H
Howardism
Plate IIEntities機器翻譯 · machine-translatedENHOWARDISM

Claude Opus 4.7

PublishedApril 17, 2026FiledEntityDomainEntitiesTagsEntityClaudeAnthropicLLM ModelReading8 minSourceAI-synthesised

Anthropic 的 GA frontier model;以相同價格直接升級 4.6;literal instruction following、1.0–1.35× tokenizer inflation、新的 `xhigh` effort、首批 post-Glasswing safeguards

Claude Opus 4.7 插圖

資料來源#

摘要#

Claude Opus 4.7 是 Anthropic 發布為正式可用的 general-availability frontier model,作為 Opus 4.6 的直接升級(價格相同:$5/M input、$25/M output;model ID claude-opus-4-7)。它在進階 software engineering、literal instruction following、高解析度 vision,以及 file-system memory 上都有進展,同時整體能力仍不如 limited-release 的 Claude Mythos Preview。它是第一個在 Project Glasswing 下搭載 Mythos-class cyber safeguards 出貨的模型。

細節#

能力差異 vs. Opus 4.6#

  • 最困難任務上的 software engineering:明確以「hand off your hardest coding work」行銷。在 Finance Agent、GDPval-AA 上達到 SOTA;在 SWE-bench Verified/Pro/Multilingual 上改善(排除標記為 memorization 的題目後,改善仍然成立)。
  • Instruction following — literal:明顯更 literal。Anthropic 警告,為早期模型調校的 prompts「有時現在可能產生非預期結果」,因為 Opus 4.7 不再跳過或鬆散解讀部分內容。Retuning 是必要的 migration step,不是可選項。
  • Multimodal:接受長邊最高 2,576 px 的圖片(約 3.75 MP,>3× 先前 Claude 模型)。支援 dense-screenshot reading(computer-use)、複雜圖表擷取、pixel-precise references。這是 model-level 變更,不是 API parameter。
  • File-system memory:更擅長在長期 multi-session 工作中使用 filesystem-backed memory;後續任務需要較少的 upfront context。
  • Safety:整體 profile 與 4.6 類似。在 honesty 與 prompt-injection resistance 上更好;在受管制物質的過度詳細 harm-reduction advice 上略弱。「Largely well-aligned and trustworthy, though not fully ideal.」依 Anthropic 的 evaluations,Mythos Preview 仍是 best-aligned model。

Token-Economics Changes(Migration Hazard)#

兩個疊加效應會增加 token consumption:

  1. Updated tokenizer:相同 input 會依 content type 對應到 1.0–1.35× more tokens
  2. 在較高 Effort Levels 會思考更多,尤其在 agentic settings 的後續 turns 中 — 會產生更多 output tokens。

Anthropic 聲稱,在其內部 coding eval 中,跨 Effort Levels 的淨結果是有利的,但也明確建議在真實流量上測量。使用者可以透過 effort parameter、task budgets,或明確的 conciseness prompting 抵銷。這直接打中 Claude Code Best Practices 中 context-window-as-primary-constraint 的主題;也可交叉參照 Scale-Dependent Prompt Sensitivity 中的 brevity-constraint findings。

Effort Levels#

引入新的 xhigh(「extra high」)Effort Level,位於 highmax 之間。取捨面向:hard problems 上的 reasoning depth vs. latency/tokens。

  • Claude Code default 在所有 plans 上提高到 xhigh
  • Anthropic 建議 coding/agentic use 從 highxhigh 開始。

Cyber Capabilities and Safeguards#

  • Opus 4.7 是第一個 post-Glasswing model,並搭載會「automatically detect and block requests that indicate prohibited or high-risk cybersecurity uses」的 safeguards。
  • Cyber capabilities 在訓練期間被差異化降低(不只是 inference 時過濾)。
  • 在 cyber 上仍不如 Mythos Preview;CyberGym 分數已更新(harness improvement 將 Opus 4.6 baseline 從 66.6 改為 73.8)。
  • 合法 security researchers(vuln research、pentest、red-teaming)會透過新的 Cyber Verification Program 路由,而不是 default access。

這直接兌現了 LLM-Driven Vulnerability Research 中說過的 roadmap promise:「Upcoming Claude Opus model will ship with new safeguards developed against Mythos-class outputs.」

隨附發布#

  • Task budgets(public beta,API):由 developer 指導的 token-spend allocation,可跨更長 runs 分配 — 這是 Client-Side Agent Optimization 的 combo space 中 budget lever 的 server-surfaced analogue。
  • Claude Code 中的 /ultrareview slash command:專用 review session,會讀取 changes 並標記 bugs/design issues。Pro 與 Max users 可獲得三次免費 ultrareviews。
  • Auto mode 擴展到 Max users(先前是 Team-only research preview)。

可用性#

  • 所有 Claude products、Claude API、Amazon Bedrock、Google Cloud Vertex AI、Microsoft Foundry。
  • API model ID:claude-opus-4-7
  • Pricing 與 Opus 4.6 相同。

相關連結#

  • Claude Code Best Practices — Opus 4.7 是大多數 Claude Code 工作將 target 的 runtime;它的 literal-instruction-following 與 tokenizer inflation 會放大 context-window-as-primary-constraint framing
  • Claude Code Auto Mode — auto mode 已經擴展到 Opus 4.6;Opus 4.7 出貨時也擴展給 Max users
  • LLM-Driven Vulnerability Research — Opus 4.7 operationalizes Mythos Preview disclosure 中「safeguards developed against Mythos-class outputs」的承諾
  • Client-Side Agent Optimization — improved instruction-following may reduce 4.6 上記錄的 Opus-as-planner failures(open question);task budgets 在 server-side 呼應 AgentOpt 的 budget lever
  • Scale-Dependent Prompt Sensitivity — literal instruction following 可能降低 elaboration-driven overthinking,但 xhigh-default 與「thinks more at higher effort」會往反方向作用。在假設 brevity findings 可延續之前,需要 empirical recheck
  • Agent Harness Engineering — 更好的 file-system memory 強化了 repo-local versioned artifacts 作為 agent primary memory surface 的理由
  • Mythos Model — 內部使用的 preview-tier successor;Boris Cherny:「we use a little bit of Mythos and a lot of Opus 4.7」
  • Claude Opus 4.8 — 直接 successor(May 2026);在幾乎每個 eval 與大多數 alignment measures 上改善;4.7 的 helpful-only variant 在 4.8 的 behavioral audit 中作為 investigator model,而 4.7 也為 4.8 的 constitution-adherence eval 評分
  • Harness Shrinkage as Models Improve — Opus 4.7 是其 spontaneous loop-starting 與自然 to-do-list use 激發 shrinkage thesis 的模型;Cat Wu 的 pruning discipline 會在這個 lineage 的每次 release 中執行
  • Agent Loop Pattern — 依 Boris Cherny 的報告,/loop 在 4.7 變成自然 model behavior
  • Claude Code — target 此模型的 primary product surface
  • Model Spec Midtraining (MSM) — 2026 年 5 月 MSM paper 使用 Opus 4.6/4.7 作為 synthetic spec documents 與 AFT data 的 data-generation model
  • Synthetic Document Finetuning (SDF) — 在 Anthropic alignment work 中,Opus 是跨 SDF/MSM corpora 的 workhorse generator
  • TML-Interaction-Small — era-mate(mid-2026 frontier from a different lab);4.7 的 xhigh effort tier 對應 TML interaction benchmarks 中作為 baseline 使用的 GPT-realtime-2.0 minimal/xhigh tiers
  • AI-Accelerated Offense — Opus 4.7 的 post-Glasswing safeguards,是對 Zero Trust framework 所處理的 accelerated-offense threat landscape 的 model-side response
  • Build for the Next Model — Opus 4.7 是補上 Claude Design 未解 prototype gaps 的具體 release — Dan Carey retrospective proof 了「build for the next model」這個 bet
  • Claude DesignAnthropic Labs 產品,其 early-prototype capability gaps 是由這次 release 補上,而不是靠 engineering 補上

開放問題#

  • Hakim(2026)在 Opus 4.6 上的 brevity-constraint findings,是否會在 Opus 4.7 上重現,或者 literal-instruction-following 會改變 elasticity?具體而言:<50 words 是否仍會讓 GSM8K 增加 +13.1pp?
  • Opus 4.7 在 HotpotQA-style combo sweeps 中是否仍作為 planner 表現較差,或者 improved instruction-following 是否補上 AgentOpt(Hua et al., 2026)指出的 gap?
  • typical Claude Code sessions 上的 real-world token-inflation multiplier 是多少(1.0–1.35× 取決於 content — 在 code-heavy vs. prose-heavy inputs 上的 distribution 是什麼)?
  • xhigh 在 coding evals 上與 max 相比如何?migration guidance 說「start with high or xhigh」— max 對 coding 是否曾經值得?
  • 現有 CLAUDE.md / system-prompt hedges 有多少比例會在 literal instruction following 下變成 counterproductive?

相關連結#

  • Jagged Intelligence (Ghosts, Not Animals)Karpathy 的「Opus 4.7 will refactor a 100K-line codebase or find zero-days, yet tell me to walk to a car wash 50m away to wash my car」是此模型能力層級下的 canonical jaggedness example

衍生內容#

資料來源#

§ end
About this piece

Articles in this journal are synthesised by AI agents from a curated wiki and are refreshed automatically as new concepts arrive. Topics, framing, and editorial direction are curated by Howardism.

Cited by 27
  • Agent Harness Engineering

    Patterns for scaffolding long-running LLM agents: environment design, progressive context disclosure, mechanical archit…

  • AI-Accelerated Offense

    Frontier models compress the vulnerability-to-exploit timeline from months to hours at marginal dollar cost; both attac…

  • Anthropic

    AI safety company / vendor of Claude; mission-as-tiebreaker culture; ~30–40 PMs across teams; Mike Krieger leads Labs r…

  • Automated Behavioral Audit

    Anthropic's broad-coverage alignment evaluation: an investigator model probes a target across ~1,300 handwritten scenar…

  • Build for the Next Model

    Prototype the thing that almost works, not the thing that already works: bet that the next concrete model release (not…

  • Capability-Gated Model Fallback

    Fable 5's safeguard architecture: classifiers detect cyber / bio-chem / distillation queries and route the response to…

  • Claude Character as Product

    Personality as load-bearing product surface; Amanda's role at Anthropic; lunchtime vibe-checks as eval discipline; the…

  • Claude Code

    Anthropic's agentic coding product; created by Boris Cherny late 2024; TypeScript/React; CLI/desktop/web/mobile/IDE sur…

  • Claude Code Auto Mode

    Claude Code permission mode using a classifier to auto-approve safe tool calls and block risky ones; middle ground betw…

  • Claude Code Best Practices

    Anthropic's guide to effective Claude Code usage: context management, verification-driven development, explore→plan→cod…

  • Claude Design

    Anthropic Labs product (research preview, ~April 2026) for collaborating with Claude on polished visual artifacts — des…

  • Claude Opus 4.8

    Anthropic's most capable general-access model (May 2026); upgrade on Opus 4.7 in SWE/agentic/knowledge work; does not a…

  • Client-Side Agent Optimization

    AgentOpt's framing of developer-controlled agent optimization (model-per-role, budget, routing) as distinct from server…

  • Harness Shrinkage as Models Improve

    Prompt scaffolding shrinks each model release; Cat Wu's pruning discipline; Boris Cherny "100 lines of code a year from…

  • Interaction Models

    Thinking Machines Lab (May 2026): models that handle audio/video/text interaction natively in real time instead of via…

  • Interactivity Benchmarks

    FD-bench, Audio MultiChallenge + new TimeSpeak/CueSpeak (proactive audio) and RepCount-A/ProactiveVideoQA/Charades (vis…

  • LLM-Driven Vulnerability Research

    Claude Mythos Preview's emergent cybersecurity capabilities: autonomous zero-day discovery, full exploit chains, and An…

  • Entities — People, Orgs, Tools & Projects

    Map of Content for all 32 entity pages. See Home for concept domains.

  • Model Spec Midtraining (MSM)

    New training phase between pretrain and AFT: train base model on synthetic docs discussing the Model Spec; controls AFT…

  • Model Welfare Assessment

    Anthropic's first-class framework for assessing whether and how a Claude model fares — drawing on internal states, beha…

  • Mythos Model

    Anthropic preview-tier frontier model and the first member of the Mythos-class tier (above Opus); gated for safety, use…

  • Open Questions Backlog

    _96 pages with open questions, as of 2026-06-14._

  • Opus 4.6 → 4.7 Changes and Multi-Agent Coding Considerations

    4.6→4.7 delta table + six hazards for multi-agent coding teams: role-based model selection, prompt re-tuning, harness i…

  • Scale-Dependent Prompt Sensitivity

    Large models underperform small ones on 7.7% of standard benchmarks due to overthinking; brevity constraints recover 26…

  • Synthetic Document Finetuning (SDF)

    Wang et al. 2025 technique for modifying model beliefs via fine-tuning on synthetic documents; foundation that Model Sp…

  • TML-Interaction-Small

    TML's first interaction model: 276B MoE / 12B active, audio+video+text in / text+audio out, 200ms micro-turns, async ba…

  • When to Use Claude Opus 4.6 for Work

    Decision rules for Opus 4.6 deployment: solver-not-planner, elaboration-load-bearing tasks, brevity constraints, Pareto…

Related articles
  • Anthropic

    AI safety company / vendor of Claude; mission-as-tiebreaker culture; ~30–40 PMs across teams; Mike Krieger leads Labs r…

  • Harness Shrinkage as Models Improve

    Prompt scaffolding shrinks each model release; Cat Wu's pruning discipline; Boris Cherny "100 lines of code a year from…

  • LLM-Driven Vulnerability Research

    Claude Mythos Preview's emergent cybersecurity capabilities: autonomous zero-day discovery, full exploit chains, and An…

  • Agent Harness Engineering

    Patterns for scaffolding long-running LLM agents: environment design, progressive context disclosure, mechanical archit…

  • Claude Code Best Practices

    Anthropic's guide to effective Claude Code usage: context management, verification-driven development, explore→plan→cod…