May 2026 Summaries

In May 2026, the landscape of AI models in terms of performance and pricing shows a significant shift, with open-source models closing the gap with their closed-source counterparts in specific contexts. The Artificial Analysis Intelligence Index reveals that premium closed-source models like GPT-5.5 and Claude Opus 4.7 remain ahead in overall intelligence and reasoning tasks, scoring 60 and 57, respectively. However, open-weight models such as Kimi K2.6 and DeepSeek V4-Pro have made strides, particularly in coding benchmarks, offering competitive performance at a fraction of the cost. Kimi K2.6, for instance, ties with GPT-5.5 on the SWE-Bench Pro for coding tasks but is about six times cheaper. The cost-effectiveness of open-source models is highlighted in tasks like autonomous coding and agentic web development, where they provide near-frontier quality at significantly lower prices, making them an attractive option for teams prioritizing budget without sacrificing too much capability. For high-stakes applications where the utmost reliability and reasoning depth are necessary, the closed-source premium remains defensible. The strategic decision in 2026 hinges on effectively routing tasks to appropriate models, utilizing closed-source models for complex reasoning and open-source models for routine tasks, optimizing both performance and cost at scale.

In 2026, DeepInfra highlighted significant security concerns surrounding OpenClaw, an AI agent runtime that faced vulnerabilities due to its architecture, which allowed attackers to exploit prompt injection and supply chain attacks. Despite patches, the core issue remained unresolved as OpenClaw's design inherently lacked a mechanism to differentiate between trusted and untrusted inputs, making it susceptible to indirect prompt injection attacks. The ClawHavoc campaign exemplified the exploitation of these vulnerabilities through the manipulation of skills available in the ClawHub registry. DeepInfra proposed solutions like the external wrapper NemoClaw, which places security measures outside the agent process, offering a more robust defense against such attacks. By implementing an isolated inference service, leveraging DeepInfra's open-weight models, and employing a layered security approach, organizations can better protect OpenClaw deployments from potential threats.

Mixture of Experts (MoE) models have significantly transformed the economics of large language models (LLMs) by allowing them to be larger yet cheaper to operate compared to traditional dense models. This architectural approach involves using a collection of smaller networks, known as experts, activated selectively for each token via a gating network, thus reducing the compute cost per token while maintaining high total model capacity. MoE models like DeepSeek V4-Pro and Kimi K2.6 can operate economically at trillion-parameter scales because they only activate a small portion of their total parameters per inference. This decoupling of total capacity from per-token compute cost makes them financially viable for API-level serving, though they require substantial memory resources. Consequently, MoE models offer competitive performance at a fraction of the cost of dense models, reshaping API pricing in the AI landscape by enabling more capability per dollar of compute.

DeepInfra's OpenClaw platform offers a cost-effective solution for running autonomous workflows by utilizing open-weight models over closed-weight APIs, significantly reducing expenses associated with high-frequency tasks. The platform categorizes use cases into morning briefings, developer workflows, multi-agent operations, and continuous monitoring, each benefiting from tailored model selection for efficiency. Open-weight models on DeepInfra, such as DeepSeek-V3.2 and Qwen3 family models, optimize performance and cost by allowing per-task model routing and leveraging context caching. This approach not only enhances agent autonomy but also makes continuous operations feasible and affordable, transforming what might be a costly experiment into a sustainable tool. By integrating DeepInfra's models with OpenClaw through a simple configuration, users can adapt workflows without the need for extensive code changes, thus maintaining operational flexibility and scalability.

DeepInfra's approach to optimizing OpenClaw AI API costs involves strategic use of different model tiers and prompt caching to significantly reduce expenses. By understanding the cost drivers of token usage, including system prompts, conversation history, and output, users can implement a two-tier model strategy, utilizing a smart primary model for main tasks and budget models for sub-tasks. This method reduces unnecessary costs by routing requests to the most economical models capable of completing the tasks. The guide also emphasizes the importance of maintaining compact system prompts and conversation histories, alongside the benefit of prompt caching to cut input costs by up to 60%. Users are advised to regularly audit their SOUL.md files and tool registrations to minimize overhead, adjust heartbeat frequencies to suit task needs, and consider local or free-tier cloud providers for non-critical agents. Overall, these strategies provide a potential 90% reduction in AI API costs, making it financially viable for users to maintain efficient and cost-effective AI operations.

In 2026, several providers offer APIs for GLM-5.1, a model used in various applications, with significant differences in cost, latency, and feature support. DeepInfra stands out for cost optimization and budget-conscious large-scale deployments, offering the lowest prices and competitive performance metrics. Fireworks excels in throughput for output-intensive tasks, while Wafer offers a balanced approach with strong performance at a moderate cost. Z.ai provides direct access to the model's full feature set, ideal for enterprises needing first-party access, and MindStudio caters to non-technical teams with a no-code platform that integrates GLM-5.1 into existing systems. SiliconFlow, although competitively priced, lacks JSON mode, limiting its use to batch processes, while Puter targets web developers with a user-pays model for frontend app integration. Each provider's suitability varies based on the specific needs of the workload, emphasizing the importance of choosing the right API provider for optimal results.

GLM-5.1, developed by Z.AI and released under the MIT license, is a next-generation Mixture-of-Experts model featuring 754 billion parameters and designed to excel in long-horizon autonomous tasks. Unlike its predecessor, GLM-5, it maintains performance across extended workloads by continuously improving through iterative cycles of planning, execution, and optimization. It has demonstrated significant results, such as autonomously building a Linux desktop environment and enhancing database query throughput. While it excels in agentic engineering tasks, its performance in pure reasoning benchmarks lags behind competitors like GPT-5.4 and Gemini 3.1 Pro. GLM-5.1 is accessible via DeepInfra's OpenAI-compatible API and offers usage-based pricing, with options for self-hosting requiring substantial hardware. Its open-weight and MIT licensing make it a strong candidate for developers focusing on sustained, complex coding workflows.

DeepInfra's guide explores the best Kimi K2.6 API providers for developers, emphasizing how workload priorities such as cost, latency, throughput, and feature support determine the optimal choice. The guide highlights DeepInfra as the recommended provider for cost-optimized production deployments, offering the lowest cached-token pricing ideal for agentic loops. Parasail is noted for its affordability in latency-tolerant batch workloads, while Fireworks excels in low-latency interactive applications with the fastest time to first token. Clarifai and CoreWeave are recommended for maximum throughput, with CoreWeave providing enterprise-grade infrastructure. Moonshot AI offers native multimodal inputs and first-party access, making it suitable for teams needing vendor support. OpenRouter ensures high availability through intelligent request routing, and Atlas Cloud integrates Kimi K2.6 into coding environments while maintaining compliance standards. The guide underscores that DeepInfra is a strong starting point for most production deployments due to its balanced cost and deployment flexibility.

DeepInfra has announced the release of Z.AI's GLM-5.1, an agentic engineering model designed for long-horizon coding tasks, terminal operations, and repository-level work, which surpasses its predecessor GLM-5, as well as competitors like Claude Opus 4.6 and GPT-5.4 in key benchmarks such as SWE-Bench Pro and CyberGym. The model's architecture allows it to sustain performance improvements across numerous rounds and tool calls by iteratively revising strategies, making it particularly suited for complex engineering tasks that involve iterative, multi-step problem-solving. It features a 202,752-token context window and supports JSON and function calls, offering deployment flexibility through an MIT license and competitive pricing. GLM-5.1 is available via DeepInfra's platform with an OpenAI-compatible API, making it accessible for developers aiming to automate engineering applications like autonomous PR triage and repo-scale refactoring tools.

Gemma 4, developed by Google DeepMind and available on DeepInfra, is a family of AI models designed to offer significant improvements over its predecessor, Gemma 3, particularly in areas like mathematics, coding, and agentic tasks. The models, ranging from sub-5B edge-optimized variants to a 31B dense model, leverage a Mixture-of-Experts (MoE) architecture, which activates only a fraction of the total parameters during inference, making them efficient and scalable. Notably, the 26B A4B variant outperforms the previous version with nearly tripled scores on benchmarks like AIME 2026 and LiveCodeBench v6. These models support a wide range of capabilities, including native function calling, extensive multimodal processing, and a 256K token context window, all under an Apache 2.0 license that allows for unrestricted commercial use. DeepInfra provides a straightforward pricing model and an OpenAI-compatible API for seamless integration, appealing to developers who require powerful AI solutions without complex infrastructure setups.

DeepInfra's GLM-5.1 is an advanced reasoning model optimized for long-horizon agentic engineering, released in April 2026, featuring a 754-billion parameter Mixture-of-Experts architecture. It was benchmarked across ten API providers, showing a significant variation in pricing and performance, with blended pricing ranging from $0.74 to $1.70 per million tokens and output speed differences of up to 5.2 times between providers. DeepInfra is highlighted as the best option, offering the lowest costs across all metrics and tying for the fastest time to first token. Fireworks leads in raw output speed, while Wafer provides a balanced alternative. The model is designed for sustained improvements across extensive runs, demonstrated by its ability to autonomously build a Linux desktop environment. DeepInfra's FP8 deployment stands out for its cost-efficiency and practical cached input pricing, making it ideal for complex, agentic workloads.

DeepInfra's Nemotron 3 Super is a cutting-edge model developed by NVIDIA, featuring a 120 billion parameter architecture that combines Mamba-2, Mixture-of-Experts routing, and attention layers under a novel LatentMoE framework, activating only 12 billion parameters per token for efficiency. The model's prowess is demonstrated by its impressive performance on the RULER benchmark, especially at long context lengths of up to 1 million tokens, surpassing competitors like GPT-OSS-120B. Pre-trained on 25 trillion tokens across diverse domains, Nemotron 3 Super is designed for both deep reasoning and conversational tasks, with a configurable reasoning mode that can be toggled as needed. It offers significant advantages in multi-agent pipelines and complex workflows due to its efficient compute budgeting and agentic scaffolding capabilities. Available on DeepInfra's platform, it supports various API integrations and is priced on a usage-based model, making it accessible for scalable deployment in diverse applications.

Gemma 4, developed by Google DeepMind and released in April 2026, is a versatile family of open-weight models designed for diverse deployment contexts, ranging from edge-optimized variants for mobile devices to a 31 billion dense model for server-side tasks. These models, available under the Apache 2.0 license, support multimodal input, built-in reasoning, and an extensive context window of up to 256K tokens, with the 26B A4B Mixture-of-Experts variant and the 31B dense model accessible on DeepInfra. All models use a hybrid attention mechanism and are equipped with a reasoning engine that processes input step-by-step before generating responses, supporting over 140 languages and compatible with various fine-tuning frameworks. The 26B A4B model achieves near-flagship benchmark performance at inference speeds similar to a 4B dense model and is offered at competitive pricing on DeepInfra. This new generation of models represents a significant advancement in reasoning, multimodal capabilities, and context handling, making it suitable for most production workloads.

Gemma 4 26B A4B is a model from Google DeepMind's Gemma 4 family, designed to provide efficient reasoning and multimodal input capabilities, supporting over 140 languages with a hybrid attention mechanism. As of May 2026, seven API providers offer access to this model, with significant variations in performance and pricing. DeepInfra emerges as the optimal choice for production deployment due to its lowest time to first token (TTFT) of 0.68 seconds, competitive pricing, and full context window support of 262K tokens. Clarifai offers the highest output speed, making it suitable for batch processing, while GMI provides a unique 1M token context window for tasks requiring extensive context. Google AI Studio provides a free tier for prototyping, making it an excellent starting point for development. The benchmark highlights DeepInfra's balanced combination of low latency, cost efficiency, and technical features, positioning it as the best overall provider for the Gemma 4 26B A4B model.

Gemma 4 is an open-weight reasoning model developed by Google DeepMind, featuring a 262K context window and available under Apache 2.0 licensing. It is offered by multiple providers, including Cloudflare, DeepInfra, and Google AI Studio, with pricing for token use varying from $0.10 to $0.70 per 1M tokens, depending on the provider and the nature of the workload. DeepInfra emerges as a cost-effective choice for input-heavy applications, benefiting from its low input token pricing and fast initial token latency, making it ideal for prompt-heavy, cost-sensitive workloads like RAG support bots and multimodal document assistants. In contrast, Cloudflare offers competitive pricing for applications where output tokens predominate, such as coding assistants. Gemma 4's extensive context window and reasoning capabilities provide valuable features, yet they also present risks of unintentional token expenditure, emphasizing the importance of testing under realistic conditions. Overall, choosing the right provider involves aligning the workload's token dynamics—input versus output and repeated context usage—with the provider's pricing structure.

DeepInfra's blog post highlights various platforms suitable for deploying the Gemma 4 model in 2026, catering to diverse needs such as cost, latency, data privacy, and infrastructure management. The guide categorizes platforms based on use cases, emphasizing that DeepInfra offers the most cost-effective and low-latency managed API solution without infrastructure setup, making it ideal for most deployments. Google Cloud is recommended for enterprises requiring deep integration and privacy, while Hugging Face is suited for experimentation and model fine-tuning. Clarifai supports on-premise execution with API accessibility for organizations with strict data governance, and Red Hat caters to self-hosted enterprise environments. SiliconFlow provides a managed API without the need for infrastructure provisioning, Ollama facilitates local development across multiple operating systems, Docker integrates Gemma 4 into CI/CD workflows, and MindStudio offers a no-code solution for non-technical teams. The blog underscores the importance of selecting the right platform based on specific optimization goals and provides a detailed cost comparison in the Gemma 4 pricing guide for workload modeling.

DeepSeek V4 is offered by various API providers, each catering to different application needs with distinct pricing and performance options. The model is available in two versions: the V4 Pro, which is a larger model designed for complex tasks, and the V4 Flash, which is optimized for faster and cost-effective inference. DeepInfra is highlighted as the best option for balancing low latency and competitive pricing across both model variants, offering a comprehensive OpenAI-compatible API. Other notable providers include DeepSeek Official API, which offers significant cache savings, Together AI for enterprise-grade infrastructure with global endpoints, Fireworks AI for high throughput and fast token generation, and OpenRouter for avoiding vendor lock-in with automatic fallback routing. For compliance-heavy industries, Atlas Cloud provides SOC 2 Type II certification and HIPAA alignment, while Clarifai offers a fully managed infrastructure with auto-scaling and an interactive UI. The choice of provider should align with specific workload priorities such as pricing, latency, throughput, and regulatory compliance.

DeepInfra's Nemotron 3 Super, an open-weight reasoning model by NVIDIA with 120 billion parameters, is designed for reasoning, tool use, and instruction following, making it suitable for production workloads. The cost of using Nemotron 3 Super varies significantly depending on the provider and workload requirements, with OpenRouter offering the lowest token pricing at $0.09/$0.45 per million input/output tokens, and DeepInfra providing competitive pricing at $0.10/$0.50 with additional features like prompt caching, JSON and function calling support, and private endpoint options. The model's verbose output nature and a 5x output-to-input cost ratio necessitate careful management of token generation to avoid escalating expenses. DeepInfra is highlighted as a preferred choice for production deployments requiring structured outputs and secure, controlled environments, with its infrastructure offering lower latency and predictable costs. The document emphasizes the importance of choosing a provider based on specific application needs and production requirements rather than solely on token cost.

DeepInfra's announcement of raising $107 million in Series B funding highlights its expansion plans for scaling its inference cloud services. The document explores various providers offering APIs and deployment platforms for the Nemotron 3 Super 120B model, which has 120 billion parameters. Each provider is evaluated based on specific use cases, such as DeepInfra for overall value, CoreWeave for interactive applications, Baseten for latency-critical tasks, and Amazon Bedrock for AWS integration. DeepInfra stands out with the lowest blended price, robust API support, and reliable infrastructure, making it a preferred option for production deployments. The document also discusses the Nemotron 3 Super's pricing and performance metrics, emphasizing the importance of choosing the right provider based on workload optimization needs.

NVIDIA Nemotron 3 Super is an advanced 120-billion parameter hybrid Mixture-of-Experts model developed by DeepInfra, optimized for high efficiency and accuracy in AI tasks. It is particularly suited for multi-agent applications and complex reasoning, utilizing a Latent Mixture-of-Experts framework to activate only 12 billion parameters at a time, thus enhancing performance while maintaining agility. The model supports a massive context window of up to 1 million tokens, making it ideal for long-document retrieval-augmented generation and multi-turn operations. It demonstrates superior capabilities in agentic workflows, scientific reasoning, and autonomous software engineering, consistently outperforming peer models in its class. DeepInfra offers the model through an OpenAI-compatible API with competitive pricing and scalability options, allowing developers to integrate it into their applications efficiently. Additionally, the model is designed for deployment on modern hardware like NVIDIA H100 and Blackwell systems, providing flexibility for both cloud-based and local infrastructure implementations.

DeepInfra's GLM-5.1 Pricing Guide outlines the economic considerations and provider options for deploying the GLM-5.1 model, which was released in April 2026 by Z.AI. This model is optimized for long-horizon, tool-using engineering tasks with a large context window of approximately 203,000 tokens. Across 10 benchmarked API providers, pricing varies significantly from $0.74 to $1.70 per million tokens, with DeepInfra offering the lowest blended price and explicit cached input pricing, making it an attractive choice for cost-sensitive, input-heavy, and cache-friendly workloads. Fireworks stands out for speed, Wafer for balance, and OpenRouter for managed access. The guide highlights the importance of understanding token costs, especially for workloads with repeated input patterns, and advises modeling token costs before selecting a provider to avoid unexpected expenses. DeepInfra is recommended for engineering and agentic applications due to its competitive pricing structure, including a unique cached input pricing feature, while Fireworks is preferred for latency-sensitive tasks.

DeepInfra has announced the release of the Nemotron 3 Super 120B A12B model, an open-weight reasoning model from NVIDIA, which is designed for reasoning, tool use, agentic workflows, and long-context instruction following across various languages. The model employs a hybrid Mamba-Transformer Mixture-of-Experts architecture with 120B parameters and features such as LatentMoE for routing accuracy and Multi-Token Prediction layers for native speculative decoding. The model has been benchmarked against providers Lightning AI, CoreWeave, and Nebius, showing a wide range in performance and costs. Lightning AI offers the fastest output speed and lowest time to first answer token, but at the highest cost and without function calling support. CoreWeave provides the best cost efficiency with the lowest Time to First Token (TTFT), while Nebius presents a balanced option with competitive pricing and function calling. DeepInfra itself offers the lowest blended cost and full feature support, including function calling and private endpoint deployment, making it a competitive choice for Nemotron 3 Super deployments.

DeepInfra has raised $107 million in Series B funding to expand its inference cloud infrastructure, aiming to address the growing demand for AI inference workloads. Co-led by 500 Global and Georges Harik, with participation from other notable investors including NVIDIA and Samsung Next, this funding will help DeepInfra enhance its global compute capacity and developer tools. The company emphasizes that inference, rather than training, has become the primary driver of AI workloads, with open-source models reaching parity with proprietary systems and agent-based systems increasing token demand. DeepInfra's platform is designed to meet these challenges with a vertically integrated approach, owning its GPU infrastructure across multiple U.S. data centers and collaborating with NVIDIA for enhanced cost efficiency. The company focuses on high-throughput, low-latency inference using specialized hardware and software, positioning itself as a leader in the AI infrastructure space as it plans to support the next generation of open-source and agentic models.