Claude Opus 4.5 Launches: Safer SOTA Coding and Agents, Now Cheaper and More Efficient

Anthropic released Claude Opus 4.5, a safer, more capable, and more efficient flagship model for coding, agents, and complex computer use, available across apps, API, and major clouds at reduced pricing. The model posts state-of-the-art results on key benchmarks, offers controllable effort levels for speed/cost versus depth, and shows better token efficiency and long-horizon performance. Upgrades to the developer platform and apps (Claude Code, Chrome, Excel, long-chat) aim to translate these gains into everyday productivity.

Key Points

Claude Opus 4.5 launches with state-of-the-art coding, agentic, and computer-use capabilities, broad availability, and lower pricing ($5/$25 per million tokens).
Token efficiency and control improve via a new effort parameter, context compaction, memory, and advanced tool use—matching or exceeding Sonnet 4.5 with dramatically fewer tokens.
Benchmark results and internal tests show significant gains: SOTA on SWE-bench Verified, leadership on multiple agentic and coding benchmarks, and superior performance on an internal engineering exam.
Safety is emphasized: Opus 4.5 is presented as Anthropic’s most robustly aligned model, with industry-leading resistance to strong prompt injection attacks.
Product ecosystem upgrades (Claude Code Plan Mode and desktop app, long-chat persistence, Chrome and Excel integrations) leverage Opus 4.5’s long-horizon and computer-use strengths.

Sentiment

The community is largely positive about Opus 4.5, with the price reduction generating the most enthusiasm. However, sentiment is tempered by persistent frustrations with rate limiting, skepticism about whether improvements translate to meaningful real-world gains, and concerns about model degradation patterns. Claude maintains strong loyalty among coding-focused developers but faces genuine competitive pressure from Gemini and OpenAI on price, availability, and specific use cases.

In Agreement

The 3x price reduction is the buried lede — Opus stops being reserved for important tasks and becomes viable for production workloads
Token efficiency improvements (76% fewer tokens at medium effort matching Sonnet's best) may matter more than raw benchmark gains
Claude Code remains the best agentic coding tool, with strong speed, tool use, and developer experience compared to Codex and Gemini CLI
SWE-bench Verified approaching 80% represents remarkable progress from roughly 49% a year ago
Removing Opus-specific caps for Max subscribers and making it the default model is a significant product decision
The safety focus and SOTA prompt injection resistance are genuinely significant for anyone deploying agents with tool access
Multi-agent architectures using Opus as orchestrator with cheaper sub-agents show promising efficiency gains

Opposed

Rate limiting remains the primary pain point — multiple users cite it as the reason for switching to Codex or Gemini CLI
Reports of Sonnet 4.5 degradation before the launch fuel suspicions of deliberate nerfing or serving quantized models under load
Real-world improvement feels incremental rather than groundbreaking, especially with large enterprise codebases
AI coding agents still fail hard on moderately-sized tasks spanning multiple files and systems — they only work well for isolated, well-defined problems
The 'most aligned model' framing does heavy lifting without sufficient third-party red team verification
Frontier models are converging toward commodity status with no real switching costs, threatening any moat Anthropic might have
Output token pricing at 5x input creates perverse incentives for verbose model outputs