V4 Pro · high thinking
Agendo harness · DeepSeek API
20/20
One repair fixed post-lock spawning and repeated hard drops.
Model time9m 09s
Actual cost$0.07219
First audit18/20
Repairs1
Final result
All three games now pass
Final score combines a 16-check Playwright behavior audit with four source-verified game paths: bag randomization, line scoring, level-speed progression, and game-over/high-score persistence.
GPT‑5.6 Sol · high reasoning
Agendo harness · ChatGPT subscription
20/20
Passed the smoke suite and deeper audit without a repair turn.
Model time5m 49s
Direct charge$0.00*
First audit20/20
Repairs0
Ling 3.0 Flash Free
Native OpenCode runner · limited-time free model
20/20
Three repair prompts fixed state snapshots, spawning, rotation, locked rendering, and level progression.
Model time7m 09s
Actual cost$0.00
First audit15/20
Repairs3
Efficiency
Time and token traffic
Time is the sum of recorded successful build/repair session spans. Token traffic includes uncached input, cached input, and output; OpenCode reasoning is shown separately below.
Recorded usage
Tokens, timing, and actual cost
Counts come from Agendo wire records and OpenCode’s local session database. No credentials or prompt contents are included.
| Route | Uncached input | Cached input | Output | Reasoning† | Time | Actual cost |
|---|---|---|---|---|---|---|
| DeepSeek V4 Pro Build + repair | 77,834 | 1,448,704 | 38,020 | included | 548.7s | $0.072186742 |
| Codex GPT‑5.6 Sol Build, no repair | 171,194 | 248,320 | 16,872 | included | 349.0s | $0 direct* |
| OpenCode Ling Free Successful build + repairs | 220,951 | 3,691,200 | 38,248 | 36,193 | 429.2s | $0 |
| OpenCode all attempts Includes unavailable/failed free models | 260,640 | 4,120,000 | 53,870 | 36,318 | 529.6s‡ | $0 |
DeepSeek cost: 77,834 × $0.435/M + 1,448,704 × $0.003625/M + 38,020 × $0.87/M = $0.072186742. *Codex: this run used ChatGPT subscription quota, so there was no per-run API charge; the monthly plan fee is not allocated. † OpenCode reports reasoning separately. ‡ Provider-active session time excludes a CLI hang between model attempts.
Quality path
What it took to reach 20/20
“Repair” means a new model prompt after an independent Playwright or source audit found a concrete failure.
- Build
Polished game, but the next piece lacked finite spawn coordinates after lock. - Repair
Fixed promoted-piece spawning, repeated hard drops, and defensive state. - Final
16/16 Playwright checks; 20/20 rubric.
- Build
Complete DOM game with 7-bag, wall kicks, responsive controls, and immutable test state. - Audit
No remediation requested. - Final
16/16 Playwright checks; 20/20 rubric.
- Selection
DeepSeek Flash produced a truncated file; North stalled; MiMo had no provider; Ling completed. - Repair 1
Fixed mutable state snapshots and promoted-piece coordinates. - Repairs 2–3
Fixed rotation teleport, invisible locked blocks, and non-linear level thresholds. - Final
16/16 Playwright checks; 20/20 rubric.
Feature completeness
Requirement-by-requirement
Green checks passed on the first audited delivery. Orange repair marks identify areas that needed a later model turn. Every final URL passes.
| Requirement area | DeepSeek first delivery | Codex first delivery | OpenCode Ling first delivery | Final result |
|---|---|---|---|---|
| One self-contained HTML; no runtime dependencies | Pass | Pass | Pass | 3/3 |
| 10 × 20 inspectable DOM board | Pass | Pass | Pass | 3/3 |
| Seven tetrominoes and randomized bag | Pass | Pass | Pass | 3/3 |
| Collision and valid promoted-piece spawning | Repaired | Pass | Repaired | 3/3 |
| Rotation with wall kicks | Pass | Pass | Repaired | 3/3 |
| Line clearing | Pass | Pass | Pass | 3/3 |
| Scoring | Pass | Pass | Pass | 3/3 |
| Levels and increasing speed | Pass | Pass | Repaired | 3/3 |
| Next-piece preview | Pass | Pass | Pass | 3/3 |
| Ghost piece | Pass | Pass | Pass | 3/3 |
| Hard drop, lock, render, and continue | Repaired | Pass | Repaired | 3/3 |
| Game-over state | Pass | Pass | Pass | 3/3 |
| Saved high score | Pass | Pass | Pass | 3/3 |
| Start, pause/resume, restart | Pass | Pass | Pass | 3/3 |
| Keyboard controls | Pass | Pass | Pass | 3/3 |
| Visible working touch controls | Pass | Pass | Pass | 3/3 |
| 375 × 812 responsive layout | Pass | Pass | Pass | 3/3 |
| Accessible names and visible focus | Pass | Pass | Pass | 3/3 |
| Reduced-motion mode | Pass | Pass | Pass | 3/3 |
| Fresh defensive test-state snapshot | Pass | Pass | Repaired | 3/3 |
Artifacts
Smallest file, biggest repair bill
All files are intentionally self-contained. SHA‑1 fingerprints are the locally audited versions used for deployment verification.
| Build | Bytes | SHA‑1 |
|---|---|---|
| DeepSeek | 21,151 | 1f9beedb694abe8f809008a0bcf5a9b95b1dbb25 |
| Codex | 35,404 | 90df6c42532715066fc958892d4b2cf2a4a989e4 |
| OpenCode | 17,543 | 35386f4a526fb941256cc2ac783cccbebcb052a0 |
Identity exploration
Agendo icon candidates
Six directions on one selection sheet. The product header currently uses the Agendo wordmark while a final mark is selected.
Method & sources
How the numbers were produced
Pricing was checked on 27 July 2026. Provider prices and limited-time free availability can change.
DeepSeek
Official V4 Pro direct API rates: $0.003625/M cache-hit input, $0.435/M cache-miss input, $0.87/M output.
Official DeepSeek pricing ↗Codex
Authenticated through the existing ChatGPT subscription. Codex and ChatGPT Work share plan usage; this was not an API-key run.
Official Codex pricing ↗OpenCode
Ling 3.0 Flash Free is listed at free input, output, and cached-read pricing for a limited time.
Official OpenCode Zen pricing ↗
Agendo is a customized Kimi Code fork. Upstream Kimi Code copyright and license references are preserved.
Audit: Chromium + Playwright · local/remote SHA verification · no external runtime assets.