What “cheapest AI models” really means for developers
The cheapest AI models are language or multimodal models that deliver useful output at the lowest blended price per token, accounting for cache hits, input, and output usage in real-world applications rather than headline list prices alone. In June, the AI pricing comparison story is shaped by two signals: OpenRouter’s leaderboard of token consumption, and Artificial Analysis’ blended pricing table. OpenRouter’s data reflects what thousands of developers and apps are actually running, while blended prices reveal where providers are cutting costs through architecture and caching. Together, they show that lowering token rates is only part of the picture; adoption depends on speed, reliability, and how models behave under production load. For teams planning large-scale deployments, understanding this blended view of model cost analysis is essential before committing to a single provider.
DeepSeek V4 Flash: price champion and usage leader
DeepSeek V4 Flash sits at the centre of both conversations: it is the most popular model on OpenRouter by token consumption and also the cheapest AI model on the blended pricing list. According to Artificial Analysis, DeepSeek V4 Flash (Max) is priced at USD 0.06 (approx. RM276) per 1 million tokens on a blended metric that assumes a 7:2:1 ratio of cache hits, input, and output tokens. On OpenRouter’s leaderboard, the same model records 10.9 trillion tokens and nearly 10x month-on-month growth, a sign that developers value its mixture-of-experts design, 1M context window, and speed. The trade-off is clear: it hallucinates 96% of the time when it does not know an answer, which can be risky for accuracy-sensitive workloads. For bulk document generation or internal tools, though, the price-to-throughput story is hard to ignore.
Blended pricing and the rest of the low-cost pack
Beyond DeepSeek V4 Flash, the cheapest AI models cluster into a tight band of low blended prices that still offer capable performance. GPT-OSS-20B (High), OpenAI’s small open-weight model, comes in at USD 0.07 (approx. RM322) per 1 million tokens blended, trading intelligence for volume and making sense for high-throughput classification or summarisation. At USD 0.18 (approx. RM828), DeepSeek V4 Pro (Max) and MiMo-V2.5-Pro share a slot as budget-friendly reasoning options, with V4 Pro activating 49B parameters per pass despite a 1.6 trillion parameter backbone. GPT-OSS-120B (High) at USD 0.20 (approx. RM920) fills the mid-tier, offering deeper reasoning while staying far below frontier prices. These figures show how MoE designs, smaller parameter counts, and caching let providers drive down blended rates without turning models into toys.
Why the cheapest AI models are not always the most used
OpenRouter’s leaderboard highlights an important tension: the cheapest AI models by blended price are not automatically the most widely adopted across all segments. DeepSeek V4 Flash dominates token usage, but the rankings also feature higher-priced closed-source options like Claude Opus 4.7 and Claude Sonnet 4.6 holding strong positions. Their steady growth reflects enterprise demand for reliability and benchmark-leading reasoning even when cheaper alternatives exist. Meanwhile, OpenRouter’s own Owl Alpha climbs the chart as a likely default or routing model, reminding developers that platform integration can matter as much as raw cost. Hy3 Preview displays explosive growth through strong agentic performance and low nominal input prices, yet it is not listed among the absolute cheapest models on the blended table. Cost, capability, and routing convenience together decide adoption, not headline prices in isolation.
Turning AI pricing comparison into real savings
For businesses, the lesson from the OpenRouter leaderboard and blended pricing tables is that cost optimisation starts with better matching workloads to models. Output-heavy tasks like document generation may benefit from DeepSeek V4 Flash’s ultra-low blended rate, while complex reasoning in production might justify paying more for models such as Claude Opus 4.7 or DeepSeek V4 Pro. Developers should watch the difference between official input and output prices and real-world blended costs shaped by cache hits, especially when context windows reach hundreds of thousands of tokens. Routing architectures that send simple prompts to smaller models like GPT-OSS-20B and reserve larger models for hard problems can dramatically cut the effective cost per useful token. The pricing war has shifted power to buyers; the next step is building systems that use that advantage with intent.
