Krafton’s PUBG Ally tests put Claude Opus 4.8’s median API cost at $7.10 per English-language match replay. For studios building AI squadmates, that turns the choice of language model into a question of operating costs as well as capability.
The project underpins Ella, the AI companion introduced in the PUBG Ally Duo beta. Players teamed up with her on Sanhok, using voice communication to coordinate with a computer-controlled partner. Making that interaction affordable and responsive presents a different challenge from producing an impressive conversation in a demonstration.
What Krafton’s AI cost comparison measured
Krafton’s September 24 research preprint, reviewed by TalkEsport, reports median API costs across 30 replayed sessions. Selected English-language figures from Appendix G appear below, in US dollars.
| Model | API provider | Cost per replayed match |
|---|---|---|
| Claude Opus 4.8 | Anthropic | $7.10 |
| Claude Sonnet 4.6 | Anthropic | $5.17 |
| DeepSeek-V4-Pro | OpenRouter | $0.17 |
| Gemma-4-31B-it | OpenRouter | $0.02 |
Krafton chose Gemma 4 31B as its teacher for data collection, finding Opus impractical at scale. Claude provided a quality reference, not training labels. These are research charges, not player fees or proof of equivalent performance.
Why PUBG Ally runs locally
The player-facing system takes a different approach. In a June engineering interview published by NVIDIA, Krafton described running a compact, two-billion-parameter language model on the player’s GPU. The team said cloud processing introduced delays that made live squad communication feel too slow.
Ally combines a transcription of the player’s speech with information supplied by the game engine. Its language model handles communication and broader decisions, while a separate gameplay system manages movement, aiming and immediate combat reactions. That division lets the character react without waiting for a spoken response to finish processing.
Local processing also has limits. Krafton’s engineers explained that PUBG already occupies a substantial share of GPU memory, leaving limited room for an AI companion. They compressed the model to fit within that remaining capacity.
The hardware trade-off behind AI teammates
Those constraints reached the public beta. Krafton’s official Update 42.1 patch notes restricted Ally Duo to PCs with NVIDIA GeForce RTX graphics cards. The notes also advised players experiencing delayed responses to lower their graphics settings or resolution. That guidance shows how an AI companion can compete with the game for a player’s available computing resources.
TalkEsport previously covered PUBG Ally’s place in Krafton’s wider AI strategy. Its implementation illustrates the practical decisions behind that ambition: a studio must consider recurring cloud charges, response speed and the hardware its audience owns. Running inference locally avoids that particular API bill, but development and computing still carry costs. For players, the useful test remains whether the companion can help during a match without compromising the experience around it.
