Multiple Dials, Not One: Reading an Open Model Release

2026-07-27 · Source: Gradient Flow · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure · Depth: Intermediate, quick

Summary

The landscape of "open" model releases has fundamentally shifted, with the term no longer serving as a reliable indicator of cost or capability. Historically associated with affordability and second-tier performance, current open models demonstrate a wide spectrum, exemplified by a top-ranked open coding model costing approximately thirteen times more per million tokens than its cheapest credible open counterpart. This divergence means openness, capability, and price are now distinct factors, often leading to procurement decisions based on incomplete understanding. To navigate this complexity, the article outlines four critical checks: evaluating cost per completed task, assessing cacheable input share (where cached input costs one tenth of uncached), inquiring about peak service availability, and thoroughly reading the specific license terms rather than assuming "open" implies certain rights.

Key takeaway

For AI Architects or Directors making model procurement decisions, relying solely on an "open" label is insufficient and risks costly missteps. You must evaluate models based on actual cost per completed task, understand the impact of cacheable input on your budget (cached input costs one tenth of uncached), confirm peak service reliability, and meticulously review specific license terms. This structured approach ensures your choices align with performance, cost, and legal requirements, moving beyond buzz to informed deployment.

Key insights

"Open" in model releases no longer predicts cost or capability; evaluate models on distinct criteria.

Principles

• Cost per token is misleading; measure per completed task.
• Caching is an architectural decision, not a billing detail.
• "Open weights" doesn't mean self-hosted or free licensing.

Method

Before model selection, perform four checks: measure cost per task, determine cacheable input share, confirm peak service availability, and read the specific license.

In practice

• Design prompts to maximize cacheable input.
• Negotiate pricing based on "open weights" leverage.
• Scrutinize licenses for specific usage rights.

Topics

• Open Models
• AI Model Procurement
• LLM Licensing
• Cost Optimization
• Model Performance
• Caching Strategies

Best for: Director of AI/ML, AI Architect, Consultant

Related on AIssential

• Open weights are quietly closing up - and that's a problem — The erosion of open weights LLM availability threatens market competition and affordability, shifting power to a few frontier labs.
• 20VC x SaaStr This Week : Apple Sues OpenAI, the Token-Maxing Era Begins, and the TAM Question Hanging Over AI Coding — AI's rapid growth faces TAM limits in coding, shifting focus to cost-per-task and broader software integration.
• The Sequence Radar #837: Last Week in AI: From Model Releases to Market Structure — AI is evolving into an infrastructure play, prioritizing scale, integration, and open models over raw product demos.
• Companies turn to Chinese AI models to cut costs — Companies are shifting to Chinese AI models for cost savings and flexibility over premium Western alternatives.
• Open Models Will Absorb Most of the AI Spend — Open models offer superior economics, control, and resilience, driving a shift in AI spend away from proprietary frontier models.
• Breaking the Single-Model Lock-In: What Microsoft Adding Claude to Copilot Teaches Us — Microsoft's multi-model AI strategy for Copilot, despite its deep OpenAI investment, underscores the inherent risk of single-vendor lock-in.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Gradient Flow.