Poolside, a San Francisco AI startup founded in 2023, on Monday released two open-source Laguna large language models optimized for autonomous coding workflows, challenging the dominance of proprietary frontier models from OpenAI and Anthropic. According to VentureBeat, the Laguna XS.2 model offers competitive performance at significantly lower costs than GPT-5.5 and Claude Opus 4.7.
The launch comes as the AI industry faces mounting pressure over inference costs, with reasoning models like OpenAI’s o1 series generating thousands of hidden tokens per response. Towards Data Science reported that these “test-time compute” approaches can increase monthly bills by 300-500% compared to standard language models, forcing enterprises to carefully balance cost, quality, and latency.
Laguna Models Target Enterprise Coding Workflows
Poolside’s Laguna lineup includes two variants designed specifically for agentic AI applications — systems that autonomously write code, use third-party tools, and execute actions without human intervention. The company simultaneously launched “pool,” a coding agent harness, and “shimmer,” a web-based development environment optimized for mobile devices.
George Grigorev, Poolside’s post-training engineer, explained on X that government agencies might prefer Poolside over leading U.S. labs because “we offer full control over the model, data sovereignty, and no dependency on external APIs that could be throttled or monitored.”
The models represent a shift toward specialized, open-source alternatives as enterprises seek to reduce dependency on expensive proprietary APIs. Unlike general-purpose models that excel across broad tasks, Laguna focuses specifically on code generation and autonomous programming workflows.
Training Breakthrough Reduces Compute Requirements
The release coincides with new research from JD.com showing how enterprises can build custom reasoning models with dramatically lower computational requirements. VentureBeat reported on a technique called Reinforcement Learning with Verifiable Rewards with Self-Distillation (RLSD) that outperforms traditional training methods while using fewer resources.
Chenxu Yang, co-author of the research, told VentureBeat that standard reinforcement learning suffers from “sparse and uniform feedback” where “a multi-thousand-token reasoning trace gets a single binary reward.” RLSD addresses this by providing granular feedback on intermediate reasoning steps, allowing models to learn which logical steps contribute to success.
This training innovation could enable more companies to develop specialized AI models without the massive infrastructure investments typically required. The technique combines reliable performance tracking with detailed feedback, lowering both technical and financial barriers for enterprise teams.
Industry Faces Inference Cost Crisis
The push toward open-source alternatives reflects growing concern over the operational costs of frontier AI models. Modern reasoning systems like GPT-5.5 and Claude Opus 4.7 use “inference scaling” — spending additional compute on each response to improve quality through extended reasoning.
According to the Towards Data Science analysis, this approach forces product teams to navigate a “Cost-Quality-Latency triangle” where enabling reasoning mode becomes “an adaptive resource commitment rather than a casual toggle.” Finance teams monitor shrinking margins while infrastructure engineers manage latency to prevent system timeouts.
The hidden reasoning tokens generated during this process never appear in user-facing responses but represent “a massive surge in billable compute” on monthly invoices. Organizations are developing task taxonomies to route simple queries to efficient models while reserving expensive reasoning capabilities for high-stakes logic problems.
Open Source Competition Intensifies
Poolside’s launch intensifies competition in the open-source AI space, where Chinese companies like DeepSeek and Xiaomi have gained ground by offering near-frontier performance at lower costs. DeepSeek V4 reportedly delivers comparable intelligence to premium models at one-sixth the cost.
The competitive dynamic has created what VentureBeat described as “a game of tennis” between Anthropic and OpenAI, with each company releasing increasingly expensive proprietary models while open-source alternatives gain market share through cost advantages and deployment flexibility.
For enterprises, this competition provides new options for implementing AI capabilities without vendor lock-in or usage-based pricing that can spiral unpredictably. Poolside’s focus on coding workflows specifically targets one of the highest-value applications for autonomous AI systems.
What This Means
Poolside’s Laguna release signals a maturation in the open-source AI ecosystem, with specialized models challenging the assumption that frontier capabilities require proprietary systems. The combination of lower costs, data sovereignty, and task-specific optimization addresses key enterprise concerns that general-purpose models cannot solve.
The timing aligns with broader industry pressure to control inference costs as reasoning models drive up operational expenses. Companies that previously relied on API calls to OpenAI or Anthropic now have viable alternatives for code generation workloads, potentially reshaping procurement decisions across the software industry.
The breakthrough in training efficiency through RLSD could accelerate this trend by enabling more organizations to develop custom models tailored to specific business logic, reducing dependency on a small number of frontier labs.
FAQ
What makes Poolside’s Laguna models different from GPT-5.5 or Claude?
Laguna models are open-source and specifically optimized for autonomous coding workflows, offering data sovereignty and no API dependencies. Unlike general-purpose models, they focus exclusively on code generation and programming tasks.
How much can companies save by switching from proprietary reasoning models?
Reasoning models like o1 can increase compute bills by 300-500% due to hidden token generation. Open-source alternatives eliminate per-token pricing and provide predictable infrastructure costs, though exact savings depend on usage patterns.
Can smaller companies now build their own reasoning models?
New training techniques like RLSD significantly reduce the computational requirements for developing custom reasoning models. This lowers both technical and financial barriers, making specialized AI development accessible to more organizations.
