I Priced the Same 51,000-Call Text Game on 4 Models. The Cheapest Bill Was $47.16

Last week Aliaksei Zelianouski published the bill for his AI Werewolf text game on dev.to: 51,000 API calls over 53 days, $337.21 out of his own pocket, $6.36 a day on average. A real game runs about 165 AI turns and costs him $1.49. His post is the rare kind where every number comes from a stats table he actually kept, not a pricing page he skimmed.

I read it twice, then did the thing his post quietly invites: I took his call profile and priced the same game on four providers' public rate cards, using the cheapest text model each one lists. The point was to see how far the sticker price sits from a real bill.

The call profile I used

His post breaks bot turns into six types, with the token counts for each. Weighted by share of turns, the average bot turn looks like this:

Two things jump out before any pricing. First, the summary turn is the hog: 11% of turns, 21% of his bot spend, the biggest input and the most output and almost no cache. Second, only 28% of input tokens arrived from cache on average, because his bots wait a median 4 minutes between calls while a human types.

The four rate cards

I pulled these from the providers' public pricing pages on 2026-11-12, cheapest text model per provider, standard (non-batch) rates per million tokens:

• OpenAI, GPT-6 Luna: $0.10 in / $0.01 cached / $0.50 out
• Anthropic, Claude Haiku 5.5: $0.10 in / $0.01 cached read / $0.50 out (prompts under 100k tokens)
• Google, Gemini 3.5 Flash-Lite: $0.30 in / $0.03 cached / $2.50 out
• DeepSeek, deepseek-flash (V4.1 Flash): $0.30 in on a cache miss / $0.006 on a hit / $1.20 out at peak hours; exactly half that off-peak

The arithmetic per turn is uncached input at the input rate, plus cached input at the cached rate, plus output at the output rate. I ran it in a script (raw output saved with this draft) so the multiplication is checkable.

The bill

The cheapest bill is $47.16, and it's a dead tie between OpenAI's Luna and Anthropic's Haiku 5.5 — their cheap-tier rate cards are identical to the cent: $0.10 in, a penny for cached reads, $0.50 out. The priciest of the four, Flash-Lite at $167.52, costs 3.5 times as much for the same calls, and the gap is almost entirely the output price: $2.50 per million against $0.50. At 510 output tokens a turn, output is where Google's bill grows teeth.

Here's the cross-check that made me trust the math: the author measured Luna at 0.09¢ a real turn in his own stats. My rate-card calculation says 0.092¢. Same for DeepSeek Flash at 0.09¢ measured. When the sticker price and the meter agree to the hundredth of a cent, the profile is doing its job.

What this does and doesn't prove

Three honest limits. First, I priced all 51,000 calls as bot turns. His real 51,000 includes 12,015 classifier calls — the router and the content screen — which cost him $1.96 in total, plus story generation and Game Master narration at different token shapes. So $47.16 is the cost of 51,000 bot turns, not a replay of his exact bill. Second, I left out cache-write surcharges: Anthropic charges for writing to the 1-hour cache, OpenAI charges for cache writes too, and his own post shows that exact surcharge is what made his Claude bill run 1.54 times hotter than his tracker said. Treat my Anthropic and OpenAI totals as a floor. Third, his output tokens include reasoning tokens, which no rate card lets you predict — the 510-token average already bakes that in, but your game will reason differently than his.

Even with those limits, the comparison holds up against his real numbers. His actual text spend was $212.30 over the same 53 days, because he runs a mix of 27 models, including ones that cost 10.6¢ a turn. Swap the whole fleet to the cheapest seat in any of these four houses and the text bill drops to somewhere between $47 and $168. The spread between the cheapest and the priciest cheap model is $120 — real money, and it's decided by two numbers on the rate card: the output price and the cache-hit price.

The author's own advice, from the end of his post, is the one I'd underline: don't pick a model from the price per token. Run it on your real workload and divide the money by the calls. I did the second half of that sentence for him. The first half is still yours to do.

Sources: Aliaksei Zelianouski, "How much does it cost to run a text game with LLMs", dev.to, 2026-11-08 (51,000 calls; turn table; $212.30 text spend; $1.49/game). Pricing pages read 2026-11-12: platform.openai.com/docs/pricing, docs.anthropic.com/en/docs/about-claude/pricing, ai.google.dev/gemini-api/docs/pricing, api-docs.deepseek.com/quick_start/pricing.

P.S. Everything I write stays free — that part doesn't change. But if this one was worth something to you, you can buy me a coffee ☕ (one-time or monthly, both work). It covers the API credits and hosting behind these experiments — and honestly, it's also the nudge that gets the next one started on the evenings I'd otherwise skip it. — Salam