Claude Opus 4.7 Just Got Way Smarter at Reading Your Mind

Claude Opus 4.7 just launched. Here’s what’s actually different for builders.

Anthropic released Claude Opus 4.7 as a direct upgrade to Opus 4.6, and this one is not a cosmetic rename. The headline is better performance on hard software engineering and long-running agent workflows, plus a major vision upgrade for high-resolution image tasks.

If you build with frontier models, the practical question is simple: does this reduce retries, tool failures, and supervision time enough to justify migration? Based on Anthropic’s published numbers and early partner evals, for many coding-heavy teams the answer is probably yes.

But there are also tokenization and cost-behavior changes you need to plan for, especially if you run long autonomous flows.

What changed at a glance

• Model: Claude Opus 4.7 is now generally available.
• API model ID: claude-opus-4-7.
• Pricing: unchanged from Opus 4.6 at $5/M input and $25/M output tokens.
• Core claim: stronger advanced software engineering, especially difficult and long-horizon tasks.
• Vision: image input now supports up to 2,576 px on long edge (~3.75 MP), over 3x prior Claude image detail.
• New effort level: xhigh (between high and max) for finer latency vs reasoning control.

The benchmark movement that matters

Anthropic’s post mixes internal and partner-reported evaluations, so treat this as directional but useful. The deltas are still big enough to pay attention to.

• CursorBench: Opus 4.7 at 70% vs Opus 4.6 at 58%.
• Sourcegraph (93-task coding benchmark): +13% resolution over Opus 4.6, including 4 tasks neither Opus 4.6 nor Sonnet 4.6 solved.
• Notion Agent multi-step workflows: +14% vs Opus 4.6, with fewer tokens and roughly one-third of tool errors.
• Finance agent benchmark: tied top overall score at 0.715 across six modules; General Finance module 0.813 vs Opus 4.6 0.767.
• BigLaw Bench (Harvey): 90.9% at high effort.
• XBOW visual-acuity benchmark: 98.5% vs Opus 4.6 54.5%.
• Databricks OfficeQA Pro: 21% fewer errors than Opus 4.6.
• Factory Droids: reported 10–15% task-success lift over Opus 4.6.
• Rakuten-SWE-Bench: reported 3x more production tasks resolved than Opus 4.6.

These are not tiny improvements. If your product depends on tool-calling reliability, long context, or agent follow-through, this is a serious upgrade signal.

New capabilities (plain English version)

The model appears better at finishing difficult work without constant babysitting. That sounds abstract, but in production it means fewer “stopped halfway” runs and better recovery when tools fail.

• Hard coding tasks: better at catching its own logic faults and validating output before claiming completion.
• Long-running autonomy: more coherent multi-step execution over longer sessions.
• Instruction following: substantially more literal compliance with prompt instructions.
• High-res vision: better reading of dense screenshots, diagrams, chemical structures, and technical docs.
• File-based memory use: stronger cross-session continuity in workflows that store notes/state in files.

Anthropic also claims improved “taste” on professional artifacts (UI, decks, docs). That’s harder to quantify, but relevant if your team uses model output directly in client-facing deliverables.

What could break when you upgrade

This is the part teams skip and regret later.

• Prompt behavior changes: because Opus 4.7 follows instructions more literally, old prompts that relied on “fuzzy interpretation” may behave differently.
• Tokenization shift: same input can map to ~1.0–1.35x tokens depending on content type.
• More reasoning at higher effort: later turns in agentic runs can produce more output tokens.
• Vision token cost: higher-resolution image processing can increase token spend unless you downsample when fine detail isn’t needed.

So even with unchanged posted pricing, workload-level cost can move if your traffic leans long, visual, or high-effort.

Safety and cyber changes you should know

Opus 4.7 ships with real-time cyber safeguards intended to detect and block prohibited or high-risk cybersecurity requests. Anthropic positions this as part of a staged rollout after Project Glasswing, before broad release of Mythos-class capability.

For legitimate security use cases (red teaming, vuln research, pentesting), Anthropic is funneling access through a Cyber Verification Program. If your product includes security workflows, plan for policy gating and verification requirements in your user journey.

Who should care immediately

• Agent builders: especially teams running multi-tool orchestration where loop resistance and error recovery drive UX.
• Coding platforms: IDE copilots, PR review, CI/CD assistants, and long-horizon debugging agents.
• Enterprise doc/reasoning apps: legal, finance, and data-heavy workflows where high-res vision and structured analysis matter.
• Teams fighting reliability issues: if Opus 4.6 needed too much supervision on hard tasks, 4.7 is worth immediate A/B testing.

Who probably shouldn’t rush

• Simple chat or low-complexity support bots: you may not monetize the extra reasoning depth.
• Strict budget-sensitive apps without routing: token behavior changes can surprise you.
• Teams with brittle parsers/prompts: literal instruction following can surface hidden prompt design flaws.

What to do this week

Run a controlled migration instead of a full flip.

• Step 1: add claude-opus-4-7 behind a feature flag with instant rollback to Opus 4.6.
• Step 2: test your top 25 real production tasks at high and xhigh effort.
• Step 3: track pass rate, tool-call error rate, tokens per successful task, and human correction time.
• Step 4: re-tune prompts for stricter instruction following.
• Step 5: set task budgets and cap output verbosity for long agent loops.
• Step 6: downsample images by default unless high-detail vision is essential.

If your metrics show higher success per run with lower intervention, expand rollout. If costs rise without quality gains, route selectively by task type.

Bottom line

Claude Opus 4.7 looks like a meaningful frontier-model upgrade, not a minor patch. The strongest evidence is in coding and agentic reliability deltas: higher task completion, fewer tool errors, better long-context execution, and much stronger high-res vision performance.

The catch is operational: literal prompt adherence and tokenization differences can change behavior and spend if you migrate blindly.

For builders, the right move is immediate, structured evals. Don’t decide on launch vibes. Decide on your own pass-rate, supervision-time, and compute-economics data.

Now you know more than 99% of people. — Sara Plaintext