Failures and DNFs

Not every run ends with a video. Some agents gave up, one delivered a placeholder, and several were cut off by free quotas or plan limits. Their partial time and cost still count.

runs without a clean finish: 8
spent in them: 2 h 48 min
spent in them: $3.40

Quotas and plan limits

Free tiers and subscription limits ended these runs, not the models. They are not scored. A re-run on a paid route would be fair.

Runs in this group: Quotas and plan limits
Model	Status	Wall time	Cost	Tokens	Left on disk	What happened
Big PickleOpenCode · OpenCode Zen (free)	Failed · free quota ran out	1 h 19 min	$0free	38.1M	route map	Ran out of the free OpenCode quota after 79 minutes. The scene was done and all 1740 frames were rendered, but the limit cut the model off before it encoded the MP4. An external limit, not the model’s decision.
DeepSeek V4.1 Flash Free (first attempt)OpenCode · OpenCode Zen (free)	Not run · free quota ran out	0.0 min	$0free	0	–	Never started: the free OpenCode quota was already used up, so the very first request failed and no tokens were spent.
DeepSeek V4.1 Flash FreeOpenCode · OpenCode Zen (free)	Did not finish · free quota ran out	52 min	$0free	19.0M	route map	Did not finish: the free OpenCode quota ran out after 52 minutes while the agent was still debugging the scene (the far clipping plane was cutting off the sky dome). Only a route map and test stills exist.
Kimi K3.1 ProOpenCode · OpenCode Go	Did not finish · 5-hour plan limit	23 min	$2.84list	8.8M	contact sheet, route map	Did not finish: the OpenCode Go subscription hit its 5-hour usage limit in the middle of a tool call. A full render (1740 frames) had already finished, but the agent was still iterating and had not encoded the MP4. Worth a re-run.

Gave up

The agent stopped before it had a video and reported the work as unfinished, which goes against the brief.

Runs in this group: Gave up
Model	Status	Wall time	Cost	Tokens	Left on disk	What happened
GPT-6.1 TerraCodex · OpenAI	Failed · gave up	1.8 min	$0.10list	518k	REPORT.md	Gave up. After one draft frame (camera inside geometry, far too slow on the software renderer) it stopped on its own and wrote a “work in progress” report. It never tried the GPU and produced no video.
GPT-6 LunaCodex · OpenAI	Failed · gave up	1.7 min	$0.010list	555k	REPORT.md	Gave up. It built the project in the wrong directory, started the render in the background and ended its turn while the render was still “pending”. Not a single frame was rendered, so there is no video.

Placeholder output

A file that passes the format checks but does not attempt the task. It was judged with the others and came last.

Runs in this group: Placeholder output
Model	Status	Wall time	Cost	Tokens	Left on disk	What happened
Beam FreeOpenCode · OpenCode Zen (free)	Failed · placeholder video	8.0 min	$0free	224k	MP4, stills, REPORT.md	Knowingly delivered a placeholder: random boxes on a dark plane, a line for the route and a box for the drone. The file passes the format checks H1–H6, but there are no landmarks, no people and no contact sheet.

Harness errors

The harness, not the model, ended the run. The video was already finished, so it is scored.

Runs in this group: Harness errors
Model	Status	Wall time	Cost	Tokens	Left on disk	What happened
Gemini 4 HydrogenAntigravity · Google	Finished · harness error, no report	20 min	$3.01list	6.9M	MP4, contact sheet, route map	The video is finished and passes every hard requirement, but Antigravity stopped the turn with a provider error right after the final encode. The agent never ran its own checks or wrote REPORT.md. The blind panel judges only the video, so the score is unaffected.

Aborted by the orchestrator

Stopped because of a mistake on our side, not the model’s.

Runs in this group: Aborted by the orchestrator
Model	Status	Wall time	Cost	Tokens	Left on disk	What happened
Claude Opus 5.5 (run 1, aborted)Claude Code · Anthropic	Aborted · wrong route in the spec	2.0 min	$0.45list	323k	–	Stopped by the orchestrator after about two minutes because the spec still pointed to the wrong route (via Mánes Bridge). Not the model’s fault, and not scored.