Claude Opus 4.8 vs Composer 2.5

Claude Opus 4.8 launched: code defect omission rate reduced to 25% of Opus 4.7’s, hallucination probability dropped to 10%; new Dynamic Workflows enable hundreds of sub-agents in parallel—Bun migration case produced 750K lines of Rust with 99.8% test pass rate.

Opus 4.8 scores 69.2% on Sweet Bench Pro—~5 pts above Opus 4.7, ~10 above GPT-4.5—but real-world coding reveals persistent ‘last 10%’ failures and hallucinations; pricing is steep at $5/k input tokens.

llm-anthropic 0.25.1 发布，新增 Claude Opus 4.8 模型及快速模式选项，优化默认最大输出令牌数。

Anthropic raised $65B in Series H at a $965B post-money valuation, with $47B annualized revenue; simultaneously launched Claude Opus 4.8 (fixing 4.7 issues, SOTA on economic benchmarks) and Dynamic Workflows (ultracode), enabling hundreds of parallel subagents for coding—demonstrated by rewriting 750k LOC of Bun in 6 days.

AI boosts white-collar productivity, sparking 4-day workweek proposals—but gains mostly captured by capital; YouTube auto-labels realistic AI videos; Opus 4.8 shows modest improvements, with community favoring GRAM-enhanced small models; LLM fact-checking remains inconsistent; Win10 can run SimCity 3000 at 4K.

Anthropic released Claude Opus 4.8, significantly improving performance: 69.2% on SWE-bench Pro (+5 pts vs 4.7), 2.5× faster inference (~250 tokens/sec), plus new dynamic workflows and long-horizon autonomy—all at the same price.

Claude Opus 4.8 introduces Ultra Code effort level and enhanced agents, enabling long-running sessions, hundreds of parallel sub-agents, output self-verification, and end-to-end codebase migrations across 100k+ lines; its ‘honesty’ manifests in disclosing limitations and hiding features like Ultra Code by default.

iOS 27 reveals dual-entry Siri with standalone app; Claude Opus 4.8 cuts fast-mode cost to one-third and reduces undetected code defects to 1/4 of Opus 4.7; HarmonyOS ecosystem exceeds 1.3 billion devices; DeepSeek suffered a 22-minute outage; Xiaomi ranks 7th globally in NEV sales, surpassing Volkswagen and Toyota.

The article discusses Cursor's release of the Composer 2.5 model, claiming it is currently the best coding model with excellent performance and cost-effectiveness.

The "smart, fast, cheap" trilemma limitation of AI models has been broken by Cursor's Composer 2.5, which can simultaneously achieve all three characteristics.

Cursor has released an SDK supporting Python and TypeScript for building custom AI agents using Composer 2.5, offering a 90% discount on Composer usage during a limited-time promotion.

The Cursor team achieved Composer 2.5 through reinforcement learning training rather than prompt engineering, with their large-scale RL program running inference on Fireworks, indicating that self-trained models will be the only way to maintain competitive moats after 2027.

Cursor released Composer 2.5, using Kimi as a base with 85% compute for self-training. It matches Claude Opus 4.7 performance at 1/10th the cost via targeted RL and 25x synthetic data.

The Cursor SDK now supports programmatic access to Composer 2.5's frontier intelligence features via Python and TypeScript, allowing developers to build their own AI agents with 90% discount during the long weekend.

Composer 2.5 is having a moment. Worth a look at how the team actually got here.

Cursor's self-developed coding model Composer 2.5 matches Opus 4.7 performance with less than 1-point score gap, but costs 10-30x less: input tokens drop from $15 to ~$1.5 per million, output from $75 to ~$2.5. It shows clear improvements over Composer 2 in long-context tasks, complex instruction following, and collaboration smoothness.