ক্লড অ'পাছ 4.8

The previous Opus flagship, built for complex agentic coding and enterprise work. It improved on Opus 4.7 across benchmarks at the same price, and early testers found it more reliable and sharper in its judgment on agentic tasks. It offers the same low-to-max effort ladder as Opus 5, but thinking is off by default here and must be requested.

Opus 4.8 was the top of Anthropic's line for about two weeks. It shipped in late May 2026, Claude Fable 5 arrived on 9 June and Claude Opus 5 on 24 July, so the model now reads as the last of the previous Opus generation rather than a current flagship - which is also how Anthropic framed it at the time, calling it a modest but tangible improvement on Opus 4.7 and saying openly that a cheaper Opus-capability tier and a class above Opus were both in progress. Both arrived.

The numbers Anthropic published are incremental and consistent: 69.2% on SWE-bench Pro against 64.3% for Opus 4.7, 83.4% on OSWorld-Verified against 82.8%, 1890 on GDPval-AA against 1753, 53.9% on Finance Agent v2 against 51.5%, and 49.8% on Humanity's Last Exam without tools and 57.9% with them against 46.9% and 54.7%. The price did not move from Opus 4.7's.

What the lab put first was not a score but honesty. Models tend to jump to conclusions and claim progress on thin evidence; Anthropic reports Opus 4.8 as around four times less likely than Opus 4.7 to let flaws in code it wrote itself pass unremarked, and early testers described a model that flags its own uncertainty, asks the right questions, and pushes back on a plan that does not hold together. The alignment assessment found substantially lower rates of misaligned behaviour than Opus 4.7, close to Claude Mythos Preview, alongside new highs on the lab's measures of prosocial traits.

Opus 4.8 is also where the effort control became a user-facing dial in Anthropic's own apps rather than an API-only setting. The lab ships the model at high effort as its own default, which it judges the best overall balance of quality and experience, and notes that on coding tasks that setting spends about as many tokens as Opus 4.7's default while scoring better; it recommends the extra rung for difficult tasks and long-running asynchronous workflows. Thinking is off here unless a call asks for it, so the ladder is something to reach for rather than something to turn down.

Two other things launched with it that say what the model was built for: dynamic workflows in Claude Code, which let Claude plan a large job, run hundreds of parallel subagents in one session and verify its own output before reporting back, aimed at codebase-scale migrations measured in hundreds of thousands of lines; and a fast mode running at 2.5 times the speed, which Anthropic priced at a third of what the equivalent cost on previous models.