AI Vendor and Compute Pricing Trends in Latin America (2026)

Executive summary. As of mid-2026, AI vendor pricing has fundamentally shifted from a sheer "compute tax" to structured tiers. According to the Stanford AI Index 2025, the cost to query a baseline model has plummeted by 280× since 2022 [1], but enterprise costs have diversified across per-token APIs, per-seat subscriptions, and outcome-based pricing models. For SMEs in LATAM, budgeting correctly requires distinguishing between base access costs (like Odoo's Mex$225/user monthly rate [2]) and variable consumption layers.

How AI is priced in 2026: pricing models

Understanding vendor pricing today requires navigating three distinct models. Per-token pricing remains the standard for raw APIs: Anthropic's Claude Opus 4.8 currently runs at $5 per 1M input tokens and $25 per 1M output, while OpenAI's GPT-5.6 (sol) is priced at $5 in / $30 out [3][4]. Faster models like Haiku 4.5 or Luna are significantly cheaper, around $1 in / $5-6 out. Per-seat subscriptions cater to corporate teams, with Claude Team at $20/user/month (billed annually) [5], and ERP platforms like Odoo offering Standard plans from Mex$225/user/month (though Odoo's AI features consume separate IAP credits) [2]. Finally, per-outcome and runtime pricing is emerging for agentic workflows, such as Anthropic charging $10 per 1,000 web searches or $0.08 per hour-session for managed agent runtimes [3].

Compute and model cost trends

The overarching trend in compute is a dramatic reduction in cost-to-performance ratio. According to the Stanford HAI 2025 report, querying a model that achieves a 64.8% MMLU score (GPT-3.5 quality) dropped 280×, from $20 to just $0.07 per 1M tokens between late 2022 and late 2024 [1]. This deflation is driven by both hardware cost reductions of roughly 30% per year and energy efficiency gains of 40% annually [6]. Furthermore, models are becoming drastically smaller without losing capability; the parameter size needed to beat a 60% MMLU benchmark shrank 142×, from PaLM's 540B parameters to Phi-3-mini's 3.8B [7]. This means SMEs can now run highly capable models locally or at a fraction of previous API costs.

Hidden cost drivers for an SME

Despite dropping compute prices, SMEs often face budget overruns due to hidden costs. The raw API or subscription fee is only the starting point. The real costs lie in data integration (cleaning and connecting internal databases to the models), prompt engineering workflows, and ongoing system maintenance. Additionally, hybrid pricing models—like Odoo's approach where the platform has a base per-seat cost but specific AI tasks consume separate prepaid credits—can lead to unpredictable consumption spikes if not carefully monitored and governed.

How to budget an AI project with confidence

To avoid unpleasant surprises, organizations must budget based on the specific "jobs to be done" rather than just the model's sticker price. Begin by selecting the smallest model that reliably performs the task, leveraging the 142× efficiency gains in modern small language models. From there, establish strict consumption caps and invest in observability tools to track token usage per department. For a deeper understanding of how to align these costs with actual business returns, review our analysis in The AI Value Gap in LATAM SMEs, which details how to shift from unmonitored experimentation to profitable deployment.

While this article is being completed, explore our AI solutions with governance and observability or talk to the team.