Saad Ullah
⬤ Anthropic launched Claude Opus 4.6, their latest Opus-class model packing a 1M token context window and upgraded reasoning capabilities. The model crushed MRCR v2 testing, which challenges AI to dig out eight specific facts buried deep inside massive prompts.
⬤ On MRCR v2's 256k token test, Claude Opus 4.6 hit between 91.9% and 93.0%. GPT-5.2 managed around 63.9%, while Gemini 3 Pro landed at roughly 45.4%. When they cranked it up to 1M tokens, Claude Opus 4.6 pulled 78.3% in one setup and 76.0% in another—Gemini 3 Pro only scraped 24.5%. GPT-5.2 didn't even show up in the 1M token results.
⬤ The model showed solid gains across other benchmarks too. Terminal-Bench agentic coding jumped from 59.8% to 65.4%. ARC AGI-2 problem solving rocketed from 37.6% to 68.8%. BrowseComp agentic search hit 84.0% versus GPT-5.2's 77.9%. A few metrics stayed flat or dipped slightly—SWE-bench Verified held steady around 80.8%, and Scaled Tool Use dropped from 62.3% to 59.5%.
⬤ The release brings adaptive effort levels and context compaction while boosting large-context processing power. These results show real progress in AI systems built to analyze massive documents, pull scattered information together, and handle complex reasoning across huge inputs.
Saad Ullah
