CodingSoftware & developer tools
Claude Rebuilds a 16,000-Line Program From Its Behavior Alone as Anthropic Restricts Game Cloning in Claude Code
Across four Claude Opus versions, behavior-only reimplementation of one 16,000-line program rose from 15% to 99.95% of tests passed, yet broader benchmarks show far weaker results and Anthropic has begun steering Claude Code away from cloning existing games.
From the Two Minute Papers video of November 10, 2026 and cited sources: On Epoch AI's MirrorCode benchmark, Claude Opus 4.0, 4.1, 4.5 and 4.6 passed 15%, 24%, 63% and 99.95% of hidden tests when rebuilding gotree, a bioinformatics toolkit, without seeing its source. [2] [3] [7]
SOURCES[2] Epoch AI: MirrorCode preliminary results epoch.ai[3] arXiv 2606.30182: MirrorCode arxiv.org[7] Epoch AI: MirrorCode epoch.ai
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- On Epoch AI's MirrorCode benchmark, Claude Opus 4.0, 4.1, 4.5 and 4.6 passed 15%, 24%, 63% and 99.95% of hidden tests when rebuilding gotree, a bioinformatics toolkit, without seeing its source. [2] [3] [7]
- Original source
- Two Minute Papers video, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 10, 2026 video published
- Covered
- Nov 11 edition
- Evidence
- 12 sources across 9 sites; transcript sourced
RisingHow it’s measured
YouTube video · 154.2K views · 2.8K likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
On Epoch AI's MirrorCode benchmark, Claude Opus 4.0, 4.1, 4.5 and 4.6 passed 15%, 24%, 63% and 99.95% of hidden tests when rebuilding gotree, a bioinformatics toolkit, without seeing its source. Epoch estimates the task would take a human engineer 2–17 weeks. Across all 25 target programs, though, the best model scores 56%, and one large attempt cost $2,600 over 19 days. [2] [3] [7]
Under stricter conditions, results drop sharply. ProgramBench gives agents only a compiled binary and documentation, bans decompilation and scores with more than 248,000 fuzz-generated tests. No public model fully solves any of its 200 tasks. Opus 4.7 gets 3.0% of tasks to at least 95% of tests passed. [10]
Related work extends reconstruction beyond whole programs. REMEND, a model of up to 12 million parameters, recovers math equations from compiled ARM and x86 code with 89.8–92.4% accuracy. It works by reading binaries, not by observing how a program behaves. Separately, WebUI-95 rebuilt real web pages that reached a mean visual-similarity score of 0.938 on a 10,000-page sample. [8] [9]
Anthropic added instructions to Claude Code on Opus 5.5 discouraging the recreation of specific existing games. Developers report that the rule also blocked modding, debugging and crash-dump analysis. A later report says Claude now lets users name a game to describe its mechanics, then proposes an original game in that genre, but Anthropic has not announced this change. [4] [5] [6]
Evidence shows agents can now reproduce well-specified, testable programs at high fidelity on some targets.
Read the full assessment
The implication, still inferred, is that observable behavior offers weaker protection for software, while assets, ecosystems and hard-to-specify functionality matter more.
Executive brief
On one task, Claude's ability to rebuild a program without its source code went from almost useless to almost perfect across four model versions. In Epoch AI's MirrorCode preliminary results, Claude Opus 4.0 passed 15% of the tests for gotree, a 16,000-line bioinformatics toolkit. Opus 4.6 passed 99.95%. A Two Minute Papers–style commentary video uses this result and two other papers to ask whether "software is over." Its answer is that the claim is "a bit dramatic." The capability is real but uneven, a harder benchmark shows near-zero full success, and Anthropic has started steering Claude away from cloning games.
What changed and event timeline
Epoch publishes MirrorCode early results
Successive Claude Opus versions climbed from 15% to 99.95% of tests passed on gotree, rebuilt with no access to its source code. Epoch puts a human engineer's time for the task at 2–17 weeks.
MirrorCode paper on arXiv
The benchmark has 25 target programs and was co-developed with METR. The best model scores 56% overall, and one large attempt cost $2,600 over 19 days.
Claude Code restricts game recreation
Anthropic added two sentences to Claude Code's prompt on Opus 5.5 telling it not to recreate specific existing games. Developers reported that legitimate reverse-engineering and crash-analysis work was blocked too.
More detail
Rule reportedly loosened
Foro3D reports that Claude now lets users name a game to describe its mechanics, then proposes an original game in that genre instead. No official Anthropic announcement was found.
Commentary video published
The video argues that capability is rising sharply but that "software is over" overstates it, and it calls for local models.
Capabilities and access
- Behavior-only reimplementation: On MirrorCode, Claude Opus 4.6 and 4.7 each rebuilt gotree (Epoch). Opus 4.7 did it in 14 hours for $251.
- Binary to math equations: REMEND uses a model of up to 12 million parameters and recovers equations from compiled code with 89.8–92.4% accuracy across ARM32, ARM64 and x86-64 (REMEND on Hugging Face).
- Rebuilding web pages: UI2Code_N was applied to the WebUI-75k corpus of real web pages and produced 63,824 rebuilt pages, the WebUI-95 dataset (WebUI-95 at CHI 2026).
Read the full section
- Behavior-only reimplementation: On MirrorCode, Claude Opus 4.6 and 4.7 each rebuilt gotree (Epoch). Opus 4.7 did it in 14 hours for $251.
- Binary to math equations: REMEND uses a model of up to 12 million parameters and recovers equations from compiled code with 89.8–92.4% accuracy across ARM32, ARM64 and x86-64 (REMEND on Hugging Face).
- Rebuilding web pages: UI2Code_N was applied to the WebUI-75k corpus of real web pages and produced 63,824 rebuilt pages, the WebUI-95 dataset (WebUI-95 at CHI 2026).
- Access: Claude Code with Opus 5.5 now carries the game-clone rule.
Technical analysis for researchers and developers
- How MirrorCode scores: An agent probes the original program in a sandbox and must match its output exactly on hidden end-to-end tests.
- ProgramBench, a harder setup: Agents get only a compiled binary plus documentation, and decompilation is banned.
- REMEND method: A transformer encoder-decoder is combined with algorithmic symbolic processing.
Read the full section
- How MirrorCode scores: An agent probes the original program in a sandbox and must match its output exactly on hidden end-to-end tests. Programs are screened for memorization, and inference budgets are far larger than usual (Epoch).
- ProgramBench, a harder setup: Agents get only a compiled binary plus documentation, and decompilation is banned. Scoring uses more than 248,000 fuzz-generated tests across 200 tasks. No public model fully solves any task, and Opus 4.7 gets 3.0% of tasks to at least 95% of tests passed (BenchLM explainer).
- REMEND method: A transformer encoder-decoder is combined with algorithmic symbolic processing. It is trained on synthetic compiled equations and validated on code from real open-source programs. It reads binaries, so it does not infer rules from observing behavior, as the video suggests at 2:24.
- Reproducibility: MirrorCode's code is public (arXiv).
Claims and evidence
- "15% to nearly 100% in nine months" (video, 1:35): Epoch reports 15% → 24% → 63% → 99.95% across Opus 4.0, 4.1, 4.5 and 4.6 on one program (Epoch).
- Equations recovered with about 90% accuracy by a 12M-parameter model (video, 2:02): consistent with REMEND's reported results.
- 0.938 visual similarity with 95% less code (video, 2:35): WebUI-95 reports a mean UIClip score of 0.938 on a 10,000-sample subset (WebUI-95).
Read the full section
- "15% to nearly 100% in nine months" (video, 1:35): Epoch reports 15% → 24% → 63% → 99.95% across Opus 4.0, 4.1, 4.5 and 4.6 on one program (Epoch). Those results are for gotree only. Across the full benchmark the best score is 56%.
- Equations recovered with about 90% accuracy by a 12M-parameter model (video, 2:02): consistent with REMEND's reported results.
- 0.938 visual similarity with 95% less code (video, 2:35): WebUI-95 reports a mean UIClip score of 0.938 on a 10,000-sample subset (WebUI-95). The video's idea that the savings came from stripping out ads and tracking is not stated in the paper.
- The Anthropic restriction: Anthropic confirmed it (80.lv). The reports of over-blocking come only from developers.
Context and prior work
- Earlier work recovered math from executables using symbolic execution instead of a neural model.
- Screenshot-to-code research such as DCGen uses CLIP-based visual similarity metrics (DCGen).
- On the law, the video says software functionality isn't necessarily protected by copyright, while assets like models, textures and characters can be.
Read the full section
- Earlier work recovered math from executables using symbolic execution instead of a neural model. REMaQE recovered a matching equation for all 25,096 test binaries (REMaQE, arXiv).
- Screenshot-to-code research such as DCGen uses CLIP-based visual similarity metrics (DCGen).
- On the law, the video says software functionality isn't necessarily protected by copyright, while assets like models, textures and characters can be. It includes a not-a-lawyer disclaimer (video, 1:00). No legal source was reviewed.
Limitations, safety and contested findings
- Contamination: Models have probably seen the original codebases during training. The authors cannot say how much this contributes to the scores (Epoch).
- Easier-than-normal setting: MirrorCode depends on precise specifications that can be checked automatically, which is unusual in real development (Epoch).
- Conflicting evidence: Full-task success on ProgramBench is 0%, which contrasts sharply with MirrorCode's headline results (BenchLM).
Read the full section
- Contamination: Models have probably seen the original codebases during training. The authors cannot say how much this contributes to the scores (Epoch).
- Easier-than-normal setting: MirrorCode depends on precise specifications that can be checked automatically, which is unusual in real development (Epoch).
- Conflicting evidence: Full-task success on ProgramBench is 0%, which contrasts sharply with MirrorCode's headline results (BenchLM).
- Over-refusal: Developers report the game rule also blocks modding, debugging and crash-dump analysis (Tom's Hardware).
- Sponsorship: The video carries a Lambda sponsorship (5:52).
Business and practitioner implications
- Vendors: If a product's behavior can be fully observed and tested, it can now be imitated by an AI agent.
- Legacy and migration work: A "test oracle" loop (run the original program, compare outputs, iterate) is a working method for porting and modernizing software (video, 3:31).
- Policy risk: Vendor guardrails can change overnight. Teams doing security or reverse-engineering work should plan for fallback models.
Read the full section
- Vendors: If a product's behavior can be fully observed and tested, it can now be imitated by an AI agent. What still protects it is assets, ecosystem lock-in and the parts that are hard to specify.
- Legacy and migration work: A "test oracle" loop (run the original program, compare outputs, iterate) is a working method for porting and modernizing software (video, 3:31).
- Policy risk: Vendor guardrails can change overnight. Teams doing security or reverse-engineering work should plan for fallback models.
- Skills and adoption: The video argues that custom, home-built tools can win if they use file formats others can open. Employers still expect skills in standard tools such as Adobe's apps (4:52).
The source trail.
Sources (12)
Is Software Over?
Transcript retrieved via youtube_auto_captions; language en. Timestamped text, not direct audiovisual review. Source: https://www.youtube.com/watch?v=vug73goPJaM. Automatic captions/transcription may contain errors.
www.youtube.com