Anthropic Unveils Claude Opus 5, a Fresh Challenger to GPT‑5.6

Anthropic has just rolled out Claude Opus 5, its latest large‑language model aimed squarely at developers, researchers and anyone who needs a reliable AI partner for complex, multi‑step tasks. While it sits just below the flagship Fable 5 in Anthropic’s own hierarchy, the company says Opus 5 is built to be the workhorse that most professionals will actually use day‑to‑day.

One of the headline claims is a noticeable jump in coding ability. In Anthropic’s internal Frontier‑Bench tests, Opus 5 delivered more than twice the performance of its predecessor, Opus 4.8, on software‑engineering prompts while using fewer compute resources. When pushed to its highest effort setting on the CursorBench 3.2 suite, it flirted with Fable 5’s top score – coming within a whisker (0.5 %) – yet it still cost roughly half per completed task.

What makes Opus 5 especially interesting is the new “adjustable effort” knob. Developers can now pick from low, medium, high, x‑high or max effort levels, letting them trade off speed, token consumption and raw capability on a per‑request basis. Need a quick answer or a short snippet? Dial it down. Tackling a thorny debugging session or a deep research query? Crank it up.

Another practical upgrade is the expanded context window. Opus 5 can hold up to one million tokens in memory, and it can generate responses that stretch to 128,000 tokens. In real terms, that means you can feed it massive codebases, long research papers or sprawling conversation histories without chopping them up.

Anthropic is also tinkering with the way its models interact with tools during a chat. In beta, developers can add, swap or drop tools on the fly without breaking the existing prompt cache. A companion beta feature, “automatic fallbacks,” silently reroutes a request that hits a safety block to a fallback model instead of throwing a hard error. Both ideas aim to make AI‑driven agents sturdier in production.

Pricing, of course, is where the rivalry with OpenAI becomes stark. Claude Opus 5 is priced at $5 per million input tokens and $25 per million output tokens – the same rates as the older Opus 4.8. By contrast, Anthropic’s own Fable 5 commands $10 and $50 respectively, while OpenAI’s GPT‑5.6 sits even higher. For developers who need the extra horsepower only on occasional peaks, there’s also a “Fast” mode that runs about 2.5× quicker for double the standard rate.

Claude Opus 5 is already available through Anthropic’s paid tiers – Claude Code, Claude Cowork and the Claude API – and it has become the default model for Claude Max subscribers. Whether it can consistently out‑perform GPT‑5.6 in diverse, real‑world workloads will still need validation from independent benchmarks, but the early numbers suggest a compelling alternative for anyone watching costs as closely as capabilities.