Claude 1M Context Performance: Opus vs Sonnet vs Competitors

Long-context performance on the Claude family is not uniform. On MRCR v2, an 8-needle benchmark at 1M tokens that requires connecting multiple facts (harder than single-needle retrieval), Sonnet 4.5 scores 18.5% — near random — and is effectively unusable past about 600K tokens. Opus 4.6 scores 76 to 78.3% on the same benchmark, a 4-9x improvement that Anthropic itself describes as a "qualitative shift." Opus 4.7 ships with what Anthropic's release notes call "retrieval over long contexts completely overhauled," with higher multi-needle scores deep into documents and a +13% gain over 4.6 on a 93-task coding benchmark. Degradation curves matter as much as headline numbers. Opus 4.6 drops only 17 points across the 750K tokens between 256K (93%) and 1M (76%) — a structurally flatter curve than Sonnet 4.5's cliff. Single-needle retrieval at 1M is roughly 90% for Opus 4.6 and higher for 4.7. Competitors look worse on real usable context. GPT-5 claims 400K but is effective at only 50-65% per the RULER benchmark. Gemini 2.5 Pro claims 2M but shows severe degradation past about 600K. As of April 2026, Opus 4.6 and Opus 4.7 are best-in-class for usable long context. Pricing timeline: in August 2025 Sonnet 4 1M was a beta with a 2x surcharge; in February 2026 Sonnet 4.6 and Opus 4.6 shipped with native 1M; in March 2026 1M went GA with the long-context premium removed and flat pricing of $5/$25 per million tokens on Opus and $3/$15 on Sonnet, making a 900K-token Opus research session cost about $5.25; Opus 4.7 followed in April 2026, and the legacy Sonnet 4/4.5 1M beta retires April 30, 2026. See Context Rot in Long AI Coding Sessions: Why Agents Get Worse as Context Fills for why benchmark scores do not directly predict agentic-coding behavior.