Jev Model First Look: The AI Fast Enough to Run Doom in Real Time

An AI Model That Claims to Be 10x Faster

What would it feel like to use an AI model that's 10x — or even 20x — faster than anything currently available? One content creator posed exactly that question while putting a model called Jev to the test. According to the video, Jev was built by a core contributor to ChatGPT — someone described as having "co-invented ChatGPT" — a backstory that immediately draws attention.

Unlike the many models on the market focused primarily on conversation, the creator emphasized that Jev is positioned for "real work" rather than casual chat. That means the emphasis is on execution efficiency, response speed, and actual task throughput — not on polished small talk.

Blazing Speed and Pricing That's Almost Free

Jev's most striking selling point is its combination of speed and cost. The video claims the model is "so efficient, so fast, and so cheap that they don't even charge for output tokens." In other words, output tokens are free — users only pay for input.

And the input pricing is equally jaw-dropping: roughly four cents per million input tokens. The creator described it as "nothing" — effectively free. For developers and businesses that make heavy API calls, process large volumes of text, or need high-frequency interactions, this pricing structure could offer a dramatic cost advantage.

If accurate, this pricing would directly disrupt the cost model of mainstream AI services. Most commercial LLMs charge separately for input and output tokens, with output typically costing three to five times more — because autoregressive generation is more computationally intensive and memory-heavy than reading context. For reference, Claude Sonnet charges around $15 per million output tokens and $3 per million input tokens. Jev's claim of zero output cost and $0.04 input cost would represent a reduction of several hundred times in overall call expenses.

From a business perspective, there are two likely explanations for pricing this aggressive: either there's a fundamental architectural or inference breakthrough (such as sparse activation or speculative decoding) that dramatically reduces generation costs, or this is a subsidized early-adoption pricing strategy whose long-term sustainability is uncertain. For developers evaluating whether to migrate their workflows, distinguishing between these two scenarios is critical.

Performance Benchmarked Against Top-Tier Models

Speed and price are compelling — but it all comes down to whether the model is actually good enough. The creator's assessment: Jev's quality is on par with Luna, Terra, and Sonnet 5, at a fraction of the price.

That's a bold claim. If a model can come close to flagship-level performance while slashing costs to almost nothing, it would fundamentally reshape how developers and teams approach model selection. That said, these conclusions come from a single source's initial impressions and would need systematic benchmarking to be verified.

Among the reference models mentioned, "Sonnet 5" refers to Anthropic's Claude Sonnet series (a mid-to-high-end product line). "Luna" and "Terra," however, are not widely recognized mainstream model names at the time of writing — they may be internal names from a specific platform or emerging vendor, or informal labels used by the creator. This is worth flagging: when the reference points themselves have limited public visibility, claims of "comparable performance" are difficult to independently verify. There are no public benchmarks like MMLU, HumanEval, or LMSYS Chatbot Arena Elo scores to cross-reference. Rigorous model evaluation typically requires blind testing on standardized datasets or large-scale user voting systems — a single video demo has inherent limitations in persuasive power.

Fast Enough to Play Doom in Real Time

To illustrate the speed in tangible terms, the video presents a striking example: Jev is fast enough to "play Doom in real time."

Using an AI model to "play" or "generate" gameplay from a classic like Doom is essentially a stress test of inference throughput — the model must continuously produce large volumes of output within extremely short time windows to sustain the frame rate a real-time game requires. Achieving this means inference latency has been pushed to remarkably low levels. This also aligns logically with the free output token claim: only when generation costs are low enough and speeds are high enough does it make commercial sense to offer output for free.

"Running Doom on AI" has become an informal benchmark for inference speed, rooted in Google DeepMind's 2024 GameNGen project — the first demonstration of a neural network simulating Doom gameplay in real time without relying on a traditional game engine, requiring per-step inference latency under 33 milliseconds (equivalent to 30 fps). Since then, "can it run Doom in real time?" has evolved into a community shorthand for testing a model's raw generation speed. Each frame requires the model to output a large number of pixels or action instructions based on the current state; the higher the frame rate, the more demanding the requirement on tokens-per-second throughput. This is fundamentally different from a typical chat scenario where users can tolerate delays of several seconds — it's a simultaneous test of both low latency and high throughput. So when a model claims it can "run Doom in real time," the implication is that its inference speed has entered a new millisecond-level tier, not merely that it "answers questions faster."

Promising — But Independent Testing Is Still Needed

The creator wrapped up by saying they plan to "heavily test" the model, and might even make it their "daily driver." That framing captures both excitement and caution — after all, every highlight showcased so far is still at the demo and first-impression stage.

For readers, there are a few signals worth watching: first, model competition is shifting from "more capable" toward "fast, cheap, and good enough"; second, aggressively priced strategies like free output tokens may foreshadow a broader compression of industry cost structures; third, demos like "real-time Doom" are becoming an intuitive yardstick for measuring inference speed.

A word of caution, though: all information in this article comes from a single video source. The specific performance against reference models, the long-term stability of the pricing, and reliability in real production tasks all lack independent third-party verification. Before adopting it as a productivity tool, waiting for more systematic benchmarking data would be the prudent path.