Claude 4.7 vs GPT-5 for business automation — which one to use in n8n

March 2026 gives us two thrones in AI automation: Claude 4.7 Opus from Anthropic and GPT-5 from OpenAI. Both claim the crown. Both cost comparably. Both plug into n8n in 5 minutes.

This isn't a synthetic benchmark roundup (MMLU, GPQA — useless in practice). It's 4 real tasks from workflows we ship to clients, with cost and a per-case winner.

Task 1: B2B lead scoring (0–100)

Workflow: lead from LinkedIn Sales Navigator → Apollo enrichment → AI returns score and justification.

• Claude 4.7: more conservative scoring (avg 58/100), with better "why that number" reasoning. Very good at spotting negative signals (freelancers, students, retirees).
• GPT-5: more optimistic (avg 71/100), better at inferring buying intent from behavior data. But it tends to default to "always 70+" when the lead isn't obviously bad.
• Winner: Claude, if you care about quality over volume. GPT, if you want more "yellow-flag" leads into sales.

Cost per lead (2,000 input / 300 output tokens): Claude 4.7 Opus ~$0.024, GPT-5 ~$0.020.

Task 2: Cold-email drafting from a LinkedIn profile

Workflow: profile → AI generates 3 variants → A/B tested, best wins.

• Claude: emails sound more human. Fewer superlatives, more specifics. Downside: sometimes too short, you have to force length.
• GPT-5: polished emails, but often detectable as AI — overuses "I noticed that…", "I'd love to…". Claude has the edge on non-English markets.
• Winner: Claude for cold B2B outreach.

Task 3: PDF invoice parsing and line-item extraction

Workflow: email with PDF invoice → OCR → AI structures into JSON → pushes to accounting.

• Claude: very stable JSON output. On 500 test invoices: 0 structural errors, 3 numeric errors (from bad OCR on small line items).
• GPT-5: on the same 500: 7 structural errors (broken JSON), 2 numeric. With response_format: json_object, structural errors drop to 0, but it sometimes skips line items.
• Winner: Claude. JSON stability is kingmaker in automation — 1 error per 500 is a monthly manual fix, 7 is constant firefighting.

Task 4: Tool use / multi-step agent

Workflow: an agent that autonomously calls APIs (CRM, calendar, Slack) to book a meeting with a lead.

• GPT-5: more mature tool calling. Better at parallel function calls, faster in agent loops.
• Claude: tool use works but is slower. Offsetting that: hallucinates parameters less often — won't invent field names that aren't in the schema. In production that matters more than latency.
• Winner: tie. Fast agents — GPT-5. Agents touching critical systems (billing, payments) — Claude.

Stability and rate limits

• Claude: tier 1 = 50 req/min, tier 4 (needs $400 spend) = 4,000 req/min. More stable API the last 6 months — fewer incidents.
• GPT-5: tier 1 = 500 req/min, tier 5 = 30,000 req/min. Scales faster but logs ~2 material incidents/month on status.openai.com.

n8n integration

Both have native nodes since n8n 1.80+. GPT additionally supports the Assistants API (persistent threads). Claude doesn't, but Claude has the better prompt caching (up to 90% off on repeated context) — if your system prompt is large, Claude is effectively ~3× cheaper.

Concrete recommendations

• Automations with natural-language writing (emails, product copy, summaries) especially non-English: Claude 4.7 Sonnet (cheaper than Opus, quality plenty).
• Scoring and classification: Claude Sonnet.
• PDF / invoice / contract parsing: Claude Opus (quality worth the premium).
• Fast tool-use agent (chatbot, customer service): GPT-5.
• Tight budget: GPT-5 nano or Claude Haiku — both 5× cheaper than flagships and good enough for 80% of tasks.

n8n tip: multi-model routing

In serious workflows, don't pick one model. Add a router: easy tasks (classification, scoring) → Haiku/Nano. Medium → Sonnet/GPT-5. Critical (outbound to clients, invoices) → Opus. You can cut API costs 60–70% without losing quality.