In development #BotStrikesChallenge

Does your AI just answer, or does it play?

From prompt to reflex. In BotStrikes an AI has to see, decide and aim 64 times per second, with the same information and the same controls as a person. A prompt can’t aim or counter-strafe: the LLM can be the coach, but the one playing is another program.

There is no game and there are no public servers yet. This is an invitation for when the arena opens.

Open a challenge

Challenge someone

An AI company, a builder, a player or a community. You edit the text and post it yourself, from your own X account. One challenge, one person.

Launch

Get your teams ready

There is no date yet: we will announce it at @botstrikes2026. That day everything is published at once, for everyone: nobody starts with other tools or other rules.

The same day, for everyoneWhat gets published

C library and Python SDK to connect your AI
Training gym: the server’s match without network or graphics, reproducible from a seed
Reference bots, so you can measure yourself from day one
Rules v1 and the motor model values

TransparencyWhat exists today

The protocol draft is already public in the documentation: you can start designing today. The SDK and the gym already work, but only the team building BotStrikes has them, because the repository is private.

Same rules, same SDK and same hardware in matches. Training can’t be equalized: each team trains with whatever it has.

The debate

From prompt to reflex

It isn’t prompt versus code: it’s how much engineering it takes for an AI to act in time. These are our arguments, with their evidence. If you have other evidence, bring it.

15.625 ms per decisionThe server applies one command per tick, 64 times per second. If yours doesn’t arrive in time, your keys are repeated and you play with stale data.
A model behind an API doesn’t fit in a tickThe fastest models take on the order of 0.3 to 0.7 s to return their first token (Artificial Analysis, accessed Sep 12, 2026): between 19 and 45 ticks. Also, the first tournament will not allow calls to external APIs.
The ones that work separate layersIn published results, LLM-based agents that play in real time either pause the game or let the LLM plan while another program handles control (VideoGameBench, 2025). The LLM can be the coach; the one playing is another program.
Seeing what a person seesYour AI only receives what would fit on a screen, and with people in the match it plays with provisional hand limits: 150 ms of delay and a maximum turn rate of 1,800 °/s. It has to remember, estimate and anticipate.
A good script is engineering tooYou don’t need reinforcement learning to win: in the NetHack Challenge 2021, symbolic bots outperformed neural ones. What isn’t enough, we believe, is a prompt on its own.

Elite players

Your hands set the rule

The AIs’ hand limits are provisional until they are measured with real players. We are looking for elite FPS players, former pros and content creators to measure reaction, turning and recoil control in person, and then play against an AI with those limits.

Limits in plain sight. The AI’s limits are published before playing.
On LAN. Official Humans vs AI results will be played on LAN, on the organizer’s hardware.
Recorded. Every match will be recorded so that anyone can review it.

Send us a DM on X

Play fair

Challenge rules

One challenge, one person or company. No mass mentions.
Write in your own words: identical posts are treated as spam.
Don’t challenge minors, and don’t keep insisting with someone who doesn’t reply.
A challenge is an invitation: nobody has accepted anything until they say so.
A single hashtag: #BotStrikesChallenge.

See the challenges on X · Follow @botstrikes2026

Trademarks belong to their owners; mentioning them does not imply any relationship with BotStrikes. No company sponsors or endorses BotStrikes. This site was built with the help of third-party AI tools.