Claude in 2026: the engineer's field guide

Opus 4.7 costs $5 per million input tokens. Cache it: 50¢. Batch it: 25¢. Same model, same prompt, a twentieth the spend.

Anthropic shipped Opus 4.7 on April 16, 2026 — a 13% coding uplift over 4.6 and 87.6% on SWE-bench Verified. The actually-useful number is the one above. If you build with Claude and you don't think about caching and batching, you are paying twenty times more than you need to.

This is the field guide I send new hires when they ask "what do I need to know about Claude?". Just the load-bearing parts: the lineup, the decision tree, the pricing math, the gotchas, the migration calendar.

TL;DR — Sonnet 4.6 is the default. Opus 4.7 is for agents and hard reasoning. Haiku 4.5 is for high-volume work. Cache anything reused, batch anything async, and check the deprecation calendar before June.

The lineup, April 2026#

Three current models. One sentence each.

• Opus 4.7 (claude-opus-4-7) — $5 / $25 per MTok. 1M context, 128k output. Ship complex agents, hard reasoning, anywhere a wrong answer is worse than a slow answer.
• Sonnet 4.6 (claude-sonnet-4-6) — $3 / $15 per MTok. 1M context, 64k output. Best price-to-quality. Production workloads, daily chat, most code generation.
• Haiku 4.5 (claude-haiku-4-5) — $1 / $5 per MTok. 200k context, 64k output. Classification, extraction, batch labelling — high volume where latency matters more than depth.

Prices haven't moved since Sonnet 4.6 launched in February. The capability has — Opus 4.7 is the same dollars as Opus 4.6 with measurably better coding. Pricing-stable upgrades are the most user-friendly thing Anthropic has done this year.

One detail worth knowing: Opus 4.7 ships with a new tokeniser. The same English text yields fewer tokens than on 4.6. If you have token-budgeting code anywhere, audit it — your "max 100k input tokens" guard now passes for slightly longer prompts.

For reference: OpenAI's GPT-5.5 launched April 23, 2026 at $5 / $30 per MTok. Opus 4.7 matches it on input and undercuts on output by 17%.

Pick the model in four questions#

The decision tree we wrote after six months of building with the API. Read top to bottom, take the first branch that fits.

1. Is the workload a multi-turn agent that edits real code? → Opus 4.7. Don't try to save money here. The cost of one bad agent run dwarfs the model bill.
2. Is it a user-facing feature where p50 latency matters? → Sonnet 4.6 with adaptive thinking off. Opus is moderate-latency; Sonnet is fast.
3. Is it volume work — classifying 100k support tickets, extracting fields from a million PDFs? → Haiku 4.5 with the Batch API. The 50% asynchronous discount stacks on top of Haiku already being the cheapest model. You will save 80%+ versus calling Sonnet synchronously.
4. Are you sitting in a Sonnet output and wondering if Opus would do better? → It probably would. Run 50 representative examples through both before you migrate the workload.

The trap most teams fall into is using Opus everywhere. Opus is overkill for 80% of LLM work. Sonnet 4.6 is the default. Reach for Opus when you actually need it.

The pricing math#

Two features turn the headline sticker into something five-to-twenty times cheaper. They compose.

Prompt caching#

Cache reads cost 10% of the base input price. Writes cost 1.25× for a 5-minute TTL, or 2× for a 1-hour TTL.

Breakeven is two reads — even on the 1-hour write. After that, every cached request runs at a tenth of base.

For 100k tokens of system prompt reused across 10 calls on Opus 4.7:

Mode	Cost
No caching	$5.00
5-minute caching	$1.07
1-hour caching	$1.45

That's a 78% saving on one of the most boringly common LLM patterns: the same instructions and few-shot examples sent every request.

The SDK call:

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();
const res = await client.messages.create({
  model: "claude-opus-4-7",
  max_tokens: 1024,
  system: [
    {
      type: "text",
      text: SYSTEM_PROMPT, // ≥ 4096 tokens for Opus / Haiku
      cache_control: { type: "ephemeral", ttl: "5m" },
    },
  ],
  messages: [{ role: "user", content: userInput }],
});

console.log(res.usage.cache_creation_input_tokens); // > 0 on first call
console.log(res.usage.cache_read_input_tokens);     // > 0 on subsequent

Always log those two usage fields. If both are zero on a call you expected to hit cache, your prefix is too short or a tool definition changed.

Three things to know before you ship caching:

1. Minimum cacheable prefix is 2,048 tokens for Sonnet 4.6, 4,096 for Opus and Haiku 4.5. Below the minimum, caching fails silently.
2. Tool definition changes invalidate the entire cache. If your tool schema is regenerated dynamically per-request, nothing caches. Pin the schema.
3. Cache lifetime is wall-clock. The 5-minute TTL refreshes on every read, so a steady-stream workload stays warm forever. The 1-hour TTL doesn't refresh — it expires hard.

Batch API#

50% off input AND output. You submit a JSONL of requests, results land within 24 hours (usually much faster). Compatible with prompt caching.

Picking the right model + caching + batching is the trifecta. Teams running real volume pay 5% of what naive synchronous usage would cost. Worth a Tuesday afternoon to set up.

Five things the docs don't shout about#

Five gotchas every team I've seen has hit, in order of frequency:

1. Adaptive thinking and extended thinking are different mechanisms. Adaptive (Opus 4.7) decides for itself how long to think. Extended (Sonnet, Haiku) takes a budget you set. Different latency profile, different billing line. Read the thinking docs before you toggle either on.

2. The synchronous Messages API caps Opus output at 128k tokens. The Message Batches API takes a beta header to extend that to 300k. If you've hit a hard 64k limit, you're either on Sonnet or you need to migrate to batch.

3. Knowledge cutoffs differ by model. Sonnet 4.6 is reliable through August 2025. Opus 4.7 is reliable through January 2026. Haiku 4.5 only through February 2025. If you ask Haiku about a library released last summer, expect hallucinations.

4. The May 2025 models retire on June 15, 2026. If you have claude-sonnet-4-20250514 or claude-opus-4-20250514 hard-coded anywhere, calendar that migration before May. The retirement date returns a hard error — not a warning.

5. There's a fourth model you've probably never heard of. Claude Mythos Preview is invitation-only, defensive cybersecurity research as part of Project Glasswing. You can't sign up. Mentioned only because someone will namedrop it in a meeting.

Migration calendar (2026)#

What to action and when:

• Today — anything on claude-sonnet-4 or claude-opus-4 migrates to 4.6 / 4.7. Same prices, better models. No reason to delay.
• Before May 2026 — full migration off the May 2025 models. Add a CI check that fails the build if those IDs appear in code or env vars.
• June 15, 2026 — May 2025 models retire entirely. Outbound calls return a hard error.
• Anytime convenient — Opus 4.5 and 4.1 are legacy but not yet deprecated. Opus 4.1 in particular drops from $15 → $5 input when you migrate to 4.7. That's a 3× cost cut for the same workload.

Anthropic's model deprecations page is the source of truth. Bookmark it; check it monthly.

What you don't need to learn#

A few things show up in every "Claude guide" post and aren't actually load-bearing:

• Constitutional AI as a concept. It's the training philosophy. As a builder, it doesn't change your prompts. Skip the philosophy posts.
• Context window theatre. Opus and Sonnet have 1M tokens of context. You will rarely use more than 200k. Don't pay for retrieval architecture you don't need.
• Whether Claude is "better than GPT-5.5". Depends on the task. Run a 50-example evaluation on your workload. Vendor benchmarks are noise.

The takeaway#

Three models, three prices. Sonnet 4.6 by default, Opus 4.7 for hard problems, Haiku 4.5 for volume. Cache anything reused. Batch anything asynchronous. Migrate off the May 2025 models before June.

That's everything that pays you back this quarter.