notes
November 24, 2025 - Initial Analysis Complete
What Just Happened
Claude Opus 4.5 was released today (November 24, 2025) - the last model in Anthropic’s 4.5 series, following Sonnet 4.5 (Sept 29) and Haiku 4.5 (Oct 15).
The Quick Answer to Your Question
Performance (Speed/Latency) vs Sonnet:
- Sonnet 4.5: ~63 tokens/second, 1.80s time-to-first-token (TTFT)
- Opus 4.5: ~45-50 tokens/sec (estimated), ~2.5s TTFT (estimated)
- Winner: Sonnet is ~40% faster in raw speed
- But: Opus uses 76% fewer output tokens, which can make total response time faster
Performance (Quality/Benchmarks) vs Sonnet:
- Opus 4.5: 80.9% SWE-bench (first Claude >80%)
- Sonnet 4.5: 77.2% baseline (82% with parallel compute)
- Gap: +3.7pp for Opus on single-attempt, but Sonnet can match with parallelization
Quality vs Sonnet:
- Opus wins on real-world coding (+10.6% Aider Polyglot)
- Opus wins on long tasks (+29% Vending-Bench)
- Opus wins on cross-language support (7/8 languages vs 1/8)
Cost vs Sonnet:
- Opus 4.5:
25 output = $30 per 1M tokens - Sonnet 4.5:
15 output = $18 per 1M tokens - Opus costs 67% more ($12 premium per 1M total tokens)
Key Insight: Historical Context
Opus prices dropped 67% from previous version:
- Old Opus 4:
75 = $90 per 1M tokens - New Opus 4.5:
25 = $30 per 1M tokens
This makes Opus 4.5 much more accessible despite being 67% more expensive than Sonnet.
Decision Framework: When to Use Each
Use Opus 4.5 When:
- Complex reasoning needed (agents, orchestration)
- Code quality is critical (production systems, security)
- Tasks run >30 minutes (Opus +29% better)
- Cost is secondary to quality
Use Sonnet 4.5 When:
- Routine development (API, CRUD, docs)
- High-volume requests (save 67% per request)
- Speed matters (latency-sensitive)
- Using parallelization (matches/exceeds Opus at lower cost)
Optimal Deployment: 80/20 Split
Recommended allocation:
- 80% Sonnet 4.5 (routine work)
- 20% Opus 4.5 (complex work)
Economics (1B input + 500M output monthly):
- Pure Sonnet: $9,000
- 80/20 blend: $10,200
- Pure Opus: $15,000
- Savings vs Opus: $4,800/month (-32%)
Quality Metrics Summary
| Metric | Opus | Sonnet | Winner |
|---|---|---|---|
| Speed | ~45-50 tok/s | ~63 tok/s | Sonnet (40% faster) |
| TTFT | ~2.5s | 1.80s | Sonnet (33% faster) |
| Token Efficiency | -76% tokens | baseline | Opus (fewer tokens) |
| SWE-bench Verified | 80.9% | 77.2% | Opus (+3.7pp) |
| Aider Polyglot | +10.6% | baseline | Opus |
| Vending-Bench (long tasks) | +29% | baseline | Opus |
| Language Coverage | 7/8 | 1/8 | Opus |
| Cost Efficiency | $30/M | $18/M | Sonnet (67% cheaper) |
| With Parallel Compute | 80.9% | 82% | Sonnet |
Effort Parameter (Game-Changer for Direct API)
Opus 4.5 has an effort parameter that changes everything:
| Effort | Quality | Output Tokens | Effective Cost |
|---|---|---|---|
| Medium | = Sonnet (77.2%) | 76% fewer | ~$11/M (39% cheaper than Sonnet!) |
| High | +4.3pp (80.9%) | 48% fewer | ~$18/M (same as Sonnet) |
Implication: With direct API access, Sonnet is obsolete!
Integration with Claude Code
⚠️ CRITICAL LIMITATION: Claude Code does NOT support effort/thinking parameters!
- Opus runs at HIGH effort (default) - cannot use medium
- Haiku runs with thinking OFF (default) - cannot enable
Claude Code-Specific Strategy:
| Task Type | Model | Cost | Why |
|---|---|---|---|
| Simple | Haiku 4.5 | $6/M | Fast, cheap |
| Complex | Opus 4.5 | $18/M | Same cost as Sonnet, +3.7pp quality |
| Sonnet | SKIP IT | $18/M | No advantage over Opus in Claude Code |
Key insight: Since Opus high = Sonnet cost in Claude Code, always prefer Opus for complex tasks!
Files Created
- README.md - Comprehensive overview with all metrics
- performance-vs-sonnet.md - Detailed benchmark analysis
- cost-analysis.md - Complete pricing scenarios and economics
- quality-characteristics.md - Code quality, reasoning, reliability
- deployment-strategy.md - Implementation patterns and recommendations
- speed-analysis.md - Latency, TTFT, and throughput comparison
- claude-code-strategy.md - Claude Code-specific limitations and strategy
Key Takeaways
For Direct API Users:
- Opus medium effort = 39% cheaper than Sonnet for equivalent quality
- Opus high effort = same cost as Sonnet, +4.3pp better
- Sonnet is obsolete when effort parameter is available
For Claude Code Users:
- Effort parameter NOT available - Opus runs at high by default
- Opus = Sonnet cost, so always use Opus over Sonnet
- Use Haiku for simple tasks (3x cheaper)
- Skip Sonnet entirely - no advantage in Claude Code
Sources
- Anthropic official announcement: https://www.anthropic.com/news/claude-opus-4-5
- Anthropic Haiku 4.5 announcement: https://www.anthropic.com/news/claude-haiku-4-5
- Claude Code documentation (model selection, custom commands)
- SWE-bench performance: First Claude >80%
- Pricing:
25 output (vs Sonnet 15)
Next Steps
For Direct API deployment:
- Use Opus medium for complex tasks (39% cheaper than Sonnet)
- Use Opus high for critical tasks (same cost as Sonnet, better quality)
- Use Haiku for simple tasks
- Skip Sonnet entirely
For Claude Code projects:
- Use Haiku for simple execution tasks (fast, cheap)
- Use Opus for complex tasks (NOT Sonnet - same cost, better quality)
- Configure Task subagents with
model: "haiku"ormodel: "opus" - Skip Sonnet - it has no advantage in Claude Code
Status: Research complete. Two distinct strategies documented:
- Direct API: Opus medium (39% cheaper) + Opus high (quality) + Haiku (simple)
- Claude Code: Opus high (default) + Haiku (simple) - skip Sonnet