Anthropic Launches Claude Opus 5.5 with Lower Costs

Anthropic has introduced Claude Opus 5.5, the first model in its new Claude 5.5 family. The model matches the performance of Claude Fable 5.1 on most tasks while reducing typical running costs by 40 percent compared to Opus 5. Standard pricing is set at $4 for input and $20 for output per million tokens, a 20 percent drop from Opus 5, while prompt caching costs fall 60 percent to $0.20. The model is 30 percent faster, and a fast mode is available at $8 for input and $40 for output to boost speeds up to 2.5 times.

Artificial Analysis ranks Claude Opus 5.5 first in overall intelligence with a score of 58, ahead of OpenAI's GPT-6 Astra. Opus 5.5 leads Astra on the AA-Briefcase, GDPval-AA SciCode, AA-Omniscience Index, and HLE benchmarks, and scored 46 on Omniscience compared to 43 for both Fable 5.1 and Astra. On the Terminal-Bench-Science 0.1 test, Astra scored 63 percent while Opus 5.5 on its extra-high setting reached 62 percent. On WeirdML v3, Astra leads at 42.2 percent, followed by Opus 5.5 at 31.2 percent, Fable 5.1 at 26 percent, and Kimi K3 at 7.3 percent. In vision and 3D tasks, Opus 5.5 scored 73 percent on BenchCAD Vision2Code, or 96.2 percent with tools, compared to Astra's 95.9 percent. It also hit 83 on Furniture Assembly versus Astra's 80, and 64.4 on Chartography compared to Fable's 44.8.

For practitioners, the model addresses common complaints about previous Claude versions by offering more natural, concise writing and stronger agentic coding. Early enterprise testing by Box showed a 63 percent reduction in tokens used, 42 percent less verbosity, and a 30 percent speedup compared to Opus 5. Box reported accuracy gains of 39 percent in financial due diligence, 65 percent in cloud cost analysis, 17 percent in client account analysis, and 15 percent in clinical diagnostics. Meanwhile, OpenAI has responded by cutting prices for its GPT-6 Sol to $2 for input and $10 for output, and Luna to $0.10 for input and $0.50 for output, encouraging developers to mix and match models.