Claude Opus 4.5
A variant of Anthropic's Claude Opus 4 series of models, intended for AI-assisted programming and complex engineering tasks.
Timeline (last 90 days)
Simon Willison used a natural language prompt to drive Claude Opus 4.5 to implement the four-player map feature, with the prompt requiring the map to be four-way rotationally symmetric, support up to four AI players, be compatible with the non-audio WebSocket multiplayer mode, allow up to four human players to join with empty slots filled by AI.
The prompt used to implement the four-player map simultaneously constrained four levels: map geometry, player count, network protocol compatibility, and matchmaking logic.
Claude Opus 4.5 modified the WebSocket multiplayer networking code in Battledrones v14 development to support four players battling simultaneously.
Battledrones v14's success case was built on the project's already clear architecture, as incremental extension on a mature codebase; for projects starting from scratch, the same prompt may not produce usable results.
In head-to-head user preference comparisons among early testers, 59% preferred Claude Sonnet 4.6 over Claude Opus 4.5.
The model outputs only a risk level (such as RISK_LEVEL: 5) with no explanation; under a prompt of about 7000 tokens, Opus's P50 latency is about 1.5 seconds.
The anti-fraud model that Coinbase fine-tuned from open-source Qwen3.5-9B beat Claude Opus 4.5 on its own risk-control benchmark.
Qwen3-235B performed the worst on injection tests, failing 5 out of 8 attacks; Gemini 3.7 Flash, Claude Haiku 5.5, and Claude Opus 4.5 achieved zero failures across all 8 injection attacks.
On strong models such as Claude Opus 4.5, DeReAct did not improve Pass@1 but made trajectories more evidence-complete and more compliant with constraints.
For Claude Opus 4.5, DeReAct's Pass@1 is roughly on par with the original ReAct baseline.
1 more timeline events
All Facts (10)
Claude Opus 4.5 modified the WebSocket multiplayer networking code in Battledrones v14 development to support four players battling simultaneously.
50%UnverifiedBattledrones v14's success case was built on the project's already clear architecture, as incremental extension on a mature codebase; for projects starting from scratch, the same prompt may not produce usable results.
50%UnverifiedThe prompt used to implement the four-player map simultaneously constrained four levels: map geometry, player count, network protocol compatibility, and matchmaking logic.
50%UnverifiedIn head-to-head user preference comparisons among early testers, 59% preferred Claude Sonnet 4.6 over Claude Opus 4.5.
50%UnverifiedThe model outputs only a risk level (such as RISK_LEVEL: 5) with no explanation; under a prompt of about 7000 tokens, Opus's P50 latency is about 1.5 seconds.
50%UnverifiedQwen3-235B performed the worst on injection tests, failing 5 out of 8 attacks; Gemini 3.7 Flash, Claude Haiku 5.5, and Claude Opus 4.5 achieved zero failures across all 8 injection attacks.
50%UnverifiedOn strong models such as Claude Opus 4.5, DeReAct did not improve Pass@1 but made trajectories more evidence-complete and more compliant with constraints.
50%UnverifiedFor Claude Opus 4.5, DeReAct's Pass@1 is roughly on par with the original ReAct baseline.
50%UnverifiedSimon Willison used a natural language prompt to drive Claude Opus 4.5 to implement the four-player map feature, with the prompt requiring the map to be four-way rotationally symmetric, support up to four AI players, be compatible with the non-audio WebSocket multiplayer mode, allow up to four human players to join with empty slots filled by AI.
50%UnverifiedClaude Opus 4.5 and GPT-5.1 were released in November 2025, and paired with their respective coding agent frameworks (Claude Code and Codex), they evolved coding agents from frequently error-prone to reliable enough for daily use.
50%