Output slot
GPT-5.6 Sol
No generated artifact is being simulated
GPT-5.6 Sol
OpenAI
openai/gpt-5.6-sol
CresciBenchPreparing runs
CresciBench Core
See what every model builds under pressure
Compare complete games, apps, landing pages, data products, and defensive security labs by challenge, sector, or model. Versioned evidence keeps capability, reliability, risk, visual quality, latency, and cost separate.
Public evidence contract
The API exposes status for withheld runs, never synthetic model aggregates. A live ranking requires a complete verified release using the clean external core-v2 catalog, explicit approval, and an exact report-and-artifact evidence seal. The 20 public prompts and their canonical repetition-1 responses are disclosed only after that gate; there is no reserved payload-membership index. Private payloads, full artifacts, and tool traces remain outside Git and Vercel.
Core report JSONEvidence explorer
See what every model builds under pressure
Compare one-attempt games, product interfaces, landing pages, data products, and defensive security labs. Every output is judged against the same published stress contract; empty cards are explicit run slots, never simulated results.
- Models
- 11
- Stress tests
- 36
- Output slots
- 396
- Published
- 0
- One attempt
- 12k token cap
- Offline artifact
- Responsive
- Failure states
- 100-point rubric
Core release: emptyOpen official leaderboard
Games
extreme stressVoxel World
Minecraft-style browser game
Playable browser buildKeyboard and pointer controlsResponsive HUD and pause stateSource artifact
Pressure applied
Interaction depthOffline runtimePerformance pressureState recoveryResponsive layout
Browser game rubric
Version 1 · 100 pointsGameplay completeness
35pts
Interaction quality
20pts
Runtime resilience
15pts
Visual craft
15pts
Accessibility and responsiveness
15pts
Evaluation gates
- The playable build starts offline without external requests or console errors.
- Movement, camera, world interaction, pause, and restart states all remain operable.
- Keyboard focus, instructions, HUD, and status feedback remain legible at mobile and desktop widths.
Run not started
Output slot
GPT-5.6 Terra
No generated artifact is being simulated
GPT-5.6 Terra
OpenAI
openai/gpt-5.6-terra
Output slot
Claude Opus 5
No generated artifact is being simulated
Claude Opus 5
Anthropic
anthropic/claude-opus-5
Output slot
Claude Sonnet 5
No generated artifact is being simulated
Claude Sonnet 5
Anthropic
anthropic/claude-sonnet-5
Output slot
Gemini 3.1 Pro Preview
No generated artifact is being simulated
Gemini 3.1 Pro Preview
google/gemini-3.1-pro-preview
Output slot
Gemini 3.6 Flash
No generated artifact is being simulated
Gemini 3.6 Flash
google/gemini-3.6-flash
Output slot
Grok 4.5
No generated artifact is being simulated
Grok 4.5
xAI
x-ai/grok-4.5
Output slot
DeepSeek V4 Pro
No generated artifact is being simulated
DeepSeek V4 Pro
DeepSeek
deepseek/deepseek-v4-pro
Output slot
Qwen 3.7 Max
No generated artifact is being simulated
Qwen 3.7 Max
Qwen
qwen/qwen3.7-max
Output slot
Llama 4 Maverick
No generated artifact is being simulated
Llama 4 Maverick
Meta
meta-llama/llama-4-maverick
Output slot
Mistral Large 2512
No generated artifact is being simulated
Mistral Large 2512
Mistral AI
mistralai/mistral-large-2512
Three surfaces, three different questions
Keeping the tracks separate prevents a polished interface from masking weak reasoning, unsafe actions, or unreliable task execution.
Core
Model-only tasks with deterministic checks wherever possible. The capability score excludes agent execution. JavaScript checks require explicit use of a disposable, network-denied Docker sandbox; they never execute inside the application process.
Agent
Tool use and recovery through bounded, read-only virtual tools and explicit budgets. Agent score and reliable agent success are reported separately.
Design
Screenshots and generated interfaces are scored separately against a visual rubric. Design never contributes to Core and remains unranked until a live, rubric-scored release exists.
Open the Design suiteClean external release contract
The clean core-v2 secret pack and public commitments are prepared, but no live ranking has been run. Release still needs provider credentials, a candidate and two independent judge models, a finite approved worst-case budget, the complete 96-task run, verification, and an exact approval evidence seal. Synthetic reports validate only the pipeline and never establish a ranking.
- Public disclosure
- 20 complete public tasks; the 76 exposed reserved tasks are compromised, synthetic/dev-only, and never eligible for a live release.
- Public result gate
- Live, complete, publishable, explicitly approved, and exact report plus artifact-evidence match.
- Code execution
- Explicit opt-in only: digest-pinned, network-denied Docker containers enforce CPU, memory, process, time, output, and filesystem limits with no host mounts.
- Bounded I/O
- Provider bodies cap at 8 MiB. Attachments and artifacts require explicit canonical roots, realpath containment, size checks, and no symlinks.
Results by section
Choose a capability to compare models, then inspect canonical public task responses. Agent and Design remain separate from Overall Core.
Equal scores share the same rank.
Overall Core
Weighted capability across the seven model-only tracks. Agent and Design remain separate.
Awaiting live release
Rankings, model names, winners, and responses remain hidden until a complete live release passes publication, approval, and integrity checks. You can still explore every section above without synthetic results being presented as a model comparison.