Originally published at devtoolpicks.com
There are three Claude models in active production use in 2026. They are not interchangeable. Picking the wrong one either burns your AI budget on capability you do not need, or saves money while quietly producing outputs that are not good enough.
The pricing is straightforward: Haiku 4.5 at $1/$5 per million tokens, Sonnet 4.6 at $3/$15, Opus 4.7 at $5/$25. What is not straightforward is which tasks belong on which tier.
Here is the breakdown.
Quick Reference
| Model | Input/Output (per 1M tokens) | Best For | Avoid When |
|---|---|---|---|
| Haiku 4.5 | $1 / $5 | Classification, routing, extraction, triage | Complex reasoning needed |
| Sonnet 4.6 | $3 / $15 | Coding, RAG responses, agentic tasks, most features | Very high volume simple tasks |
| Opus 4.7 | $5 / $25 | Hardest reasoning tasks, architecture, complex debugging | High-volume production inference |
Prompt caching cuts cached input cost by up to 90% across all three. Batch API (async, 24-hour turnaround) cuts both sides by 50%.
When to Use Claude Haiku 4.5
Haiku 4.5 is built for speed and volume. It costs $1 per million input tokens, five times cheaper than Sonnet. For tasks where throughput matters more than depth, it is the correct default.
Use Haiku for:
Routing and classification. Your app receives a user message. Before you do anything expensive, you want to know: is this a support question, a billing question, or a feature request? Is this input safe to process? Does this belong to category A or B? These decisions do not require Sonnet-level reasoning. They need a fast, cheap, accurate classifier. Haiku handles this at a fraction of the cost.
Simple extraction. Pulling a company name from a document. Identifying the date in a sentence. Extracting the price from a product description. Haiku processes these correctly and reliably. Running extraction tasks on Sonnet is spending three times as much for the same result.
Summarization at scale. You are processing 10,000 documents overnight. Each one needs a two-sentence summary. Batch this with Haiku on the Batch API ($0.50 input / $2.50 output per million tokens with the 50% discount) and the cost is negligible.
Customer support triage. First-pass handling of inbound messages. Route the simple ones to automated responses, flag the complex ones for Sonnet or a human. Haiku reads and classifies accurately at high volume.
Where Haiku falls short: Multi-step reasoning, nuanced writing, complex code, anything requiring the model to hold a chain of logic across many steps. Haiku produces plausible-sounding output on these tasks. It is just often wrong in ways that are hard to catch.
When to Use Claude Sonnet 4.6
Sonnet 4.6 is the model most indie hackers should default to for production features. Anthropic's recommended daily driver. It scores 79.6% on SWE-bench Verified, compared to Opus 4.7 at 80.8%. For the vast majority of coding and reasoning tasks, that 1.2-point gap does not show up in practice.
Use Sonnet for:
Interactive coding features. If your SaaS product helps users write, edit, or understand code, Sonnet handles this at production quality. The SWE-bench numbers confirm it: Sonnet is nearly identical to Opus on real-world coding tasks. Running coding features on Opus instead of Sonnet costs 67% more for a 1.2-point benchmark improvement that is invisible to most users.
RAG responses. Retrieval-augmented generation (fetching relevant context from a database and generating a coherent response) is Sonnet's core use case. The model reads context, reasons over it, and produces accurate answers. This is where most SaaS AI features actually live.
Agentic workflows. Multi-step tasks where the model takes actions, observes results, and decides next steps. Sonnet's performance on these is strong, and Claude Code runs on Sonnet by default for exactly this reason. See the Claude Sonnet 4.6 vs Opus 4.7 breakdown for the agentic benchmark comparisons.
Content generation. Marketing copy, documentation, email drafts, product descriptions. Sonnet produces high-quality output at a cost that scales without eating your margin.
Where Sonnet falls short: Genuinely hard reasoning chains, complex architectural analysis across thousands of lines of code, tasks that require holding and cross-referencing many constraints simultaneously. Sonnet gets most of these right, but the failure rate is higher than Opus on the hardest problems.
When to Use Claude Opus 4.7
Opus 4.7 launched on April 16, 2026 at the same $5/$25 per million token pricing as Opus 4.6. One important caveat: Opus 4.7 ships with a new tokenizer that can generate up to 35% more tokens for the same input text. The rate card says unchanged, but the effective cost per request can be meaningfully higher.
Use Opus for:
The 10% of tasks Sonnet gets wrong. You will know these when you see them. Complex debugging across a large codebase. Architectural decisions requiring deep tradeoff analysis. Tasks that require synthesizing information across many sources into a coherent, accurate conclusion. Sonnet makes confident-sounding mistakes on a small percentage of genuinely hard tasks. Opus makes fewer of them.
Evaluating other models' outputs. If you are running a pipeline where one model generates and another evaluates, Opus is the right evaluator. It catches errors that Sonnet misses when used as a judge in LLM-as-judge workflows.
One-off tasks with high stakes. Building an architecture document for a major feature. Analyzing a complex contract. Generating a detailed technical specification. When volume is low and accuracy matters, the Opus premium is justified.
Where Opus does NOT make sense: High-volume production inference. Customer-facing features running hundreds or thousands of calls per day. Classification, extraction, or routing. These are Haiku and Sonnet territory. Opus at high volume is a common early mistake that creates AI bills that make no sense relative to the value delivered.
The Three-Layer Architecture Most Solo Devs End Up Using
After building with the Claude API for a while, most indie hackers converge on the same pattern:
Layer 1: Haiku for triage. Every request goes through Haiku first. Route it, classify it, check if it's safe to process. Cost: near zero.
Layer 2: Sonnet for most features. 80-90% of user-facing features run on Sonnet. Interactive assistant, RAG responses, code help, content generation.
Layer 3: Opus on demand. A small percentage of requests get escalated to Opus. Complex debugging, architectural questions, tasks that explicitly need the best available reasoning.
The cost impact is significant. A SaaS with 50,000 API calls per month at an average of 500 input and 300 output tokens:
- All on Sonnet: (25M × $3 + 15M × $15) / 1M = $75 + $225 = $300/month
- Three-layer (70% Haiku, 28% Sonnet, 2% Opus): roughly $130/month
- Adding prompt caching to the Sonnet layer (80% cache hit rate): drops Sonnet effective input cost from $3 to ~$0.60, bringing total to roughly $90/month
The same 50,000 API calls, the same outputs, two-thirds lower bill.
The Prompt Caching Multiplier
This is where most solo devs leave the most money on the table. If you send the same system prompt with every API call (your product's instructions, your user's profile, your tool definitions), you are paying full price for the same input on every single call.
Prompt caching stores that context on Anthropic's infrastructure and charges 10% of the standard rate for cache reads. On Sonnet, cached input costs $0.30 per million tokens instead of $3.00. A system prompt that runs to 10,000 tokens, repeated across 10,000 daily calls, costs $300/month at full price. With caching, that same context costs $30/month.
Cache writes have a one-time premium (1.25x for a 5-minute cache, 2x for a 1-hour cache), but even the first cache read pays for the write cost on subsequent calls.
If you are building with the Claude API and not using prompt caching yet, this is the highest-leverage change you can make to your AI cost structure.
For a deeper look at how these models compare in a subscription context, the Claude Pro vs ChatGPT Plus breakdown covers the subscription side. For the recent pricing changes from Anthropic's June 2026 restructure, the Anthropic subscription split post has the full context.
Top comments (0)