Anthropic's three-tier lineup makes the choice harder than it should be. Opus 4.7 looks irresistible at the top, but most teams ship to production on Sonnet and never look back. The right answer is "it depends on the turn" — and routing well is where the real cost savings live.
This guide is the price-performance grid I wish someone had handed me when our team migrated off GPT-4o.
What changed in 2026
- Opus 4.7 (1M context) ships generally available — long enough to hold a full repo, multiple specs, and a transcript in one prompt.
- Sonnet 4.6 closed most of the gap on code tasks — on SWE-bench Verified it's within 6 points of Opus at one-fifth the cost.
- Haiku 4.5 is the new default for chat-tier features — fast enough for streaming UIs, smart enough for routing decisions and classification.
Sonnet 4.6 — the default
Sonnet handles the vast majority of production work: code generation, structured extraction, summarization, RAG synthesis, agentic tool calls under five steps. At roughly $3 in / $15 out per million tokens, it's the price point where you stop counting tokens and start counting features. Latency averages 800-1500ms first-token, which is fine for most non-interactive flows. The honest weakness is multi-step reasoning where intermediate errors compound — for those, Opus pays for itself by getting the answer right on the first try instead of looping three times.
Opus 4.7 — for hard reasoning
Opus earns its premium on tasks where Sonnet is almost right. Think: reading a 200-page contract and flagging unusual clauses, planning a refactor across a 50k-line codebase, or synthesizing five conflicting research papers into a coherent thesis. The 1M context window is the real unlock — you stop chunking and start putting the whole problem in front of the model. At $15 in / $75 out, it's expensive per token but cheap per correct answer if you've been retrying smaller models.
Haiku 4.5 — the silent winner
The most underrated tier in the lineup. Haiku 4.5 is fast (200-400ms first-token), cheap ($0.80 in / $4 out), and good enough for classification, routing, light summarization, and tier-one chat. The classic mistake is starting on Sonnet for everything; the better pattern is to put Haiku in front, let it decide whether the turn needs escalation, and only then call Sonnet or Opus. Production teams report 60-80% cost reduction from this pattern alone.
Tier comparison: Claude 4.x lineup in May 2026
| Tier |
Input / Output (per 1M tokens) |
First-token latency |
Best for |
| Haiku 4.5 |
$0.80 / $4 |
200-400ms |
Classification, routing, chat |
| Sonnet 4.6 |
$3 / $15 |
800-1500ms |
Code, RAG, structured extraction |
| Opus 4.7 (1M) |
$15 / $75 |
2-5s |
Long-context reasoning, planning |
Common mistakes to avoid
Defaulting everything to Opus "to be safe". It's not safer — it's slower and costs 5x. Most production traffic should never touch Opus.
Skipping Haiku because it "feels weak". It isn't 2024 Haiku. The 4.5 generation handles most chat-tier work and pays for itself in latency improvements alone.
Pricing on input tokens only. Output tokens dominate cost on coding and long-form tasks. Always model both.
FAQ
When should I use Opus 4.7 over Sonnet 4.6?
When you need a single correct answer on a hard problem and retries are expensive — long-context analysis, multi-step planning, ambiguous research. If you're streaming text or doing tool calls under five hops, Sonnet is almost always the answer.
Does prompt caching change the math?
Yes, dramatically. Cached input tokens are ~90% cheaper. If you reuse a long system prompt or document context, prompt caching can flip the cost equation in favor of Opus.
Is the 1M context worth it in practice?
For specific use cases — repo-level code review, contract analysis, multi-document synthesis — yes. For most chat or RAG use cases you're already chunking, so it doesn't matter.
Should I A/B test tiers in production?
Always. Pick five representative turns from your prod traffic and run all three tiers. The right tier per route is often surprising.
Where to go next
For related guides see Claude vs ChatGPT vs Gemini in 2026, Claude Opus 4.7 — everything new, and How to use Claude for coding in 2026.