업데이트
What is Claude Opus 4.8?
Claude Opus 4.8 is Anthropic's May 28, 2026 Opus model, built for agentic coding, computer use, and enterprise knowledge work. It follows Claude Opus 4.7 by 41 days and runs on UnificAlly as anthropic/claude-opus-4-8. The specs carry over from 4.7: a 1M-token context window, up to 128k output tokens, and adaptive thinking. The gains are in behavior. It scores higher on agentic coding, computer use, and knowledge-work benchmarks, reviews its own code more honestly, and spends fewer thinking tokens on turns that never needed them. Treat it as a real step up from Opus 4.7 for long agent runs, not a redesign of the model.
Key features of Claude Opus 4.8
1M token context window
Large codebases, long agent traces, and dense document sets fit in one call, with a callable max output of 128k tokens.
Stronger agentic coding
It scores 88.6% on SWE-bench Verified and 69.2% on SWE-bench Pro, up from 87.6% and 64.3% on Opus 4.7, with better long-context handling late in long coding runs.
Better computer use
It reaches 83.4% on OSWorld-Verified, ahead of GPT-5.5 at 78.7%, and scores 84% on Online-Mind2Web for browser tasks that click, type, and read screens.
Knowledge work gains
It scores 1890 on GDPval-AA versus 1753 for Opus 4.7, improves Finance Agent v2 to 53.9%, and is the first model past 10% on the Harvey Legal Agent all-pass standard.
More honest code review
It is about four times less likely than Opus 4.7 to let a flaw in its own code pass without comment, and its measured misaligned-behavior rate is well below Opus 4.7 and Sonnet 4.6.
Adaptive thinking and effort control
Thinking triggers only on turns that need it, and effort defaults to high, so simple steps stay fast and hard problems still get deep reasoning in the same model.
What Claude Opus 4.8 is best for
Long-horizon coding agents
Multi-hour sessions across a large repo, with fewer derailments late in the run.
Computer-use agents
Desktop and browser tasks that click, type, and read screens, with clear action limits.
Finance and legal drafts
First-pass memos and analysis built from filings, contracts, and spreadsheets.
Long document review
Contracts, filings, and research packs held together in one context.
Multi-step research
Search, synthesis, and careful citations across many sources.
Code review
Review passes that flag flaws instead of letting them slide.
Claude Opus 4.8 use cases
Build a coding agent that works through a large repo for hours, keeps its plan intact, and flags its own mistakes instead of hiding them. Use it for computer-use agents that operate dense internal tools, fill unforgiving forms, or compare options across browser tabs. It also fits knowledge-work pipelines: turn filings and spreadsheets into a first-pass finance memo, run contract review across a long document set, or draft legal analysis for human check. For research teams, Claude Opus 4.8 is a good base for agents that search, read many sources, and write up findings with usable citations.
Claude Opus 4.8 limitations
Terminal-heavy coding is not a clean win. On Terminal-Bench 2.1 with the same Terminus-2 harness, GPT-5.5 scores 78.2% to Opus 4.8's 74.6%.
Deep science Q&A did not improve. GPQA Diamond comes in at 93.6%, a touch below Opus 4.7 at 94.2%.
Chat preference lags the agentic gains. Arena's Text board ranks claude-opus-4-8-thinking at #9 with 1484±5, below the Opus 4.6 and 4.7 thinking variants. If everyday chat quality is the goal, the older Opus models still rate higher there.
The API rejects temperature, top_p, and top_k, and adaptive thinking is the only thinking mode. Steering happens through prompts and effort, not sampling knobs.
High-stakes finance, legal, medical, and security outputs still need human review. One-pass output is a draft, not final work.
Claude Opus 4.8 vs Claude Opus 4.7
Claude Opus 4.8 is the better agent. It improves SWE-bench Pro from 64.3% to 69.2%, Terminal-Bench 2.1 from 66.1% to 74.6%, OSWorld-Verified from 82.8% to 83.4%, Humanity's Last Exam with tools from 54.7% to 57.9%, and GDPval-AA from 1753 to 1890. Coding tasks use similar token counts, so the extra quality does not cost more output. It is also more honest: about four times less likely to let flaws in its own code pass without comment, with a lower misaligned-behavior rate. One caution if you tuned effort levels against Opus 4.7: the token allocation behind each level changed, so re-baseline cost and latency before adjusting.
When to use Claude Opus 4.8
Use Claude Opus 4.8 when the job is a long-running agent: hours of coding, computer use, document work, or research where follow-through and honest self-checks matter. Pick it over Opus 4.7 for any new agentic build. If the task is short, chat-style, or latency-sensitive, a Sonnet model or an older Opus is usually enough.
Claude Opus 4.8 API examples
Call Claude Opus 4.8 from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model in the UnificAlly API. Full parameter docs live at docs.unifically.com/models/llm/anthropic/claude-opus-4-8.
curl -X POST https://api.unifically.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "anthropic/claude-opus-4-8",
"messages": [
{ "role": "user", "content": "Review this diff and call out any bug it introduces." }
]
}'
The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.
FAQ
Claude Opus 4.8에 대해 자주 묻는 질문
anthropic/claude-opus-4-8, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your UnificAlly API key.
1M tokens, with a callable max output of 128k tokens. That fits large codebases, long agent traces, or many source documents in one run.
Yes. It scores 88.6% on SWE-bench Verified and 69.2% on SWE-bench Pro, and it is about four times less likely than Opus 4.7 to let a flaw in its own code pass without comment.
Yes. It scores 83.4% on OSWorld-Verified and 84% on Online-Mind2Web, so desktop and browser agents fail less often. Still set clear action limits for agents that click and type in real apps.
No. Requests that set temperature, top_p, or top_k to non-default values return a 400 error, same as on Opus 4.7. Steer the output with prompting instead.
Adaptive thinking is the only thinking mode. The model decides per turn whether to reason before answering, and the effort setting defaults to high. Fixed thinking budgets from older Claude models are not supported.
For agentic coding, computer use, and knowledge work, yes. It improves SWE-bench Pro from 64.3% to 69.2%, Terminal-Bench 2.1 from 66.1% to 74.6%, and GDPval-AA from 1753 to 1890, with more honest code review and fewer wasted thinking tokens.
Deep science Q&A did not improve, and chat-style preference boards still favor older Opus models. Keep humans in review for high-stakes finance, legal, medical, and security work.
Claude Opus 4.8 가이드 및 비교
이 모델을 실행하기 전에 가격, 제한 사항, 테스트된 결과물을 확인하세요.
