Launch 005 · September 18, 2026

Qwen 3.8 Flash vs DeepSeek V4.1 Flash vs GLM 5.3 Flash vs Gemini 3.8 Flash

There was no universal winner. Flash-model performance was brutally task-specific. Qwen led several game and 3D-scene rounds, DeepSeek produced the better submitted CAD result, GLM alone passed Apple Stem, and Gemini led several robot tasks.

14 fixtures · 54 direct clips · 54 permanent results (35 reused, 19 added) · single attempts, no aggregate score

Watch every fixture Inspect all public results Watch the full episode on YouTube

Fixture 01 · Dex Cube

Rubik’s Cube single turn

GLM and Gemini completed the single-turn task; Gemini was fastest. Qwen and DeepSeek did not complete it.

Qwen 3.8 Flash · dex-cube-turn presentation recording poster — **Qwen 3.8 Flash**Open result
$0.261648m 34.9sMaxCampaign judgment: did not pass
Record details
Campaign evidence · campaign evidence not harness ledger
Campaign-evidence record; this attempt was not generated from the public Harness ledger.
This permanent campaign-evidence record uses the handoff’s exact displayed cost and wall clock; it is not a generated Harness ledger.
Public record identity: Qwen 3.8 Flash · Qwen 3.8 Flash · Unavailable · Unavailable

DeepSeek V4.1 Flash · dex-cube-turn presentation recording poster — **DeepSeek V4.1 Flash**Open result
$0.095716m 38.0sMaxCampaign judgment: did not pass
Record details
Campaign evidence · campaign evidence not harness ledger
Campaign-evidence record; this attempt was not generated from the public Harness ledger.
This permanent campaign-evidence record uses the handoff’s exact displayed cost and wall clock; it is not a generated Harness ledger.
Public record identity: DeepSeek V4.1 Flash · DeepSeek V4.1 Flash · Unavailable · Unavailable

GLM 5.3 Flash · dex-cube-turn presentation recording poster — **GLM 5.3 Flash**Open result
$0.341440m 47.8sMaxCampaign judgment: pass
Record details
Campaign evidence · campaign evidence not harness ledger
Campaign-evidence record; this attempt was not generated from the public Harness ledger.
This permanent campaign-evidence record uses the handoff’s exact displayed cost and wall clock; it is not a generated Harness ledger.
Public record identity: GLM 5.3 Flash · GLM 5.3 Flash · Unavailable · Unavailable

Gemini 3.8 Flash · dex-cube-turn presentation recording poster — **Gemini 3.8 Flash**Open result
$2.0915m 35sHighCampaign judgment: pass
Record details
Campaign evidence · campaign evidence not harness ledger
Campaign-evidence record; this attempt was not generated from the public Harness ledger.
This permanent campaign-evidence record uses the handoff’s exact displayed cost and wall clock; it is not a generated Harness ledger.
Public record identity: Gemini 3.8 Flash · Gemini 3.8 Flash · Unavailable · Unavailable

Fixture 02 · Dex Cube

Rubik’s Cube three-move extension

Neither GLM nor Gemini completed the harder three-move extension.

GLM 5.3 Flash · dex-cube-three-move presentation recording poster — **GLM 5.3 Flash**Open result
$1.44121h 31m 15.1sMaxCampaign judgment: did not complete
Record details
Campaign evidence · campaign evidence not harness ledger
Campaign-evidence record; this attempt was not generated from the public Harness ledger.
This permanent campaign-evidence record uses the handoff’s exact displayed cost and wall clock; it is not a generated Harness ledger.
Public record identity: GLM 5.3 Flash · GLM 5.3 Flash · Unavailable · Unavailable

Gemini 3.8 Flash · dex-cube-three-move presentation recording poster — **Gemini 3.8 Flash**Open result
$5.6544m 30sHighCampaign judgment: did not complete
Record details
Campaign evidence · campaign evidence not harness ledger
Campaign-evidence record; this attempt was not generated from the public Harness ledger.
This permanent campaign-evidence record uses the handoff’s exact displayed cost and wall clock; it is not a generated Harness ledger.
Public record identity: Gemini 3.8 Flash · Gemini 3.8 Flash · Unavailable · Unavailable

Fixture 03 · Campfire

Campfire

Qwen produced one of the strongest visual outputs in the campaign’s Campfire fixture.

DeepSeek V4.1 Flash · campfire presentation recording poster — **DeepSeek V4.1 Flash**Open result
$0.52761h 10m 27.4sVision expPASS
Record details
Existing public result · PASS
No additional public caveat was supplied by this result record.
Metric scopes materially disagree. Clip handoff: $0.5276, 1h 10m 27.4s. Existing result receipt: $0.08 Provider-recorded OpenRouter cost; Not recorded. Both remain visible; neither silently replaces the other.
Public record identity: DeepSeek V4 Flash · deepseek-v4-flash-vision-exp-openrouter · Provider not separately recorded

GLM 5.3 Flash · campfire presentation recording poster — **GLM 5.3 Flash**Open result
$0.01517m 21.718sMaxPASS
Record details
Existing public result · PASS
No additional public caveat was supplied by this result record.
Clip handoff: $0.0151, 7m 21.718s. Existing result receipt: $0.02 Provider-recorded OpenRouter cost; 7m 21.718s wall-clock. More precise clip formatting is preserved without changing the receipt.
Public record identity: GLM 5.3 Flash · glm-5.3-flash-max-opencode · Provider not separately recorded

Gemini 3.8 Flash · campfire presentation recording poster — **Gemini 3.8 Flash**Open result
$4.441h 11m 27.5sHighestpass
Record details
Existing public result · complete token timing artifact and validation ledger
No blind visual quality score is attached.
Clip handoff: $4.44, 1h 11m 27.5s. Existing result receipt: $4.44 Provider-recorded usage estimate, not an itemized subscription cash charge; 4287.5s wall-clock. More precise clip formatting is preserved without changing the receipt.
Public record identity: Gemini 3.8 Flash · Gemini 3.8 Flash · OpenRouter · OpenCode 1.18.23

Fixture 04 · Fighter Blossom

Courtyard Blossom Stage 2 / Blender arena

Qwen produced one of the strongest arena outputs in the campaign.

Qwen 3.8 Flash · courtyard-blossom-stage-2 presentation recording poster — **Qwen 3.8 Flash**Open result
$1.172h 35m 9.1sUnreportedFAIL
Record details
Existing public result · FAIL
No additional public caveat was supplied by this result record.
Metric scopes materially disagree. Clip handoff: $1.17, 2h 35m 9.1s. Existing result receipt: $0.71 Provider-recorded OpenRouter cost; 1h 33m 46.679s wall-clock. Both remain visible; neither silently replaces the other.
Public record identity: Qwen3.8 Flash · qwen3.8-flash-openrouter · Provider not separately recorded

DeepSeek V4.1 Flash · courtyard-blossom-stage-2 presentation recording poster — **DeepSeek V4.1 Flash**Open result
$0.361359m 59.2sVision expPARTIAL
Record details
Existing public result · PARTIAL
No additional public caveat was supplied by this result record.
Metric scopes materially disagree. Clip handoff: $0.3613, 59m 59.2s. Existing result receipt: $0.46 Provider-recorded OpenRouter cost; Not recorded. Both remain visible; neither silently replaces the other.
Public record identity: DeepSeek V4 Flash · deepseek-v4-flash-vision-exp-openrouter · Provider not separately recorded

GLM 5.3 Flash · courtyard-blossom-stage-2 presentation recording poster — **GLM 5.3 Flash**Open result
$0.78931h 37m 48.2sMaxFAIL
Record details
Existing public result · FAIL
No additional public caveat was supplied by this result record.
Metric scopes materially disagree. Clip handoff: $0.7893, 1h 37m 48.2s. Existing result receipt: $0.68 Provider-recorded OpenRouter cost; 1h 45m 16.069s wall-clock. Both remain visible; neither silently replaces the other.
Public record identity: GLM 5.3 Flash · glm-5.3-flash-max-opencode · Provider not separately recorded

Gemini 3.8 Flash · courtyard-blossom-stage-2 presentation recording poster — **Gemini 3.8 Flash**Open result
$5.0247m 18.7sHighestpass_with_visual_caveat
Record details
Existing public result · complete token timing artifact and validation ledger
No blind visual quality score is attached.
Clip handoff: $5.02, 47m 18.7s. Existing result receipt: $5.02 Provider-recorded usage estimate, not an itemized subscription cash charge; 2838.7s wall-clock. More precise clip formatting is preserved without changing the receipt.
Public record identity: Gemini 3.8 Flash · Gemini 3.8 Flash · OpenRouter · OpenCode 1.18.23

Fixture 05 · Interactive Voxel Diorama

Interactive voxel diorama

Qwen was the editorial favorite in the interactive voxel-diorama round.

DeepSeek V4.1 Flash · interactive-voxel-diorama presentation recording poster — **DeepSeek V4.1 Flash**Open result
$0.49251h 12m 32.3sVision expPASS
Record details
Existing public result · PASS
No additional public caveat was supplied by this result record.
Metric scopes materially disagree. Clip handoff: $0.4925, 1h 12m 32.3s. Existing result receipt: $0.18 Provider-recorded OpenRouter cost; Not recorded. Both remain visible; neither silently replaces the other.
Public record identity: DeepSeek V4 Flash · deepseek-v4-flash-vision-exp-openrouter · Provider not separately recorded

GLM 5.3 Flash · interactive-voxel-diorama presentation recording poster — **GLM 5.3 Flash**Open result
$0.52241h 52m 57.216sMaxPASS
Record details
Existing public result · PASS
No additional public caveat was supplied by this result record.
Clip handoff: $0.5224, 1h 52m 57.216s. Existing result receipt: $0.52 Provider-recorded OpenRouter cost; 1h 52m 57.216s wall-clock. More precise clip formatting is preserved without changing the receipt.
Public record identity: GLM 5.3 Flash · glm-5.3-flash-max-opencode · Provider not separately recorded

Gemini 3.8 Flash · interactive-voxel-diorama presentation recording poster — **Gemini 3.8 Flash**Open result
$0.53272h 2m 44.5sHighestpass
Record details
Existing public result · complete token timing artifact and validation ledger
No blind visual quality score is attached.
Clip handoff: $0.5327, 2h 2m 44.5s. Existing result receipt: $0.53 Provider-recorded usage estimate, not an itemized subscription cash charge; 7364.5s wall-clock. More precise clip formatting is preserved without changing the receipt.
Public record identity: Gemini 3.8 Flash · Gemini 3.8 Flash · OpenRouter · OpenCode 1.18.23

Fixture 06 · JRPG

JRPG boss battle

The JRPG attempts split across completion, validation, and visual quality rather than producing one universal model ranking.

DeepSeek V4.1 Flash · jrpg-boss-battle presentation recording poster — **DeepSeek V4.1 Flash**Open result
$0.156134m 6.7sVision expPASS
Record details
Existing public result · PASS
No additional public caveat was supplied by this result record.
Metric scopes materially disagree. Clip handoff: $0.1561, 34m 6.7s. Existing result receipt: $0.23 Provider-recorded OpenRouter cost; Not recorded. Both remain visible; neither silently replaces the other.
Public record identity: DeepSeek V4 Flash · deepseek-v4-flash-vision-exp-openrouter · Provider not separately recorded

GLM 5.3 Flash · jrpg-boss-battle presentation recording poster — **GLM 5.3 Flash**Open result
$0.157152m 42.980sMaxPARTIAL
Record details
Existing public result · PARTIAL
No additional public caveat was supplied by this result record.
Clip handoff: $0.1571, 52m 42.980s. Existing result receipt: $0.16 Provider-recorded OpenRouter cost; 52m 42.980s wall-clock. More precise clip formatting is preserved without changing the receipt.
Public record identity: GLM 5.3 Flash · glm-5.3-flash-max-opencode · Provider not separately recorded

Gemini 3.8 Flash · jrpg-boss-battle presentation recording poster — **Gemini 3.8 Flash**Open result
$1.1225m 30.7sHighestpass
Record details
Existing public result · complete token timing artifact and validation ledger
No blind visual quality score is attached.
Clip handoff: $1.12, 25m 30.7s. Existing result receipt: $1.12 Provider-recorded usage estimate, not an itemized subscription cash charge; 1530.7s wall-clock. More precise clip formatting is preserved without changing the receipt.
Public record identity: Gemini 3.8 Flash · Gemini 3.8 Flash · OpenRouter · OpenCode 1.18.23

Fixture 07 · Neon Gunner

Neon Gunner

Neon Gunner again separated visual preference from recorded functional and validation states.

DeepSeek V4.1 Flash · neon-gunner presentation recording poster — **DeepSeek V4.1 Flash**Open result
$0.343653m 44.4sVision expPARTIAL
Record details
Existing public result · PARTIAL
No additional public caveat was supplied by this result record.
Metric scopes materially disagree. Clip handoff: $0.3436, 53m 44.4s. Existing result receipt: $0.54 Provider-recorded OpenRouter cost; Not recorded. Both remain visible; neither silently replaces the other.
Public record identity: DeepSeek V4 Flash · deepseek-v4-flash-vision-exp-openrouter · Provider not separately recorded

GLM 5.3 Flash · neon-gunner presentation recording poster — **GLM 5.3 Flash**Open result
$0.79263h 9m 20.539sMaxPENDING_JUDGE
Record details
Existing public result · PENDING JUDGE
No additional public caveat was supplied by this result record.
Clip handoff: $0.7926, 3h 9m 20.539s. Existing result receipt: $0.79 Provider-recorded OpenRouter cost; 3h 9m 20.539s wall-clock. More precise clip formatting is preserved without changing the receipt.
Public record identity: GLM 5.3 Flash · glm-5.3-flash-max-opencode · Provider not separately recorded

Gemini 3.8 Flash · neon-gunner presentation recording poster — **Gemini 3.8 Flash**Open result
$2.5924m 23.9sHighestpass_with_nonbenchmark_caveat
Record details
Existing public result · complete token timing artifact and validation ledger
No blind visual quality score is attached.
Clip handoff: $2.59, 24m 23.9s. Existing result receipt: $2.59 Provider-recorded usage estimate, not an itemized subscription cash charge; 1463.9s wall-clock. More precise clip formatting is preserved without changing the receipt.
Public record identity: Gemini 3.8 Flash · Gemini 3.8 Flash · OpenRouter · OpenCode 1.18.23

Fixture 08 · Sekiro

Sekiro main menu

The menu recreations differ in motion, composition, and finish; their public records preserve the evidence gaps.

DeepSeek V4.1 Flash · sekiro-main-menu presentation recording poster — **DeepSeek V4.1 Flash**Open result
$0.331758m 15.4sVision expPARTIAL
Record details
Existing public result · PARTIAL
No additional public caveat was supplied by this result record.
Metric scopes materially disagree. Clip handoff: $0.3317, 58m 15.4s. Existing result receipt: $0.49 Provider-recorded OpenRouter cost; Not recorded. Both remain visible; neither silently replaces the other.
Public record identity: DeepSeek V4 Flash · deepseek-v4-flash-vision-exp-openrouter · Provider not separately recorded

GLM 5.3 Flash · sekiro-main-menu presentation recording poster — **GLM 5.3 Flash**Open result
$0.10635m 18.685sMaxPARTIAL
Record details
Existing public result · PARTIAL
No additional public caveat was supplied by this result record.
Clip handoff: $0.106, 35m 18.685s. Existing result receipt: $0.11 Provider-recorded OpenRouter cost; 35m 18.685s wall-clock. More precise clip formatting is preserved without changing the receipt.
Public record identity: GLM 5.3 Flash · glm-5.3-flash-max-opencode · Provider not separately recorded

Gemini 3.8 Flash · sekiro-main-menu presentation recording poster — **Gemini 3.8 Flash**Open result
$0.740211m 56.6sHighestpass
Record details
Existing public result · complete token timing artifact and validation ledger
No blind visual quality score is attached.
Clip handoff: $0.7402, 11m 56.6s. Existing result receipt: $0.74 Provider-recorded usage estimate, not an itemized subscription cash charge; 716.6s wall-clock. More precise clip formatting is preserved without changing the receipt.
Public record identity: Gemini 3.8 Flash · Gemini 3.8 Flash · OpenRouter · OpenCode 1.18.23

Fixture 09 · Space Flight

Space Flight Stage 1

Qwen’s Space Flight output was one of its strongest campaign showcases.

DeepSeek V4.1 Flash · space-flight-stage-1 presentation recording poster — **DeepSeek V4.1 Flash**Open result
$0.24245m 44.4sMaxnot_run
Record details
Existing public result · partial token timing and artifact ledger
This is a two-user-turn protocol deviation: the original benchmark prompt needed one ‘continue’ rescue/follow-up turn.
Metric scopes materially disagree. Clip handoff: $0.242, 45m 44.4s. Existing result receipt: $0.25 API-equivalent usage accounting, not an itemized subscription cash charge; 5h 0m 0.3s wall-clock. Both remain visible; neither silently replaces the other.
Public record identity: DeepSeek V4 Flash · DeepSeek V4 Flash · OpenCode Go

GLM 5.3 Flash · space-flight-stage-1 presentation recording poster — **GLM 5.3 Flash**Open result
$1.073h 14m 24.2sMaxNo machine-readable validator attached
Record details
Existing public result · partial multi turn token timing and artifact ledger
This is a two-turn, 13-user-message ZCode session, not a one-shot benchmark run.
Metric scopes materially disagree. Clip handoff: $1.07, 3h 14m 24.2s. Existing result receipt: $4.46 UNVERIFIED reference scenario: GLM-5.2/GLM-5.1 official rates applied to GLM-5.3 token counts; not a verified GLM-5.3 charge or subscription marginal-cash invoice; 48m 14.1s wall-clock. Both remain visible; neither silently replaces the other.
Public record identity: GLM 5.3 Flash · glm-5.3-flash · ZCode

Gemini 3.8 Flash · space-flight-stage-1 presentation recording poster — **Gemini 3.8 Flash**Open result
$4.32h 26m 30.3sHighestpass_with_visual_caveat
Record details
Existing public result · complete token timing artifact and validation ledger
No blind visual quality score is attached.
Clip handoff: $4.3, 2h 26m 30.3s. Existing result receipt: $4.30 Provider-recorded usage estimate, not an itemized subscription cash charge; 8790.3s wall-clock. More precise clip formatting is preserved without changing the receipt.
Public record identity: Gemini 3.8 Flash · Gemini 3.8 Flash · OpenRouter · OpenCode 1.18.23

Fixture 10 · Turbofan

Turbofan CAD

DeepSeek delivered the better submitted CAD result. Gemini produced valid solids but did not complete the required delivery; Qwen timed out and GLM truncated without native geometry.

Qwen 3.8 Flash · turbofan-engine presentation recording poster — **Qwen 3.8 Flash**Open result
$0.3553121h timeoutMax0 parts · 0 assembly instances · 0 mates
Record details
Campaign evidence · campaign evidence not harness ledger
Campaign-evidence record; this attempt was not generated from the public Harness ledger.
This permanent campaign-evidence record uses the handoff’s exact displayed cost and wall clock; it is not a generated Harness ledger.
Public record identity: Qwen 3.8 Flash · openrouter/qwen/qwen3.8-flash · OpenRouter · OpenCode 1.18.23

DeepSeek V4.1 Flash · turbofan-engine presentation recording poster — **DeepSeek V4.1 Flash**Open result
$0.37563755m 52.064sMaxThree valid native solids; wrong-sign coupling and LP/HP interference
Record details
Campaign evidence · campaign evidence not harness ledger
Campaign-evidence record; this attempt was not generated from the public Harness ledger.
This permanent campaign-evidence record uses the handoff’s exact displayed cost and wall clock; it is not a generated Harness ledger.
Public record identity: DeepSeek V4.1 Flash · opencode-go/deepseek-v4.1-flash · OpenCode Go · OpenCode 1.18.23

GLM 5.3 Flash · turbofan-engine presentation recording poster — **GLM 5.3 Flash**Open result
$0.21524250m 37sMaxNo source or native geometry delivered
Record details
Campaign evidence · campaign evidence not harness ledger
Campaign-evidence record; this attempt was not generated from the public Harness ledger.
This permanent campaign-evidence record uses the handoff’s exact displayed cost and wall clock; it is not a generated Harness ledger.
Public record identity: GLM 5.3 Flash · opencode-go/glm-5.3-flash · OpenCode Go · OpenCode 1.18.23

Gemini 3.8 Flash · turbofan-engine presentation recording poster — **Gemini 3.8 Flash**Open result
$3.7937501h 0m 4.361sHighFour valid native solids from compiling FeatureScript
Record details
Campaign evidence · campaign evidence not harness ledger
Campaign-evidence record; this attempt was not generated from the public Harness ledger.
This permanent campaign-evidence record uses the handoff’s exact displayed cost and wall clock; it is not a generated Harness ledger.
Public record identity: Gemini 3.8 Flash · gemini-3.8-flash-high · Google · Antigravity CLI

Fixture 11 · Neo-Gothic Storm City

Neo-Gothic Storm City

DeepSeek was among the strongest graphics models in the Neo-Gothic city fixture.

DeepSeek V4.1 Flash · neo-gothic-storm-city presentation recording poster — **DeepSeek V4.1 Flash**Open result
$0.15961h 21m 50.825sVision expINCOMPLETE_EXHAUSTED
Record details
Existing public result · INCOMPLETE EXHAUSTED
No additional public caveat was supplied by this result record.
Metric scopes materially disagree. Clip handoff: $0.1596, 1h 21m 50.825s. Existing result receipt: $0.42 Provider-recorded OpenRouter cost; Not recorded. Both remain visible; neither silently replaces the other.
Public record identity: DeepSeek V4 Flash · deepseek-v4-flash-vision-exp-openrouter · Provider not separately recorded

GLM 5.3 Flash · neo-gothic-storm-city presentation recording poster — **GLM 5.3 Flash**Open result
$0.118235m 58.906sMaxPARTIAL
Record details
Existing public result · PARTIAL
No additional public caveat was supplied by this result record.
Metric scopes materially disagree. Clip handoff: $0.1182, 35m 58.906s. Existing result receipt: $0.45 Provider-recorded OpenRouter cost; 2h 14m 59.899s wall-clock. Both remain visible; neither silently replaces the other.
Public record identity: GLM 5.3 Flash · glm-5.3-flash-max-opencode · Provider not separately recorded

Gemini 3.8 Flash · neo-gothic-storm-city presentation recording poster — **Gemini 3.8 Flash**Open result
$2.5947m 47.906sHighestpass
Record details
Existing public result · complete token timing artifact and validation ledger
No blind visual quality score is attached.
Clip handoff: $2.59, 47m 47.906s. Existing result receipt: $2.59 Provider-recorded usage estimate, not an itemized subscription cash charge; 2867.9s wall-clock. More precise clip formatting is preserved without changing the receipt.
Public record identity: Gemini 3.8 Flash · Gemini 3.8 Flash · OpenRouter · OpenCode 1.18.23

Fixture 12 · Infinite Cathedral

Infinite Cathedral

The Infinite Cathedral attempts split by visual execution, recorded completion, and evidence quality.

DeepSeek V4.1 Flash · infinite-cathedral presentation recording poster — **DeepSeek V4.1 Flash**Open result
$0.217735m 47.810sVision expPENDING_JUDGE
Record details
Existing public result · PENDING JUDGE
No additional public caveat was supplied by this result record.
Metric scopes materially disagree. Clip handoff: $0.2177, 35m 47.810s. Existing result receipt: $0.24 Provider-recorded OpenRouter cost; Not recorded. Both remain visible; neither silently replaces the other.
Public record identity: DeepSeek V4 Flash · deepseek-v4-flash-vision-exp-openrouter · Provider not separately recorded

GLM 5.3 Flash · infinite-cathedral presentation recording poster — **GLM 5.3 Flash**Open result
$0.58951h 26m 0.031sMaxcomplete
Record details
Existing public result · complete
The run required repeated user-message continuations because the OpenCode endpoint frequently exited or stalled before task completion.
Clip handoff: $0.5895, 1h 26m 0.031s. Existing result receipt: $0.59 API-equivalent estimate at current Z.ai list pricing; not a historical invoice; 1h 26m 0.031s wall-clock. More precise clip formatting is preserved without changing the receipt.
Public record identity: GLM 5.3 Flash · glm-5.3-flash · Provider not separately recorded · OpenCode CLI 1.18.18

Gemini 3.8 Flash · infinite-cathedral presentation recording poster — **Gemini 3.8 Flash**Open result
$2.331h 6m 37.863sHighestpass
Record details
Existing public result · complete token timing artifact and validation ledger
No blind visual quality score is attached.
Clip handoff: $2.33, 1h 6m 37.863s. Existing result receipt: $2.33 Provider-recorded usage estimate, not an itemized subscription cash charge; 3997.9s wall-clock. More precise clip formatting is preserved without changing the receipt.
Public record identity: Gemini 3.8 Flash · Gemini 3.8 Flash · OpenRouter · OpenCode 1.18.23

Fixture 13 · Eiffel Tower

Eiffel Tower drawing

Gemini was the editorial winner and the fastest shown Eiffel Tower attempt at 8m 25s.

DeepSeek V4.1 Flash · eiffel-tower-drawing presentation recording poster — **DeepSeek V4.1 Flash**Open result
$0.084118m 31.4sMaxNo machine-readable validation outcome supplied
Record details
Campaign evidence · campaign evidence not harness ledger
Campaign-evidence record; this attempt was not generated from the public Harness ledger.
This permanent campaign-evidence record uses the handoff’s exact displayed cost and wall clock; it is not a generated Harness ledger.
Public record identity: DeepSeek V4.1 Flash · DeepSeek V4.1 Flash · Unavailable · Unavailable

GLM 5.3 Flash · eiffel-tower-drawing presentation recording poster — **GLM 5.3 Flash**Open result
$0.116910m 28.2sMaxNo machine-readable validation outcome supplied
Record details
Campaign evidence · campaign evidence not harness ledger
Campaign-evidence record; this attempt was not generated from the public Harness ledger.
This permanent campaign-evidence record uses the handoff’s exact displayed cost and wall clock; it is not a generated Harness ledger.
Public record identity: GLM 5.3 Flash · GLM 5.3 Flash · Unavailable · Unavailable

Gemini 3.8 Flash · eiffel-tower-drawing presentation recording poster — **Gemini 3.8 Flash**Open result
$1.368m 25sHighEditorial winner; not a machine-validated score
Record details
Campaign evidence · campaign evidence not harness ledger
Campaign-evidence record; this attempt was not generated from the public Harness ledger.
This permanent campaign-evidence record uses the handoff’s exact displayed cost and wall clock; it is not a generated Harness ledger.
Public record identity: Gemini 3.8 Flash · Gemini 3.8 Flash · Unavailable · Unavailable

Fixture 14 · Apple Stem

Apple-stem manipulation

GLM was the only pass. Gemini lifted the apple using disallowed finger surfaces and therefore did not pass.

Qwen 3.8 Flash · apple-stem presentation recording poster — **Qwen 3.8 Flash**Open result
$0.23971h 28m 21.1sMaxCampaign judgment: did not pass
Record details
Campaign evidence · campaign evidence not harness ledger
Campaign-evidence record; this attempt was not generated from the public Harness ledger.
This permanent campaign-evidence record uses the handoff’s exact displayed cost and wall clock; it is not a generated Harness ledger.
Public record identity: Qwen 3.8 Flash · Qwen 3.8 Flash · Unavailable · Unavailable

DeepSeek V4.1 Flash · apple-stem presentation recording poster — **DeepSeek V4.1 Flash**Open result
$0.09921h 28m 20.2sMaxCampaign judgment: did not pass
Record details
Campaign evidence · campaign evidence not harness ledger
Campaign-evidence record; this attempt was not generated from the public Harness ledger.
This permanent campaign-evidence record uses the handoff’s exact displayed cost and wall clock; it is not a generated Harness ledger.
Public record identity: DeepSeek V4.1 Flash · DeepSeek V4.1 Flash · Unavailable · Unavailable

GLM 5.3 Flash · apple-stem presentation recording poster — **GLM 5.3 Flash**Open result
$0.55831h 29m 33.2sMaxCampaign judgment: pass
Record details
Campaign evidence · campaign evidence not harness ledger
Campaign-evidence record; this attempt was not generated from the public Harness ledger.
This permanent campaign-evidence record uses the handoff’s exact displayed cost and wall clock; it is not a generated Harness ledger.
Public record identity: GLM 5.3 Flash · GLM 5.3 Flash · Unavailable · Unavailable

Gemini 3.8 Flash · apple-stem presentation recording poster — **Gemini 3.8 Flash**Open result
$2.691h 9m 1sHighCampaign judgment: did not pass
Record details
Campaign evidence · campaign evidence not harness ledger
Campaign-evidence record; this attempt was not generated from the public Harness ledger.
This permanent campaign-evidence record uses the handoff’s exact displayed cost and wall clock; it is not a generated Harness ledger.
Public record identity: Gemini 3.8 Flash · Gemini 3.8 Flash · Unavailable · Unavailable

Verdict

Capability moved with the task. These cheaper Flash models did not show consistent general-purpose performance, and no single fixture preference supports a universal leaderboard.

Qwen: Games and 3D scenes — several of the strongest game, arena, voxel, and space-flight outputs.
DeepSeek: Graphics and CAD delivery — the better submitted CAD result, with material CAD defects.
GLM: Specific robot success — the only Apple Stem pass and a single-turn Rubik’s Cube pass.
Gemini: Robot speed and drawing — fastest on the single-turn cube and the editorial Eiffel drawing winner.
Limit: No three-move completion — only GLM and Gemini advanced; neither completed the extension.

Method

Every clip is hash-bound to the supplied test/model mapping; costs and wall clocks are shown verbatim.
All 54 supplied attempts resolve to unique permanent result pages — 35 existing records reused, 19 added as campaign evidence.
Qwen and DeepSeek have no fabricated three-move cube runs because they did not advance.
Robot pass/fail labels remain video-campaign judgments where no machine-readable validator exists.
Gemini’s operator-created CAD assembly and presentation remain post-run work, never scored model completion.
No aggregate score, universal winner, or campaign-specific Builder package has been fabricated.

Inspect all public results Explore benchmark projects and reproduction paths Previous: Launch 004 →