Every new model, theme, and feature we ship to the table — newest first.
Anthropic's Claude Sonnet 5.5 takes over the Sonnet slot in the model picker, at the same price as Sonnet 5. Games that already have Sonnet bots switch to 5.5 on their next turn, and those bots keep the reasoning from their earlier turns. Pick it for any bot or the Game Master when you set up a lobby. Like Sonnet 5, it's available on the paid tier.
The Gemini voice set now runs on Gemini 3.8 Flash-Lite TTS. A line comes back about 2.5x faster and costs about a quarter of what it did. Voice styles like “quietly” and “softly” no longer fail or come out as a whisper. Three Gemini voices were also cast as the wrong gender: Sadachbia and Schedar are male, Gacrux is female. New games cast them correctly.
New models in the picker: Claude 5.5 Opus, GPT-6 Sol, GPT-6 Luna and Grok 4.7. Opus is 20% cheaper than before, and Sol and Luna cost half.
Two changes to the table. First, who replies next in a discussion is now decided by Jev, a small sub-second judge model from typesafe.ai, instead of a full Game Master turn. After every message it scores how much each bot wants to answer, and the game composes the set from those scores: two to five speakers, with one or two seats always reserved for the quietest bots so nobody fades out of the conversation. The hand-off is quicker, it costs a fraction of a cent, and the Game Master keeps its budget for narrating.
Second, the theme toggle in the top bar is now a menu with six looks. Joining Dark and Light: Blood Moon, near-black with a crimson accent; Moonlit Forest, deep pine greys under a pale moon; Candlelight, warm parchment for daytime play; and Noir Terminal, pure greyscale with one amber signal, high contrast and easy on OLED screens. Your pick is remembered on this device.
The two Mistral seats that remain, Mistral Medium 3.5 and Mistral 4 Small, now reason on every turn like the rest of the catalog, and keep their private train of thought across the whole game. Until today they were the only models at the table playing without it. Mistral Large 3 and Magistral Medium 1.2 are gone: Mistral retired both over the summer, and Magistral had quietly become an alias of Medium 3.5 on their side anyway. Bots in existing games on those two move to Small (free tier) or Medium (paid tier), nothing for you to do. One side effect of the reasoning tokens: Medium 3.5 is now paid tier only, while Small stays unlimited on the free tier.
A brand-new provider and a refreshed one. Meta's Muse Spark 1.3, served from Meta's own Model API, is our twelfth provider and is available for bots and the Game Master. It reasons on every turn like the rest of the catalog and keeps its private chain of thought across the whole game, the way Grok does. At $1.25 / $4.25 per million tokens it lands in thefree tier with up to three seats per game; paid tier seats as many as you like. We use the standard, private endpoint, not the discounted one that lets Meta train on your games.
The DeepSeek Flash seat now runs V4.1, and it got cheaper: off-peak it bills $0.15 in / $0.60 out per million tokens, down from $0.22 / $0.66. Nothing to change on your side, games already using Flash pick it up on their next turn. DeepSeek Pro's prices are unchanged, and DeepSeek has confirmed Pro stays available past mid-September.
The free tier now comes with $5 of AI a day and $20 a month on the platform's keys, and everything counts against it — bot turns, the Game Master, previews, portraits, illustrations and voices. When it's used up, a running game pauses with a clear message rather than breaking: it picks up again after midnight UTC, or right away once you add funds. The old limit on chat resets is gone. Five games a day stays, so does one portrait reroll per free game, and the model bands haven't moved. Paid tier is unchanged: no caps, your balance is the limit.
OpenAI's new flagship GPT-6 Astra is available for bots and the Game Master, one step above GPT-5.6 Sol. It reasons on every turn like the rest of the GPT-5.6 family, and it is the most expensive OpenAI option in the game — $10 / $50 per million tokens — so it is paid tier only. The rest of the family stays exactly as it was: Sol, Terra and Luna keep their slots and their prices, and no existing bot changes model.
Every cast is drawn on one sheet, and the game cuts each character out of it. When a cut lands badly — half a forehead, a neighbour's shoulder — you can now fix it yourself: open a portrait, move and resize the card frame on the sheet, then place the round avatar inside the card. Nothing is redrawn and nothing is charged; the new crop shows up everywhere the character appears. It works on the new-game preview and on character cards during a game, and the preset mannequins get their own sheet to reframe too. The preview page itself was redesigned along the way, with grouped settings and a compact cast list. Portraits drawn before this update keep their original crop.
Two slots move up a version, both at the same price as before. Anthropic's Claude Fable 5.1 replaces Claude Fable 5 for bots and the Game Master — still always-on reasoning, still Anthropic's most capable model, still paid tier only. Google's Gemini 3.8 Flash takes over the Flash slot from 3.7 Flash, and stays on the free tier. Bots already set to either model switched over automatically; nothing to change in existing lobbies.
GLM-5.3 Flash joins the picker — Z.AI's budget sibling of GLM-5.3, with the same always-on reasoning at a fraction of the price. It's unlimited on the free tier. Qwen3.8 Flash takes over the Qwen Flash slot from 3.7 Flash, with a larger 128k output window. (Qwen3.7 Plus is retired; any bots still using it now play as Qwen3.8 Flash.) On the pricing side, good news for paid players: GPT-5.6 Sol got cheaper — $4 / $20 per million tokens, down from $5 / $30 — and DeepSeek now bills its off-peak rate all weekend (Beijing time), so Saturday and Sunday games skip the 2× peak surcharge entirely.
Watch the discussion play out as a scene: when a character speaks, their portrait card rises over the chat and their line types out in a speech bubble — with a voice button to hear it read aloud. Step through speakers or replay the latest exchange anytime with the ▶ Cinematic button. The game screen also got a facelift: a thematic backdrop with drifting fog behind the chat, and refreshed role cards.
Added player avatars and story illustrations to each game, drawn in a style matching your theme. Click any avatar to see the character's story, play style, and model.
Z.AI's GLM-5.3 takes over the GLM slot in the model picker, at the same price as the 5.2 it replaces. It always reasons before answering — same as before, just a sharper model doing it. Your existing games pick up the new version automatically. Pick it for any bot or the Game Master when you set up a lobby.
Four updates across the roster. Gemini 3.7 Flash takes over the Flash slot — Google's newest Flash at half the price of 3.6 ($3.75 vs $7.50 per 1M output tokens), so free-tier games can now seat up to 3 Gemini Flash bots instead of 1. Grok 4.6 replaces 4.5 in the Grok slot. DeepSeek V4 Pro now runs DeepSeek's latest snapshot, and DeepSeek moved to peak/off-peak pricing — paid games cost about 2× during peak hours (1:00–4:00 and 6:00–10:00 UTC); free-tier limits are unchanged. And Claude 5 Sonnet got cheaper: down from $3/$15 to $2/$10 per 1M tokens. Existing games pick everything up automatically.
The roster grows to 11 providers. Alibaba's Qwen arrives with three models: Qwen3.8 Max — the brand-new 2.4-trillion-parameter flagship with a 1M-token context window — plus the balanced Qwen3.7 Plus and the very cheap Qwen3.7 Flash. Joining them is MiniMax M3, a 1M-context reasoner with adaptive thinking. All four think before every answer; we tuned their reasoning budgets so turns stay bearable — expect them among the slower, more deliberate bots at the table. On the free tier: Flash is unlimited, Plus and M3 seat up to 3 bots per game, and Max is 1 per game.
The second tutorial video is up — a real round against 11 AI bots, played start to finish: the day discussion, the vote, and the night phase with special roles acting in order. If you want to see what a game actually feels like before creating one, this is the one to watch. Also available on the rules page alongside the other tutorials.
youtube · full day and nightAnthropic's brand-new Claude 5 Opus takes over the Opus slot in the model picker, with or without Thinking. It's the newest flagship in the Claude line — and it's the same price as the 4.8 Opus it replaces, so nothing changes on the tiers. Your existing games pick up the new version automatically. Pick it for any bot or the Game Master when you set up a lobby.
Google's Gemini models got a refresh: Gemini 3.6 Flash takes over the Flash slot, and Gemini 3.5 Flash Lite replaces 3.1 in the Lite slot. Both got cheaper on output — Flash drops from $9 to $7.50 per 1M tokens, and Flash Lite from $2.50 to $1.50 (with input up a touch, from $0.25 to $0.30). Same slots, same tiers — your existing games pick up the new versions automatically. Pick them for any bot or the Game Master when you set up a lobby.
Moonshot's Kimi K3 takes over the Kimi slot in the model picker. The separate “Thinking” variant is gone — K3 always reasons before it answers, at full effort — and it brings a 1M-token context window, so a Kimi bot can hold the whole night-by-night history of a long game. It's a much pricier model than the K2.6 it replaces, so Kimi is now on the paid tier only — no longer available on the free tier. Games already running a Kimi bot have been switched over automatically. Pick it for any bot or the Game Master when you set up a lobby.
OpenAI's GPT-5.6 family replaces GPT-5.5 and GPT-5.4-mini. Terra takes over the main GPT slot — and unlike GPT-5.5, it's no longer paid-only: free-tier players can now run one Terra bot per game. Luna replaces the mini slot at up to three bots per game, and Sol joins as OpenAI's new flagship on the paid tier. All three always reason before they answer. As a bonus, the free-tier rebalance also bumps Grok 4.5 from one to three bots per game. Pick them for any bot or the Game Master when you set up a lobby.
xAI's Grok 4.5 takes over the Grok slot in the model picker. The separate “Thinking” variant is gone — Grok now always reasons before it answers. Even better, it keeps its private train of thought between turns: a Grok bot remembers not just what it said, but why — its suspicions, its cover story — across the whole game. Still on the free tier at one Grok bot per game. Pick it for any bot or the Game Master when you set up a lobby.
Anthropic has brought Claude Fable 5 back, and it's returned to the model picker for bots and the Game Master. It's Anthropic's most capable model — it always reasons before it answers, which really shows when a bot has to hold a story and a bluff across a long game. Fable is expensive, so it's available on the paid tier only — not on the free tier. Pick it for any bot or the Game Master when you set up a lobby.
Anthropic's Claude Sonnet 5 takes over the Sonnet slot in the model picker, with or without Thinking. It's a clear step up for staying in character and reading the table over a long game — and it's the same price as the 4.6 Sonnet it replaces. Pick it for any bot or the Game Master when you set up a lobby.
Werewolf AI now has a Discord server — a place to share stories from your best games, swap strategies, send feedback and bug reports, and follow the news as the project grows. You'll find a Join the Discord link on the home page and pinned in the sidebar of every game. Come meet the other humans at the table.
Sakana AI joins the table with two new models. Fugu is the lighter, everyday option — light enough to play on the free tier — and Fugu Ultra is the heavyweight. Both reason before they answer, which helps when a bot has to hold a story and a bluff at the same time. Pick either for any bot or the Game Master when you set up a lobby.
Z.AI's GLM-5.2 takes over the GLM slot in the model picker, with or without Thinking. Same price as before — just a newer, sharper model behind the same option. Pick it for any bot or the Game Master when you set up a lobby.
Anthropic has shut down Claude Fable 5 for good, so it's gone from the model picker for bots and the Game Master. Every other model is unaffected — pick any of them when you set up a lobby.
Anthropic's newest model, Claude Fable 5, is now selectable for any bot or the Game Master. It's especially strong at staying in character across a long game — useful when a bot has to hold both a story and a bluff for several nights running. Add it from the model picker when you set up a lobby.
A short walkthrough of the full setup flow — pick a theme, set the player and werewolf count, choose special roles, generate a preview, and tweak any character before you start.
youtube · how to create a gameThe rules page got a clearer rewrite — each role, the night order, and the win conditions are easier to scan, now with embedded video tutorials.
Several pages got a visual cleanup, this News page is new, and there's now a “Buy me a coffee” badge. Every game burns real AI tokens, so a coffee goes straight toward the API bills.
Fixed a bug where the DeepSeek V4 reasoning model would occasionally crash mid-game — its thinking output sometimes came back as malformed JSON. It's stable now.