Anthropic
Claude Haiku 5.5
最大コンテキスト
1M
最大出力
128K
リリース日
N/A
claude-opus-4.8Claude Opus 4.8 は、参照元で強調されている特徴として premium Opus-tier model, 複雑な推論, 1M-コンテキスト ワークフロー and high-autonomy agents を持ちます。主な価値は difficult 専門タスク, 長期 コーディング and 高度な 知識作業 にあります。そのため、汎用チャットモデルとしてではなく、このモデル固有の強みを中心に説明するのが自然です。
Context Window
1M tokens
Maximum Output
128K tokens
Release Date
2026年5月28日
Modalities
| 入力価格 | 出力価格 | キャッシュ読み取り |
|---|---|---|
| $5/M | $25/M | $0.5/M |
Reasoning
Tool calling
Temperature parameter
Attachments
Knowledge Base
Endpoint Protocols
Claude Opus 4.8は、長時間の専門タスクに向けて、信頼性の高いAgent判断、効率的なソフトウェア開発、マルチモーダル分析を重視しています。
複数段階の作業を計画し、進捗を自己点検し、重要な操作の前に不確実性を示します。
コーディング作業を一貫して完遂し、少ない手順でツールを使いながら自身の実装上の欠陥を検出します。
PDFや図表などの非構造化資料を横断して推論し、長いセッションでも文脈とスタイルを維持します。
Claude Opus 4.8は、自律的なソフトウェア開発、根拠重視の調査、一貫した実行が必要なコンピューター操作に適しています。
複数サービスにまたがる複雑な変更を計画・実装・検証し、調査からテスト済みコードまで進めます。
情報量の多い文書と複合メディアを確認し、根拠を追跡して構造化された分析報告を作成します。
ブラウザーベースの手順を進め、変化する画面を解釈して複数段階の操作を完了します。
import OpenAI from "openai"
const client = new OpenAI({
apiKey: process.env.TOKENHUB_API_KEY,
baseURL: "https://us-api.tokenhub.com/v1",
})
const result = await client.chat.completions.create({})
console.log(result.choices[0]?.message?.content)Claude Opus 4.8 (Adaptive Reasoning, Max Effort)
| 指数スコア | ||
|---|---|---|
| Artificial Analysis Intelligence Index | 総合能力の集約評価 | 55.7 |
| Artificial Analysis Coding Index | コーディング能力の集約評価 | 56.7 |
| 知識と推論 | ||
| GPQA | 高度な科学問題の解決 | 92% |
| HLE | 幅広い分野の専門家レベル試験 | 45.7% |
| コーディングとソフトウェア開発 | ||
| SciCode | 科学計算のコーディング課題 | 53.5% |
| Terminal-Bench Hard | 高難度のターミナル操作 | 58.3% |
| 指示追従とエージェントタスク | ||
| IFBench | プロンプト制約への準拠 | 62.2% |
| AA-LCR | 長いコンテキストでの推論 | 67.7% |
| τ²-Bench | エージェントのワークフロータスク | 94.4% |
指標の出典 Artificial Analysis
Claude Opus 4.8 の位置づけ、適した用途、主な強み、注意点、TokenHub で安全に利用するための確認事項を紹介します。
Claude Opus 4.8 は、複雑な推論、長時間動くエージェント、コーディング、専門業務向けの Anthropic 最新 Opus モデルです。提供元の公開ドキュメントに掲載される現行モデルですが、利用可否はプラットフォームによって異なります。
長時間動作する自律エージェント、エージェント型コーディングとリポジトリ作業、および専門文書と意思決定の分析に適しています。本番導入前に代表的な入力でテストし、測定可能な合格基準を定めてください。
実用上の主な強みは高いコーディング能力、ツールと関数呼び出しの有効活用、および不確実性や根拠の弱い結論への慎重な対応です。この組み合わせは、エージェント型コーディングとリポジトリ作業で特に価値があります。
小型モデルで十分な単純作業の場合、最優先事項が極めて低い遅延である場合、または重要な判断に人手レビューを入れられない場合は、別のモデルも検討してください。事実、法務、財務、医療、運用に関わる重要な出力は、適切な専門知識を持つ人が確認してください。
TokenHub では、Claude Opus 4.8 に表示される正確なモデル識別子を選び、アカウント資料に記載されたエンドポイントと TokenHub 認証情報を使用してください。アカウントで利用できる Claude の機能、コンテキスト上限、ツール対応、現在のモデル状態を TokenHub のモデルページで確認してください。
1つの API キーで Claude Opus 4.8 とその他の AI モデルにアクセスできます。
メディアと議論
このモデルに関連する公開動画と投稿を選んでいます。
X (Twitter)
Anthropic ·
New Frontier Red Team blog: Phase 2 of Project Fetch, where we test how well Claude can program a robodog. Opus 4.7, on its own, was ~20x faster than last year's best human team aided by Opus 4.1. (The robodog, alas, still failed to fetch a beach ball.) https://www.anthropic.com/research/project-fetch-phase-two
原文を見るAnthropic ·
New Anthropic Science Blog: Making Claude a chemist. To manipulate a molecule, chemists first need to understand its structure. Their main tool is NMR spectroscopy. We found Opus 4.7 matches—and on some tasks beats—dedicated NMR software. Read more: https://www.anthropic.com/research/making-claude-a-chemist
原文を見るAnthropic ·
Each time we release a model, we run the same test: give it code that trains a small AI model, ask the new model to speed it up. It takes a skilled human 4-8 hours to reach 4x faster. In May 2024, Claude Opus 4 averaged a ~3x speedup. This April, Mythos Preview achieved ~52x.
原文を見るReddit
Status Report: Claude Performance Megathread – Week of May 11 – 18, 2025
sixbillionthsheep ·
全文を読む
Prior week's Status Report is here: https://www.reddit.com/r/ClaudeAI/comments/1kjxis1/status_report_claude_performance_megathread_week/ Disclaimer: This was entirely built by AI. Please report any hallucinations. 🧠 Executive Summary (11 – 18 May 2025) Over the past week Claude users on Reddit reported a sharp rise in truncated chats, stalled or empty Artifacts, and login/uptime errors. Sentiment was strongly negative (~70 % critical). Anthropic’s status page confirms a live long-context incident and multiple shorter outages that coincide with the complaints, suggesting backend changes (Artifact rollout plus context bug) as the root cause. Some community workarounds exist but are limited. 📊 Key Performance Observations (from Reddit comments) Rank Category What users actually saw Illustrative quote 1 Context-window collapse Chats cut off after 1-5 prompts; even 159-word inputs rejected. “I haven’t been able to get more than 4 or 5 prompts in a row in a single chat for about two days.” 2 Artifact instability “Drafting…” never completes; copy button missing, files export empty. “It starts drafting the Artifact and just sits there forever.” 3 Availability & login failures “Internal Server Error”, Cloudflare loops, Android shows “offline”. “Every login just throws me back to the start page.” 4 Speed / latency drop Long thinking phases; extended-thinking time shrinks each new chat. “Claude Code hangs on a single line, then times out.” 5 Unexpected UI changes Copy button removed; emojis suddenly appear in math chats. “Claude’s sending hearts and emojis now??” 6 Project data loss Entire project folders disappeared for some paid users. “My whole project just vanished!” 😊 😐 🙁 Overall Megathread Sentiment Negative (~70 %) — words like defrauded, gut-punch, unprofessional. Neutral (~20 %) — troubleshooting and confirmations. Positive (~10 %) — a few users on Enterprise or via Cursor reported normal or better-than-usual performance after maintenance on 16 May. Sentiment turned markedly worse between 14–16 May, matching the status-page incident window. 🔄 Recurring Themes & Topics Collapsed context / premature chat limits – most frequent issue. Artifact pipeline failures – second leading complaint. Paid-tier capacity limits – Max/Enterprise users still rate-limited. Regional reliability problems – reports cluster in Australia & Scandinavia. Claude Code regressions – loops, autocompactor triggers too early. 🛠️ Possible Workarounds (community-tested or documented) Issue Workaround Source / reliability Artifact never renders Ask Claude to output Markdown directly in chat instead of Artifact. Reddit user tip — moderate reliability Android “offline” error Switch to mobile data, change DNS, or disable VPN (Cloudflare challenge). Reddit users in AU — moderate reliability Ultra-short chat limits Start a fresh chat with a short recap or switch to Claude 3.5 Sonnet. Multiple Reddit reports — low-moderate Empty Artifact **“Try fixing with Claude”**Use new button in UI. Anthropic support doc — high reliability 👍 Notable Positive Feedback “I’m blown away at how good Claude Code is compared to Roo Code!” “After they finished that maintenance…I think they found the bug; I’m hitting 200+ pages now.” 👎 Notable Negative Feedback “By the time you finish explaining things you hit limits… breadcrumbs as a paying user, it’s a slap in the face.” “Claude writes the entire code, then a network error deletes everything.” 🌐 External Context & Potential Explanations External finding Link to user reports Long-context incident (opened 15 May) — “elevated errors on long context requests”. Directly explains abrupt chat-length failures and context truncation. Outages 11–13 May on Claude 3.x Align with spikes in login errors and 500 responses. Artifact GA rollout 14 May Coincides with new UI (copy → export) and Artifact stalls. Support doc: “Try fixing with Claude” Confirms Anthropic is aware of Artifact generation failures. No official acknowledgement yet of removed copy button, personality drift (emojis), or regional uptime gaps. 🆕 Potential Emerging Issues Personality drift – emojis/hearts in technical chats may hint at prompt-template experimentation. Project data loss – isolated but severe; not yet publicly addressed. Bottom line User-reported regressions in context length, Artifacts, and uptime line up with Anthropic’s own incident logs. Until the long-context issue is resolved and the Artifact backend stabilises, the listed workarounds may help reduce workflow disruption. Key sources informing this report (in order of direct relevance) Anthropic Status Page — incident “elevated errors on long-context requests”, opened 15 May 2025, still active when checked. Anthropic Status Page — short outages for Claude 3.x on 11–13 May 2025. Anthropic Support Article — “Try fixing with Claude” for broken Artifacts, updated 18 May 2025. Anthropic blog / release note — Artifact feature moved to general availability 14 May 2025. (Anonymous Reddit comments from the May 11–18 Megathread provide all user-reported data.)
[ Removed by moderator ]
Interesting-Appeal35 ·
原文を見るLetting the AIs Judge Themselves: A One Creative Prompt: The Coffee-Ground Test
Double_Picture_4168 ·
全文を読む
I work on the best way to bemchmark todays LLM's and i thought about diffrent kind of compettion. Why I Ran This Mini-Benchmark I wanted to see whether today’s top LLMs share a sense of “good taste” when you let them score each other, no human panel, just pure model democracy. The Setup One prompt - Let the decide and score each other (anonimously), the highest score overall wins. Models tested (all May 2025 endpoints) OpenAI o3 Gemini 2.0 Flash DeepSeek Reasoner Grok 3 (latest) Claude 3.7 Sonnet Single prompt given to every model: In exactly 10 words, propose a groundbreaking global use for spent coffee grounds. Include one emoji, no hyphens, end with a period. Grok 3 (Latest) Turn spent coffee grounds into sustainable biofuel globally. ☕. Claude 3.7 Sonnet (Feb 2025) Biofuel revolution: spent coffee grounds power global transportation networks. 🚀. openai o3 Transform spent grounds into supercapacitors energizing equitable resilient infrastructure 🌍. deepseek-reasoner Convert coffee grounds into biofuel and carbon capture material worldwide. ☕️. Gemini 2.0 Flash Coffee grounds: biodegradable batteries for a circular global energy economy. 🔋 scores: Grok 3 | Claude 3.7 Sonnet | openai o3 | deepseek-reasoner | Gemini 2.0 Flash Grok 3 7 8 9 7 10 Claude 3.7 Sonnet 8 7 8 9 9 openai o3 3 9 9 2 2 deepseek-reasoner 3 4 7 8 9 Gemini 2.0 Flash 3 3 10 9 4 So overall by score, we got: 1. 43 - openai o3 2. 35 - deepseek-reasoner 3. 34 - Gemini 2.0 Flash 4. 31 - Claude 3.7 Sonnet 5. 26 - Grok. My Take: OpenAI o3’s line— Transform spent grounds into supercapacitors energizing equitable resilient infrastructure 🌍. Looked bananas at first. Ten minutes of Googling later: turns out coffee-ground-derived carbon really is being studied for supercapacitors. The models actually picked the most science-plausible answer! Disclaimer This was a tiny, just-for-fun experiment. Do not take the numbers as a rigorous benchmark, different prompts or scoring rules could shuffle the leaderboard. I’ll post a full write-up (with runnable prompts) on my blog (https://tryaii.com/blog/llms-self-scoring-coffee-benchmark) soon. Meanwhile, what do you think did the model-jury get it right?
YouTube
I Turned Claude Opus 4.8 Into My Entire AI Operating System
Nate Herk | AI Automation
原文を見るClaude Opus 4.8 + n8n MCP: Autonomous Coding Pipeline Tutorial
Derek Cheung | AI Agents Automation
原文を見るClaude Opus 4.8 Full Breakdown & Testing (AI News You Can Use)
The AI Advantage
原文を見る