Jevman AI Pac-Man Benchmark: Jev Price and Cost per Game
Jevman tests six AI decision models in Pac-Man. Compare Jev's official API rate, Opper game costs, gateway fees, and benchmark limits.
By AI Pricing Guru Editorial Team
AI Pricing Guru articles are maintained by the editorial workflow behind the site: daily pricing snapshots, provider source checks, and review passes for model launches, subscription limits, and billing changes.
TL;DR
- Opper's Jevman leaderboard reports 100 games for each of six decision models; Jev 1.13 leads its published mean score, with statistical ties possible.
- TypeSafe's direct Jev API list price is $0.042 per million input tokens, with free output tokens. Opper has separate credit-purchase fees.
- The game is a vendor-run decision-model harness, not our Labs score or proof that Jev is cheapest for every production task.
Current direct TypeSafe Jev API rate—not Opper's game bill
| Model | Provider | Input / 1M | Cached / 1M | Output / 1M |
|---|---|---|---|---|
| TypeSafe Jev 1.13.0 | typesafe | $0.042 | n/a | $0 |
Built from pricing.json at publish time.
Opper AI’s Jevman asks decision models to play Pac-Man against classic scripted ghosts. Its public leaderboard records 100 games per model for six models. Jev 1.13 has the highest published mean score, 2,750, followed by Opper’s GPT-6 Luna Decisions route at 2,568 and Clef Flash at 2,538. Opper treats overlapping 95% margins of error as ties, so the table is not a claim that every score difference is statistically decisive. The leaderboard and method are public, as is the underlying recorded summary.
What does Jev cost?
TypeSafe, the creator of Jev, publishes $0.042 per million input tokens ($42 per billion) and $0 per million output tokens for its direct System One API. Its API quick start shows the jev-latest request and a response identifying jev-1.13.0; Jev is described as early access. This is a typed decision API, not a general-purpose text generator. The live table above is the maintained direct TypeSafe rate. Opper’s GPT-6 Luna Decisions route is a separate decision product; do not substitute the standard GPT-6 Luna API token rate for it without verifying that route’s own terms. See our TypeSafe Jev pricing guide for the route and access distinctions.
If you call the models through Opper, its official price card says provider token rates pass through at cost, but credit purchases incur a 3% Gateway fee or 5.5% Control Plane fee. Those fees are charged when buying credits, not as a per-call token markup. Bring-your-own-key is free on the Gateway tier; taxes and optional checks can change the bill. Do not treat the direct TypeSafe row in our pricing feed as the final Opper invoice.
Which route pays for the game?
The project README distinguishes a signed-out hosted demo (draws on Opper’s finite shared-credit pool while funded), a signed-in player (uses that player’s Opper wallet), a local direct TypeSafe key (the game estimates cost because TypeSafe’s response does not return a charge), and a custom HTTP endpoint (the operator pays its own upstream or infrastructure costs). The free demo is not a permanent free Jev API tier. Pick the route before comparing bills.
What did the 100-game run cost?
| Model | Opper mean score | Opper-reported cost per game | 100 games at that reported average* |
|---|---|---|---|
| Jev 1.13 | 2,750 | $0.00593 | $0.593 |
| GPT-6 Luna Decisions | 2,568 | $0.00968 | $0.968 |
| Clef Flash | 2,538 | $0.00362 | $0.362 |
| Clef | 2,476 | $0.01728 | $1.728 |
| Kev 4B | 1,506 | $0.00282 | $0.282 |
| Laya | 639 | $0.00340 | $0.340 |
*Arithmetic is 100 × the rounded cost-per-game field in Opper’s published leaderboard JSON, generated October 7. These are vendor-run workload costs, not published token list rates for the other five routes and not independent invoices from AI Pricing Guru. Opper’s gateway credit fee, taxes, different workloads, and access terms can alter customer spend. A cheap game is not necessarily a cheap accepted business decision.
What the benchmark measures—and what it misses
Each model gets a structured game state at junctions and chooses a direction with probabilities. A reply that misses the two-second deadline is replaced by a greedy backup rule; Opper reports fallback rates alongside score and latency. The repo can replay recorded games and accept self-reported community submissions, but the six-model comparison was run by the company operating the benchmark and gateway. It is useful evidence about a real-time structured-decision workload, not an independent ranking of chat, coding, research, or enterprise automation.
Our Cost-per-Task Labs uses a fixed 49-task text-answer suite. Jev’s typed System One response and Pac-Man action loop do not satisfy that contract, and Opper’s game score cannot be imported as our own. The Labs coverage note lists the exact replay blocker: we would need a pinned game build, model routes, identical prompts and state, raw calls, backup-move policy, token/credit ledger, and an acceptance metric before comparing cost per successful decision.
For teams testing their own open-weight decision endpoint, the Jevman repository supports an HTTP endpoint submission. If you need GPU infrastructure for such a separate model, compare current RunPod options and include idle GPU time, serving engineering, and retries. That link may earn us a commission at no extra cost to you; it does not provide Jev’s proprietary weights or reproduce Opper’s leaderboard automatically.
Bottom line: Jev has an unusually low official direct input-token rate and the best mean score in Opper’s six-model Pac-Man sample. Choose a production route by measuring your own accepted decisions and full invoice, not by multiplying a game score into a business ROI claim. Official pricing and benchmark sources checked October 9, 2026; rates remain under daily official-source verification.