If you are evaluating Claude for a production build and trying to forecast monthly API spend, vendor calculators tend to wave their hands. This is the math we actually use when scoping Canadian client builds in 2026, with real benchmarks from production workloads.
Quick answer: Claude API cost ranges (2026)
For a typical Canadian SMB workload (1,000 to 10,000 conversations per month), Claude API costs land in these ranges:
- Claude Haiku 4.5: $20-$200/month, the cheapest, used for routing and high-volume light work.
- Claude Sonnet 4.6: $80-$500/month, the production workhorse, 80% of agent builds.
- Claude Opus 4.7: $200-$1,200/month for reasoning-heavy work and complex agents.
Most production builds use a mix: Haiku for routing, Sonnet for the bulk of work, Opus only for the hardest reasoning steps.
2026 Claude pricing (per million tokens)
| Model | Input | Output | Best for |
|---|---|---|---|
| Claude Haiku 4.5 | $1.00 | $5.00 | Classification, routing, simple Q&A |
| Claude Sonnet 4.6 | $3.00 | $15.00 | Production agents, RAG, balanced workloads |
| Claude Opus 4.7 | $15.00 | $75.00 | Complex reasoning, multi-step planning |
Prices fluctuate. Check Anthropic's pricing page for current rates. The relative ratios usually hold.
Token math (the part most articles skip)
Roughly:
- 1 token ≈ 4 characters of English text
- 1 token ≈ 0.75 words
- 1 page of typical text ≈ 500 tokens
- 1 hour of typed conversation (back and forth) ≈ 8,000-15,000 tokens
For a customer service chat:
- Average conversation: ~3,000 input tokens, ~800 output tokens.
- Cost on Sonnet 4.6: 3,000 × $3/$1M + 800 × $15/$1M = $0.009 + $0.012 = $0.021 per conversation.
For a voice call (transcribed):
- 5-minute call ≈ 750 words ≈ 1,000 tokens of input + 500 output.
- Cost on Sonnet 4.6: ~$0.011 per call. Plus speech-to-text + text-to-speech costs.
Monthly cost estimates by use case
Inbound support chat (3,000 conversations/month)
- 70% routed by Haiku, 30% handled by Sonnet
- Routing: 3,000 × $0.001 = $3
- Handling: 900 × $0.02 = $18
- Total: ~$21/month in LLM costs
Voice booking agent (500 calls/month, 4 min avg)
- STT + TTS: ~$50/month (not LLM)
- LLM (Sonnet): 500 × $0.015 = $7.50
- Total LLM: ~$8/month, total platform: ~$80-$120/month
Multi-step lead qualification (1,000 leads/month)
- ~10 LLM turns per lead, 2,000 input + 500 output per turn
- Per lead: 10 × ($0.006 + $0.0075) = $0.135
- Total: ~$135/month on Sonnet 4.6
Document analysis (200 PDFs/month, avg 20 pages each)
- ~10,000 input tokens per doc, 2,000 output
- Per doc on Sonnet: $0.03 + $0.03 = $0.06
- Total: ~$12/month
- Switching to Opus for the hardest 10%: +$25/month
Multi-agent research system (100 deep tasks/month)
- ~5 agent rounds, 20,000 tokens of context each
- Per task: 5 × ($0.06 + $0.075) = $0.675 on Sonnet, ~$2.25 on Opus
- Total: ~$70/month mixed Sonnet+Opus
How to cut costs 60-80%
1. Route by complexity
Send easy turns to Haiku, hard turns to Sonnet, only the hardest to Opus. A well-tuned router cuts cost 60% with no quality drop. Implement with a cheap classification call upfront.
2. Prompt caching
Anthropic supports prompt caching. If your system prompt is large (10,000+ tokens) and reused across calls, caching it cuts that portion to 10% of normal cost. Big win for RAG-heavy workloads.
3. Batch API for non-real-time work
For background jobs (document processing, summarization, classification), use the batch API. 50% discount on most models.
4. Tighter system prompts
Every token in the system prompt is paid for on every call. Cut yours from 4,000 tokens to 1,500 tokens (without losing capability) and your input bill drops by half.
5. Output token limits
Set max output tokens conservatively. Models will generate long answers if allowed. Most production agents need 200-500 output tokens, not 4,000.
6. Skip Opus by default
Opus is 5x the price of Sonnet for ~15-20% better quality on most tasks. Use it only where the reasoning gap actually matters (planning, hard math, ambiguous classification).
7. Compress retrieval context
RAG systems often dump 10-20 chunks into context "just in case." Re-rank and pass the top 3-5. Cuts input cost 70%, often improves quality too.
Real production numbers (from 3 Canadian SMB builds)
Plumbing contractor voice agent (Vapi + Claude Sonnet 4.6):
- ~800 calls/month, 4 min avg.
- LLM cost: $14/month
- Voice infrastructure: $320/month
- Total Claude bill: $14. Total platform: ~$420/month.
Clinic intake + scheduling (text only, Claude Sonnet 4.6):
- 2,500 conversations/month, avg 6 turns each.
- LLM cost: $42/month
- Hosting + monitoring: $80/month
- Total: ~$120/month.
Multi-agent research system (Claude Opus 4.7 + Sonnet 4.6):
- 150 deep research tasks/month.
- LLM cost: $145/month (60% Opus, 40% Sonnet)
- Hosting + observability: $200/month
- Total: ~$345/month.
Mistakes that blow up your bill
No max_tokens cap. A single runaway loop can rack up $50 in minutes. Always cap.
Repeating large context every call instead of using prompt caching.
Using Opus when Sonnet was fine. Test on Sonnet first. Only upgrade if you measure a meaningful quality gain.
No monitoring. Plug in LangSmith or your own logging. Surprise bills are almost always preventable.
Streaming in places it does not help. Streaming does not save tokens, only reduces perceived latency. Do not feel bad turning it off for background jobs.
How to estimate your build before committing
Rough formula:
Monthly cost ≈
(conversations/month) × (turns/conversation) × (avg_input_tokens × $X + avg_output_tokens × $Y)
For 1,000 conversations × 5 turns × (3,000 input × $3/$1M + 700 output × $15/$1M) on Sonnet 4.6: 1,000 × 5 × ($0.009 + $0.0105) = ~$98/month
Double it for safety margin. Subtract 50% once you tune. Land somewhere between.
Frequently asked questions
Does Anthropic offer discounts for committed spend? Yes, enterprise tiers offer volume discounts and dedicated capacity. Worth negotiating once you cross ~$2,000/month.
Is Claude cheaper than GPT-5? For input, GPT-5 is often slightly cheaper. For output, Claude Sonnet is often cheaper than GPT-5. Total cost depends on your input-to-output ratio.
Can I use Claude through AWS Bedrock or GCP Vertex? Yes. Pricing is similar to direct, with the advantage of unified cloud billing and Canadian data residency through AWS Canada Central.
Do I need to worry about cost predictability? Yes. Set hard spend caps in your Anthropic console. The biggest cost surprises come from prompt loops or rogue test scripts, not steady-state production.
Want help scoping the LLM bill for your build?
Free 30-minute consult. We will work through your projected volumes and give you a real monthly number. Book one.
