NaiveAI Prices New Flash Model at 10 Cents Per Million Input Tokens

NaiveAI's research page, dated September 27, lists Naive-N0.5-Flash at $0.10 per million input tokens, $0.40 per million output tokens, and $0.01 per million cached reads. The model is a mixture-of-experts system with 309 billion total parameters and 15.5 billion active per token, released under an MIT license. NaiveAI states a native context window of 1 million tokens and a serving throughput figure, branded NaiveRT, of up to 2,000 tokens per second. Those are the company's own numbers, posted on its own site. No independent benchmark or third-party pricing comparison accompanies the post.

For a manager approving an API budget, the active-parameter count matters more than the headline size. Running only 15.5 billion parameters per token, despite a 309 billion parameter total, is what lets NaiveAI quote $0.10 input pricing at all. That architecture choice is the mechanism behind the price, not a separate selling point. Anyone evaluating cost should ask what quantization or serving stack produces the claimed 2,000 tokens per second, since throughput claims tied to a specific runtime do not always hold on generic infrastructure the way they do on the vendor's own.

Workers integrating this into existing pipelines carry the near-term risk. An MIT license means no restriction on commercial use or redistribution, which is a real concession compared with gated or usage-capped competitor licenses. But license terms say nothing about uptime, rate limits, or how long the quoted price holds. NaiveAI's announcement does not disclose a service-level agreement, regional availability, or whether the $0.01 cache-read price applies only under specific prompt-caching conditions. Those gaps do not appear anywhere in the September 27 post.

Policy watchers tracking compute costs and model proliferation should treat the 1 million token context window as the more consequential figure. A native 1M window, if it holds under load rather than degrading in quality past some shorter effective range, changes what counts as a single query versus a multi-call workflow. That shift affects total cost calculations more than the per-token price alone. It also raises questions about data retention during long-context sessions that the announcement does not address.

Three numbers anchor the current claim: 15.5 billion active parameters, 1 million token context, and $0.10 per million input tokens. All three come from NaiveAI's own research page, not from an outside filing or audited benchmark. Teams considering adoption should ask for a reproducible throughput test under their own load, not NaiveAI's.

The next checkpoint is simple. Does the 2,000 tokens per second figure hold outside NaiveAI's own infrastructure, under a contract that specifies what happens if it does not.