Claude Opus 4.7 hits 80.5% on SWE-bench and leads GPT-5.4 on GDPVal but eats 1.35x more tokens than Opus 4.6
decrypt.co •
Revision history
2 recorded changes
Want your article here?
Promote with Leviathan NewsAnthropic shipped Claude Opus 4.7 with solid benchmark gains — 80.5% on SWE-bench Multilingual (up from 77.8%), 80.6% on OfficeQA Pro vs. GPT-5.4's 51.1% and Gemini 3.1 Pro's 42.9%, plus a 1,753 GDPVal Elo that clears GPT-5.4's 1,674. The catch: a new tokenizer inflates usage 1.0-1.35x and the model autonomously rewrites code, so Decrypt's reviewer burned through a full session quota on one coding task. Pricing holds at $5/$25 per million input/output tokens on Claude.ai, API, Bedrock, Vertex AI, and Microsoft Foundry.
TLDR by @Benthic
More on GPT
OpenAI launches GPT-Rosalind for life sciences and drug discovery, with Amgen, Moderna, and Allen Institute on early access
Openai ·
OpenAI expands cybersecurity push with GPT-5.4-Cyber access for US and UK agencies, partnering with banks and tech giants to strengthen global AI defense systems
𝕏/@OpenAINewsroom ·
ZetaChain 2.0 powers Anuma, the first AI app enabling users to import and encrypt their entire chat history while preserving private memory across models like GPT and Claude.
𝕏/@ZetaChain ·
OpenAI unveils GPT-5.4 with reasoning and agentic upgrades; users can now interrupt and redirect model mid-response
Openai ·
2025 wasn’t about one “best” LLM—it was about stacks. Users mixed models by task: Claude for coding, GPT-5.x for agents/chat, DeepSeek/Qwen for cheap scale, Gemini for research, and niche tools for creativity.
decrypt.co ·
OpenAI releases GPT Image 1.5
Openai ·
OpenAI launches GPT-Rosalind for life sciences and drug discovery, with Amgen, Moderna, and Allen Institute on early access
Openai ·
OpenAI expands cybersecurity push with GPT-5.4-Cyber access for US and UK agencies, partnering with banks and tech giants to strengthen global AI defense systems
𝕏/@OpenAINewsroom ·
ZetaChain 2.0 powers Anuma, the first AI app enabling users to import and encrypt their entire chat history while preserving private memory across models like GPT and Claude.
𝕏/@ZetaChain ·
OpenAI unveils GPT-5.4 with reasoning and agentic upgrades; users can now interrupt and redirect model mid-response
Openai ·
2025 wasn’t about one “best” LLM—it was about stacks. Users mixed models by task: Claude for coding, GPT-5.x for agents/chat, DeepSeek/Qwen for cheap scale, Gemini for research, and niche tools for creativity.
decrypt.co ·
OpenAI releases GPT Image 1.5
Openai ·