2026-09-28 — Recess Borgo v2: cost estimates for a multi-model experiment

Written by Claude Fable 5.1. Prices are OpenRouter list prices pulled live from https://openrouter.ai/api/v1/models on 2026-09-28.

Revised 2026-10-28. The plan is the player-focused design (self-play plus DeepSeek and GLM worlds) in two sizes, Flagship and Full. Earlier plan variants were removed. All plan figures are budget figures, built conservatively.

Bottom line

Budget

Game runs are the player slot plus the world slot (GM and NPCs). Analysis is the two-stage review described below. The rerun buffer is 20% of the lines above, for stalled runs, hit spend caps and relaunches. The platform fee is OpenRouter's roughly 5.5% on credit purchases.

The compute lines do not include operator time, the server, or follow-up experiments the results suggest. The optional stipend covers operator time to launch, monitor and analyze the runs, and the cost of running the server for the duration.

The two plans

Both plans use the same design: every model plays itself, and every model also plays as the player in a DeepSeek world and in a GLM world. The Flagship plan takes the top model from each major lab plus the two world models. The Full plan adds each lab's mid-tier and value models.

• The plan is the player-focused design. Every model plays itself (player and world on the same model). Every model also plays as the player in a DeepSeek V4.1 Flash world and in a GLM-5.3 world. "World" means the game master plus the NPCs. A batch is 6 runs: 2 default, 2 nice, 2 mean.
• Full plan: 19 models, 55 batches, 330 games. Compute total $2,620. Game runs $1,855, analysis $215, 20% rerun buffer $415, platform fee $135.
• Flagship plan: 9 models, 25 batches, 150 games. Compute total $1,600. Game runs $1,160, analysis $105, rerun buffer $255, platform fee $85.
• The game master is about 80% of a run's bill. Self-play on an expensive model is costly because it also pays for its own GM. Putting an expensive model in a cheap world costs a fifth to a tenth as much.
• Reading the results costs far less than producing them. A two-stage review (mid-tier model digests each batch, frontier model reads the digests) is about $215 for the Full plan.

Method

For each model, the average run's tokens are priced at that model's OpenRouter list price, role by role, then a margin is added for the risks listed under "How far to trust these numbers".

• Anthropic models use the cache pattern the Haiku 4.5 runs actually showed. For the GM that was 93% cache reads, 7% cache writes and almost no uncached input. For the player it was 96% reads and 3% writes. Anthropic cache writes cost 1.25 times the input price.
• All other models use the DeepSeek runs' pattern: 91% of GM input and 88% of player input read from cache, no write charge. OpenAI, Gemini, Grok, GLM, Kimi, Qwen and DeepSeek all cache automatically.
• Mistral Medium 3.5 lists no cache price on OpenRouter, so it is priced with no caching. That is why it looks so expensive.
• Output is taken as the same count as DeepSeek produced. Every agent runs with thinking set to "low", and thinking tokens bill as output.
• A run has two slots: the player, and the world (game master plus NPCs, always on the same model). NPCs cost 1 to 4% of a run.

Models and prices

OpenRouter list prices, dollars per million tokens. Twenty models are priced; the plan uses nineteen. Claude Opus 5 is priced for reference but excluded from the plan because Opus 5.5, its successor, is already included and costs less.

Models chosen. The table covers the current flagship and value tier of every major lab on OpenRouter:

• Anthropic. Claude Fable 5.1, Opus 5.5, Sonnet 5 and Haiku 4.5. Haiku 4.5 is still the newest Haiku. Opus 5 is priced for reference but left out of the plan because Opus 5.5 covers it. The older Fable 5, Opus 4.8 and Opus 4.7 are also listed but omitted here.
• OpenAI. GPT-6 Astra, Sol and Luna. Each also has a "pro" variant at the same per-token price.
• Google. Gemini 3.1 Pro (still "preview", and the newest Pro listed), Gemini 3.8 Flash and Gemini 3.5 Flash-Lite.
• Others. xAI Grok 4.7, Z.ai GLM-5.3 and GLM-5.3 Flash, Moonshot Kimi K3, Alibaba Qwen3.8 Max, DeepSeek V4 Pro and V4.1 Flash, MiniMax M3, and Mistral Medium 3.5.
• Left out. Meta's newest Llama on OpenRouter is Llama 4 (April 2025), so it is omitted. OpenRouter also lists :batch variants at half price, but they are asynchronous and the agent runtime makes live calls, so they do not apply.

How far to trust these numbers

The figures are budget figures. They carry a margin over the straight list-price arithmetic, because:

• Verbose models cost more. A model that writes longer narration makes every later turn's input bigger. A model whose "low" thinking still produces a lot of reasoning bills that as output. Output is 13 to 41% of a run's cost, so doubling it adds that much again.
• Caching could fail. If a provider fails to cache the GM's context, that model's cost can rise 3 to 10 times. Check the first run's usage.jsonl for cache reads before launching the rest of a batch.
• Prices change. Several of these models were released in the last month, and OpenRouter prices move.

Two effects run the other way. List prices overstate the open-weight models: DeepSeek at list comes out at $0.46 per run against $0.33 on the meter, and GLM, Kimi, Qwen, MiniMax and DeepSeek Pro are also served by many cheaper providers. And strong GMs tend to end games sooner: in the v1 runs with Haiku 4.5 as GM, games ended at 52 to 54 turns and cost about half the straight estimate. On balance the plan is more likely to come in under budget than over.

Review and analysis costs

The run costs are the whole cost of producing a run. Nothing inside the pipeline calls a model after the game ends: there is no judge, and results generate builds the transcript, metrics, stat replay and summary from the state and event logs with no LLM calls.

The written reviews, such as the GPT Astra Borgo v1 and v2 run reviews of 2026-09-25, were made separately by giving a frontier model the result bundles. They are an extra line item. Measured sizes of the five v2 bundles:

Budget to review one batch, assuming a 10k-token written report, single pass:

An agentic review session that re-reads files and runs its own checks typically costs 2 to 3 times the single-pass figure.

Two-stage design, used in the budget. Expensive models do not need to read raw transcripts.

1. A mid-tier model (Sonnet 5, GPT-6 Sol or Gemini 3.1 Pro) reads each batch's full bundle and writes a structured digest: key events, stat changes, quest outcomes, and anything unusual. About $3.60 per batch, $200 for the 55 batches of the Full plan.
2. A frontier model (Fable 5.1, Opus 5.5 or GPT-6 Astra) reads only the digests and the metric tables and writes the cross-model comparison. The 55 digests total roughly 400k tokens, about $15.

That puts analysis for the Full plan at about $215, against roughly $1,000 if a frontier model read every bundle once, or $2,000 to $3,000 if it did so agentically. The two-stage design also leaves the digests as a reusable record.

The plan: player-focused design

The question this design answers is how each model behaves as the player, holding the world constant. Each model gets three batches of 6 runs:

1. Self-play. The model is the player and the world (GM plus NPCs).
2. DeepSeek world. The model is the player; DeepSeek V4.1 Flash is the GM and NPCs.
3. GLM world. The model is the player; GLM-5.3 is the GM and NPCs. GLM is the second world because it is strong, cheap and from a different lineage than DeepSeek.

DeepSeek in the DeepSeek world and GLM in the GLM world are already self-play, so a set of N models needs 3N − 2 batches. No model is asked to be the GM against other players except DeepSeek and GLM.

The two plans

• Flagship plan (9 models): Claude Fable 5.1, Claude Opus 5.5, GPT-6 Astra, GPT-6 Sol, Gemini 3.1 Pro (preview), Grok 4.7, Kimi K3, GLM-5.3, DeepSeek V4.1 Flash. The top model from each major lab plus the two world models.
• Full plan (19 models): the Flagship set plus Claude Sonnet 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 3.5 Flash-Lite, GPT-6 Luna, Qwen3.8 Max, GLM-5.3 Flash, DeepSeek V4 Pro, MiniMax M3, Mistral Medium 3.5. Each lab's mid-tier and value models. Claude Opus 5 is left out because Opus 5.5 already covers it. Mistral stays despite its price because it is the only model from its lab.

Game-run cost per batch of 6, Full plan

Flagship-plan models are marked with an asterisk.

The row totals are the "Game runs" line of the budget below. Self-play is half of it. The GLM-world column has a floor of about $28 per batch, the cost of GLM running the world whoever the player is. Mistral's figure assumes no caching, because OpenRouter lists no cache price for it; if its first run shows cache reads, its three batches fall to roughly $50.

The DeepSeek V4.1 Flash self-play batch is nearly done: the five existing v2 runs cover 1 default, 2 nice and 2 mean, so it needs one more default run.

To change the number of runs per batch, divide by 6 and multiply by the new count.

Practical notes

• Raise the per-environment spend limits. Every current environment has a $2 limit. A Fable 5.1, GPT-6 Astra, Opus or Mistral Medium self-play run needs $15 to $35, and several others need $5 to $10. Raise the limit per environment when launching (zookeeper.py budget topup), or those runs will stop partway.
• Run the DeepSeek-world column first. It is the cheapest third of the plan, about $244 for all 19 models, and shows every model as the player before any expensive self-play batch is launched.
• Treat the first game of each batch as a pilot. Launch one game, check the meter and the cache reads in usage.jsonl, then launch the other five.