BHBench source submissions

This repository contains the preserved model-generated projects behind bhbench.ssh.codes. Each submission is kept in its own folder so the generated source, assets, build files, tests, and existing build artifacts can be inspected as a complete project.

The files inside submissions/ are imported from the preserved submission folders without source edits. The repository-level README, ignore file, and checksum manifest are organizational metadata added for this archive. Nested Git metadata and macOS .DS_Store files are not included.

Live browser adaptations

The thirteen newly imported working renderers now run live on the site, along with the existing Fable 5.1 browser port. Their WebGL2 adaptations, native lookup tables, and browser validation page are in browser/. Original files under submissions/ remain unchanged.

Current models

Benchmark entry	Folder	Notes
Gemini 4 Hydrogen	gemini-4-hydrogen	OpenCode / Vertex AI; high reasoning; Metal compute renderer; live port
Claude Fable 5.2	claude-fable-5.2	Claude Code; xhigh reasoning; 85 native physics checks passed; live Kerr port
Kimi K3.1 Pro	kimi-k3.1-pro	Kimi Code; reasoning setting not recorded; live port with Flamm embedding inset
GPT 6.1 Terra	gpt-6.1-terra-xhigh	Codex; xhigh reasoning; live port
Qwen4	qwen4	Claude Code harness; restart continuation added packaging and docs only; live port
Tencent HY4	tencent-hy4	OpenCode / OpenRouter; high reasoning; live port
DeepSeek V4.1 Flash	deepseek-v4.1-flash	Mistral Vibe; max reasoning; live port with native blackbody table
Inkling 2 Small	inkling-2-small	OpenCode; max reasoning; includes one user Continue turn that edited source; live port
Muse Ember	muse-ember	OpenCode / OpenRouter; max reasoning; live ray trace and Flamm grid
Gemini 4 Argon	gemini-4-argon	OpenCode / OpenRouter; high reasoning; source reproduced from recorded writes; live native presets
Muse Wonder 1.0	muse-wonder-1.0	Muse Code 1.4.6; model identity verified in session records; three live views
StepFun 5 Preview	stepfun-5-preview-high	OpenCode Go; high reasoning; original C++/OpenGL source; four live scenes
Claude Haiku 5.5	claude-haiku-5.5-xhigh	Claude Code; xhigh reasoning; original C++/Metal render; 25 physics checks passed
Claude Opus 5.5	claude-opus-5.5	Preserved C++/Metal source already published on the live site
GPT 6.1 Sol	gpt-6.1-sol-xhigh	Codex; xhigh reasoning
Claude Sonnet 5.5	claude-sonnet-5.5-xhigh	Claude Code; xhigh reasoning
Grok 4.7	grok-4.7	Grok Build; high reasoning
GPT 6 Luna	gpt-6-luna-xhigh	Native render unavailable; original source preserved
Mistral Large 4	mistral-large-4	OpenCode; Default reasoning
MiMo V2.6 Flash	mimo-v2.6-flash	OpenCode; Default reasoning; local model run
Qwen 3.8 27B	qwen-3.8-27b	OpenCode; Default reasoning; local model run
Ling 3.0 Flash	ling-3.0-flash	Native render unavailable; original source preserved
GPT 6 Astra	gpt-6-astra-xhigh	xhigh reasoning, generated with Codex
GLM 5.3	glm-5.3	max reasoning, generated with OpenCode
GLM 5.3 Flash	glm-5.3-flash	Originally submitted as ox-alpha; max reasoning, generated with OpenCode
Qwen 3.8 Flash Next (NVFP4)	qwen-3.8-flash-next-nvfp4	Final standalone local generation with OpenCode
Muse Spark 1.3	muse-spark-1.3-xhigh	xhigh reasoning, generated with Muse Code
Muse Spark 1.3 Max	muse-spark-1.3-max	max reasoning, generated with Muse Code
Composer 2.5	composer-2.5
Gemini 3.8 Flash	gemini-3.8-flash	Original C++/Metal project, generated with OpenCode

Old models

Superseded results are kept here with their original submission paths so existing links continue to work.

Benchmark entry	Folder	Notes
Claude Fable 5.1	claude-fable-5.1	xhigh reasoning, generated with Claude Code
Qwen 3.8 Max 0902	qwen-3.8-max-0902	Generated with OpenCode
DeepSeek V4 Flash	deepseek-v4-flash	OpenCode; Default reasoning
GPT-5.6 Terra	gpt-5.6-terra-xhigh	xhigh reasoning result
Tencent HY3	tencent-hy3
Kimi K3	kimi-k3
Inkling	inkling
GPT 6 Sol	gpt-6-sol-xhigh	Codex; xhigh reasoning
Qwen 3.7 Flash	qwen-3.7-flash	OpenCode; Default reasoning
Claude Fable 5	claude-fable-5-cpp	Default C++ result
Claude Fable 5	claude-fable-5-python	Published Python variant
Claude Opus 5	claude-opus-5
GPT-5.6 Sol	gpt-5.6-sol-xhigh	xhigh reasoning result
GPT-5.6 Sol	gpt-5.6-sol-ultra	ultra reasoning result
GPT-5.6 Sol	gpt-5.6-sol-xhigh-gargantua	Gargantua variant by zoomx64
GLM 5.2	glm-5.2
Qwen 3.8 Max	qwen-3.8-max
Grok 4.6	grok-4.6	High reasoning, generated with Grok Build
Claude Sonnet 5	claude-sonnet-5
GPT-5.6 Luna	gpt-5.6-luna-xhigh	xhigh reasoning result
Muse Spark 1.2	muse-spark-1.2
Gemini 3.7 Flash	gemini-3.7-flash	Generated with OpenCode
Gemini 3.6 Flash	gemini-3.6-flash	Previous-generation entry
Grok 4.5	grok-4.5	Previous-generation entry
Claude Opus 4.8	claude-opus-4.8	Previous-generation entry
Muse Spark 1.1	muse-spark-1.1	Previous-generation entry
Qwen 3.8 Max Preview	qwen-3.8-max-preview	Previous-generation entry
Grok Build 0.1	grok-build-0.1	Previous-generation entry
GPT-5.5	gpt-5.5-xhigh	Previous-generation xhigh result
Gemini 3.5 Flash	gemini-3.5-flash	Previous-generation failed build, preserved as submitted

Scope

BHBench compares one-shot black-hole simulation projects produced by different models. The interactive browser adaptations used by the live website are maintained separately; this repository is the source-submission archive and does not normalize the projects into one build system.

Build and runtime requirements differ by submission. Consult the README and build files inside a folder before running it. No repository-wide license is applied to the submitted projects.

October 6, 2026 refresh

Added 13 distinct model submissions and the missing Claude Opus 5.5 source. Separate quantization and harness experiments of existing models are not additional entries. New imports preserve source, assets, tests and build instructions; generated CMake build directories and nested Git metadata are excluded. Failed native rendering is recorded without repairing the model output. Catalog metadata records the new submissions and current/old grouping.

October 7, 2026 addition

Added Claude Haiku 5.5 from the original Claude Code run at xhigh effort. Its complete submitted project, seven renders and existing binaries are preserved without source or shader edits. The native validation executable passed all 25 physics checks. The distinct-model catalog now has 26 current models and 17 old models.

October 9, 2026 addition

Added StepFun 5 Preview from the original OpenCode Go run at high reasoning. All 121 assistant responses identify step-5-preview-free. The original project and four native captures are preserved without source edits, excluding generated CMake build files. Its submitted ray tracer, spacetime grid, lensing view and Flamm funnel run live in the browser. The distinct-model catalog now contains 25 current models and 17 old models (42 total).

October 12, 2026 addition

Added Muse Wonder 1.0 from its original Muse Code session. The selected model and all 72 root-session completions identify muse-wonder-1.0. Source and deterministic native captures are preserved without repairs; generated CMake files are excluded. Its submitted ray tracer and separate polar grid run live in all three original views. The distinct-model catalog contains 26 current models and 17 old models (43 total).

October 13, 2026 addition

Added Gemini 4 Argon from the original OpenCode/OpenRouter run at high reasoning, confirmed in all 44 assistant responses. Replaying the recorded writes and edits exactly reproduces all eight submitted source, build and documentation files. Its original source, executable and native capture are preserved without repairs. The run used 100,617 output tokens including reasoning, and reports about $2.45 of cost. The browser runs its submitted ray tracer, exported native funnel and photon paths, original bloom passes and five original presets plus the default view. The distinct-model catalog contains 27 current models and 17 old models (44 total).

November 9, 2026 addition

Added eight submissions: Claude Fable 5.2, Kimi K3.1 Pro, GPT 6.1 Terra, Qwen4, Tencent HY4, DeepSeek V4.1 Flash, Inkling 2 Small and Muse Ember. Each complete project is preserved without source edits; generated CMake build directories, nested Git metadata and .DS_Store files are excluded. All eight built and rendered on an Apple M4 Max, and each submitted ray tracer runs live in the browser. Output tokens, cost and rates come from each harness's session records; figures the records do not contain are marked as not recorded rather than estimated.

Follow-up turns were checked. Fable 5.2's later messages only reopened its window, and Qwen4's restart continuation added packaging and documentation without renderer changes. Inkling 2 Small's run includes one user "Continue" turn that edited src/Shaders.h, src/main.cpp and README.md; the archived project is the tree after that turn, as its record states.

These eight entries are not yet scored; the October 13 judge scores are unchanged. Seven superseded entries moved to Old models: Claude Fable 5.1, Qwen 3.8 Max 0902, DeepSeek V4 Flash, GPT-5.6 Terra, Tencent HY3, Kimi K3 and Inkling. The distinct-model catalog contains 28 current models and 24 old models (52 total).

November 12, 2026 addition

Added Gemini 4 Hydrogen, run through OpenCode on Google Vertex AI at high reasoning, confirmed in all 68 assistant responses. The run had no follow-up messages. Its complete project is preserved without source edits, excluding the generated build directory. It built and passed its own tests on an Apple M4 Max, and its Metal compute ray tracer, Kerr spacetime grid and bloom run live in the browser. The run used 118,559 output tokens including reasoning. No verifiable Vertex price for Hydrogen was published at the time, so cost is not recorded. The entry is not yet scored and supersedes no other entry. The distinct-model catalog contains 29 current models and 24 old models (53 total).