples

Open weight AI models achieved a decisive milestone in September, processing 71.5% of tokens through Vercel's AI Gateway while capturing nearly one in four gateway dollars spent2

2

. This represents a dramatic shift from June, when open models accounted for just one in twenty gateway dollars. The spending share more than quadrupled over three months, signaling that cost efficiency and token efficiency are driving enterprise AI adoption decisions2

2

. Gateway spending in September reached 13x the December 2025 level, while the average price per token dropped 47%, demonstrating how competition among AI models is reshaping market economics2

2

. Almost one in three teams changed their primary AI provider in September, suggesting the AI model market is limiting vendor lock-in2

2

.

US Labs Launch Enterprise-Ready Open Models to Challenge Chinese Dominance

Reflection AI released Beam, a 501-billion-parameter open model, positioning it as a direct competitor to Chinese open models from DeepSeek, Qwen, and Z.ai1

1

3

3

. The model launched with over 40 partners including AWS, Google Cloud, Microsoft, and Oracle, with Dell and HPE delivering it for on-premises and air-gapped environments4

4

. Nvidia released Nemotron 3.5 Super VL on October 15, a 120-billion-parameter model that reads documents, charts, and web pages, running five times faster than comparable open models4

4

. Spend management FinTech Ramp is tuning it for financial document analysis, while Scale AI trained a version for merger due diligence work4

4

. These releases represent US and European labs playing catch-up to establish viable options for enterprise use cases and sovereign AI deployments1

1

.

Token Efficiency Emerges as Key Differentiator Over Per-Token Pricing

Reflection AI is betting that Beam can finish tasks in far fewer tokens than other models its size, which lowers the cost per completed task rather than competing solely on per-token pricing3

3

. The company trained the model using intensive reinforcement learning over four weeks, running more than 100 million rollouts powered by 10,500 Nvidia GB300 GPUs3

3

. During training, Reflection penalized the model whenever it used unnecessary tokens, though Chief Technology Officer Ioannis Antonoglou explained the challenge: teams that apply penalties naively end up with models that simply avoid solving tasks3

3

. The computing fleet peaked at 883,006 virtual CPUs across 23 clusters on Google Cloud and Amazon Web Services3

3

. Artificial Analysis noted that early indicators suggest Beam will be "one of the most token-efficient open models we've seen for its level of intelligence"3

3

.

Chinese Models Capture Significant Gateway Spending Despite US Competition

Chinese AI models maintained strong positions in September gateway spending, with DeepSeek V4.1 Flash ranking third at 4.19% of spending2

2

. Moonshot AI's Kimi captured 3.6% of gateway tokens and 13.2% of spend, while DeepSeek took 58% of tokens for just 6.6% of spend, and Z.ai's GLM-5.3 took 7.2% of tokens and 4.5% of spend2

2

. More than two-thirds of September spending on Chinese models went to US providers serving those models, up from less than one-tenth in December 20252

2

. DeepSeek updated its Flash model on October 17, with the Chinese lab claiming higher benchmark scores, faster task completion, and shorter answers that lower cost per job4

4

. MiniMax launched M3.1 Flash on October 19, running nearly three times faster than its predecessor while cutting usage costs by 70%, with pricing at $0.15 per million input tokens and $0.60 per million output tokens4

4

.

Geopolitical Tensions Threaten Cross-Border AI Model Access

China proposed adding 40 technology items to its export control list on Friday, with 23 related to AI, including models like Moonshot's Kimi K3 and Alibaba's Qwen 3.8-Max that meet parameter thresholds set by Chinese authorities2

2

. Chris McGuire, senior fellow for China and emerging technologies at the Council on Foreign Relations, stated this "introduces significant uncertainty for US businesses that host or rely on Chinese models"2

2

. Given that Chinese AI labs are moving toward licensing open-source models to generate revenue, the Chinese government's ability to curtail licensing agreements should make US businesses more hesitant to rely on those models2

2

. Efforts to ban or limit open weight models considered by the current US administration would affect American businesses serving those models and their customers2

2

. For banks and insurers that cannot use Chinese models for security reasons, the choice of what to run in-house through on-premises deployment is expanding4

4

.

Platform Capabilities Trump Individual Model Performance in Enterprise Decisions

While model proliferation creates flexibility, the enduring question is whether an organization's AI platform, governance framework, agent architecture, existing operations, and desired sovereignty posture can accept a model swap1

1

. Enterprises are not purchasing benchmarking scores but seeking models that fulfill governance, security, orchestration, agents, developer tooling, evaluation capabilities, observability, sovereignty, and integration needs1

1

. This advantage swings to frontier model providers like Anthropic and OpenAI, which captured 11.03% and 4.86% of September spending respectively with Claude Opus 5.5 and GPT-6 Astra2

2

. These providers combine models with platforms delivering orchestration, agents, governance, security, observability, and development tooling1

1

. The open-source community is expected to follow a similar path, building open agent frameworks, open orchestration layers, and eventually open AI platforms that help organizations operationalize capabilities1

1

. Future enterprise AI strategies will use a mix of frontier AI and open weight models as complementary components, with open models providing more choice, potentially lower costs, and more regional and sovereign deployment options1

1

.