Benchmarks

OpenGameEval leaderboard update: June 2026

June 8, 2026 · Updated July 25, 2026


Roblox's OpenGameEval leaderboard has seen significant movement since the May update. The headline: Claude Fable 5, Anthropic's newest model, debuted straight at #1 on both leaderboards — 50.34% Pass@1 on the 87-eval code generation set and 64.67% on Debug, a full 8 points clear of the previous Debug leader. Gemini 3.5 Flash also entered the code generation set and holds the top Pass@5 score. Gemini 3.1 Pro, previously cited as the overall leader, has not been evaluated on the expanded 87-eval set and now appears only on the Debug leaderboard.

Code generation leaderboard (87 evals)

The main leaderboard now has seven evaluated models on the full 87-eval set. Two metrics were added alongside this update: Cons@5(success in at least 3 of 5 attempts) and All@5 (success in all 5 attempts), which together give a clearer picture of consistency vs. ceiling performance.

ModelPass@1Pass@5Cons@5All@5Tool err
Claude Fable 550.34%62.07%51.09%39.52%1.40%
Claude Opus 4.648.05%59.77%48.05%38.28%0.71%
Gemini 3.5 Flash48.05%63.22%49.03%33.86%3.30%
Gemini 3 Flash Preview47.82%60.92%48.84%35.12%5.51%
Claude Opus 4.743.45%58.62%43.45%32.18%1.33%
GPT-5.5 (Reasoning: M)40.69%56.32%40.13%30.62%0.91%
GPT-5.4 (Reasoning: M)40.23%55.17%40.00%29.02%1.81%

Debug leaderboard (30 evals)

The Debug leaderboard has expanded to eleven models and now also tracks Cons@5 and All@5. Claude Fable 5 leads every column: Pass@1 (64.67%), Pass@5 (73.33%, tied with GLM 5), Cons@5 (66.09%), and All@5 (54.66%). Behind it, Gemini 3.1 Pro's advantage over GLM 5 on Pass@1 remains narrow (56.67% vs. 56.00%).

ModelPass@1Pass@5Cons@5All@5Tool err
Claude Fable 564.67%73.33%66.09%54.66%1.01%
Gemini 3.1 Pro56.67%70.00%58.36%42.68%5.97%
GLM 556.00%73.33%59.87%33.98%2.39%
Claude Opus 4.752.67%63.33%53.14%43.57%4.26%
GPT-5.4 (Reasoning: M)51.33%63.33%52.08%39.70%2.98%
Gemini 3 Flash Preview51.33%63.33%51.06%43.31%4.58%
Claude Opus 4.650.67%66.67%49.52%40.85%0.96%
GPT-5.5 (Reasoning: M)50.00%66.67%51.02%35.18%1.54%
Gemini 3.5 Flash49.33%70.00%48.46%36.33%3.37%
GPT Codex 5.347.33%70.00%47.90%27.00%3.21%
Claude Sonnet 4.646.00%60.00%46.47%33.87%6.47%

What changed since May

Claude Fable 5: new #1 on both leaderboards

The most significant new entry. Claude Fable 5 tops the code generation set on Pass@1 (50.34%), Cons@5 (51.09%), and All@5 (39.52%) — the first model to cross 50% Pass@1 on the harder 87-eval set. On Debug the gap is even larger: 64.67% Pass@1, 8 points ahead of Gemini 3.1 Pro, with the highest consistency scores on the board (66.09% Cons@5, 54.66% All@5). Its tool error rates (1.40% codegen, 1.01% debug) are among the lowest tested, though Claude Opus 4.6 still holds the absolute lowest on both sets.

Roblox's June 13 detailed Fable comparison adds an important caveat: the aggregate code-generation gains over Opus 4.7 are not statistically significant at this sample size. The clearest improvement appears on bounded debug tasks. Fable uses roughly the same number of tool calls as Opus 4.7 but shifts from scene-tree traversal toward textual code search. We unpack the workflow implications in our OpenGameEval prompting analysis.

Gemini 3.5 Flash: best Pass@5 on code generation

Gemini 3.5 Flash matches Claude Opus 4.6 on Pass@1 (48.05%) and leads all evaluated models on Pass@5 (63.22%) — including Fable 5 (62.07%). In practical terms: if you run the same prompt a few times and take the best result, Gemini 3.5 Flash gives you the highest probability of at least one success. The tradeoff is a higher tool error rate (3.30% vs. Opus 4.6's 0.71%).

Gemini 3.1 Pro: Debug-only

Previous reports described Gemini 3.1 Pro as the overall leader at 55.3% Pass@1 — that figure came from the old deprecated 47-eval leaderboard. On the current 87-eval expanded set, Gemini 3.1 Pro has not yet been evaluated. It is now the second-strongest model on Debug (56.67%, behind Fable 5), but its code generation performance on the harder task set is an open question.

New metrics: Cons@5 and All@5

The leaderboard added two new columns:

Claude Opus 4.7: the detailed review

Roblox published a detailed comparison of Opus 4.6 vs. 4.7 on this eval suite. The apparent Pass@1 drop (48.1% → 43.5%) is not statistically significant (p=0.24). What is significant: Opus 4.7 makes 39% fewer tool calls per task, with the largest reductions in exploration tools (search_game_tree, script_grep, inspect_instance). It tends to fail by stopping too early rather than over-engineering, and it recovers well on well-specified tasks where the target is unambiguous.

Roblox's recommendation: for open-ended discovery tasks (e.g. “remove tutorial assets,” “make lights toggle with day/night”), add explicit instructions for Opus 4.7 to explore the full workspace before acting. For well-defined tasks it outperforms 4.6 on several evals and costs less to run.

New models: GPT-5.5, GPT-5.4, and GPT Codex 5.3

GPT-5.5 (Reasoning: M) entered both leaderboards: 40.69% Pass@1 on code generation and 50.00% Pass@1 on Debug. Its 0.91% tool error rate on code generation is the second-lowest of any model tested, behind only Claude Opus 4.6. GPT-5.4 (Reasoning: M) was also added to the code generation set (40.23% Pass@1, currently last) and its Debug results were re-run: 51.33% Pass@1, up from the previously reported 50.00%.

GPT Codex 5.3 appears on Debug only (47.33% Pass@1) with a strong Pass@5 (70.00%). It has not yet been evaluated on code generation.

What this means for your workflow

Benchmark freshness: Roblox corrected the checks for three evals on July 24. No revised leaderboard run has been published, so the tables above remain the latest official scores rather than estimated adjustments.


Back to BloxBot