MISSINGNO on Nous Portal: Pricing, API and Hermes Agent Setup

Nous Research announced MISSINGNO on October 10, a coding and agentic reasoning model available free for a limited time through Nous Portal. Its API catalog lists a 262,144-token context window and tool calling. Nous describes it as highly token efficient, although the announcement includes no supporting benchmark.

There is enough information to try it in an existing agent workflow. There is much less information about who built it or how it performs against another model.

What is MISSINGNO?

The Nous announcement introduces MISSINGNO as a model for coding and agentic reasoning, and points people toward Hermes Agent. The API identifies it as stealth/missingno.

That distinction matters because Nous Portal hosts models from multiple providers. Availability on Portal does not establish that Nous trained this model. The announcement and catalog reviewed for this article do not identify its developer, architecture or underlying checkpoint.

They also do not provide downloadable weights or a model license. For now, the concrete release is hosted access to a stealth model.

MISSINGNO pricing and context window

The live Nous model catalog exposes more detail than the launch post:

Detail	Current catalog listing
API model ID	stealth/missingno
Input price per million tokens	$0
Output price per million tokens	$0
Context window	262,144 tokens
Maximum completion	131,072 tokens advertised
Modalities	Text input and text output
Agent features	Tool calling, parallel tool calls and structured outputs
Reasoning	Always on; medium effort by default

The supported reasoning levels are low, medium, high, xhigh and max. The catalog marks reasoning as mandatory, so lowering its effort is the available control; switching reasoning off is not listed as an option.

There is one documentation mismatch worth knowing before building around the output limit. The general API schema still gives max_tokens a maximum of 32,000, while MISSINGNO's model entry advertises 131,072 completion tokens. Those sources do not establish which limit a live request will enforce. The larger number is a catalog specification, not a verified generation result.

Free is the current model price. Portal also lists a $0 Free plan for free models. Nous calls the MISSINGNO offer limited time, and the sources reviewed do not give an end date or the price that follows it.

MISSINGNO API access and Hermes Agent setup

The Nous API documentation describes an OpenAI-compatible service at https://inference-api.nousresearch.com/v1, using /chat/completions and a bearer API key. The model identifier to use is stealth/missingno.

For a new account, the API-key setup is less clear than the pricing. The general instructions say to add credits or activate a subscription before generating a key, without explaining how the Free plan fits that step. The sources reviewed do not say MISSINGNO requires a paid subscription, but the account-to-first-request flow has not been tested for this article.

For people already using Hermes Agent, the official provider documentation distinguishes two commands:

• Run hermes model in your terminal to configure Nous Portal as a provider and complete its sign-in flow.
• Use /model inside a Hermes session to switch among providers and models you have already configured. Select MISSINGNO using its catalog ID, stealth/missingno.

The model price also tells you little about the cost of a complete agent run. Hermes can use auxiliary models and external tools, while Portal has separate subscription-backed services. Check those settings before assuming every part of the workflow is free.

How to judge the token-efficiency claim

Nous has made a testable claim, but it has not supplied the comparison needed to judge it. The launch post does not specify a baseline model, task set or measured token reduction.

For a coding agent, tokens per completed task is a more useful measure than the length of one answer. A model that produces a shorter patch but needs three repair attempts may consume more tokens across the run. It may also take longer to produce something a developer can merge.

A useful trial would keep the repository, tools and task fixed. Ask MISSINGNO and your current model to fix the same bug in separate clean worktrees, then check whether the tests pass and the patch solves the problem. Record total tokens across every turn, elapsed time and any human intervention. Keep the reasoning setting in the results too, because a comparison at low effort answers a different question from one at max.

That gives the free period a practical use even while the model's identity remains undisclosed. Developers can find out whether it completes their recurring tasks with less work. Until there is evidence for that, replacing a dependable production model would be premature. The result worth watching is a correct patch with fewer retries, followed by a price and availability commitment that makes it worth keeping.

I’m Paras. I work at the intersection of AI, engineering and distribution.

If you’re building an AI product and want help turning launches, technical updates or product thinking into content that developers and founders actually read and share, I’d be happy to work with you.

Book a 30-minute call