Claude Models Explained: Opus vs Sonnet vs Haiku

Anthropic ships Claude models in three tiers: Haiku (fast and cheap), Sonnet (the balanced default), and Opus (the frontier flagship). In July 2026 that means Claude Haiku 4.5 at $1/$5 per million tokens, Claude Sonnet 5 at $3/$15, and Claude Opus 4.8 at $5/$25 — pick Sonnet unless you have a measured reason not to.

This page is our hub for the whole lineup: what each tier is for, exact pricing with every discount lever, how to decode Anthropic's version strings, and which models you actually get inside claude.ai, the API, and Claude Code. Where one comparison deserves full depth, we link out to a dedicated article rather than skim it here.

Why Anthropic ships Claude in three sizes

The three-tier structure isn't marketing — it's an engineering trade-off made explicit. Bigger models reason more reliably but cost more and respond slower. Rather than sell one model and make you accept its compromises everywhere, Anthropic lets you match the compute to the job.

The naming carries the metaphor: a haiku is short, a sonnet is structured and mid-length, an opus is the big work. In practice the tiers behave less like "good, better, best" and more like specialists. We route classification and extraction to Claude Haiku 4.5, everyday coding and writing to Claude Sonnet 5, and escalate to Claude Opus 4.8 only when a task's failure cost justifies the premium. That routing habit — not any single model choice — is what actually controls quality and spend.

If you're still deciding between vendors rather than tiers, start one level up with our comparison of Claude vs ChatGPT — this page assumes you've landed on Claude and now need to pick the right size.

Claude models at a glance: the 2026 lineup

Here is the current lineup as of July 2026, in one table:

Two things stand out. First, the price spread is now only 5× between the cheapest and most expensive tier — narrow enough that model choice is about fit, not affordability. Second, Sonnet and Opus both reach a 1M-token context window and Haiku offers 200K, so capacity rarely forces you up-tier the way it did in 2024–2025.

Current versions: release dates and what changed

Version numbers move independently per tier, so "the latest Claude" is really three different answers:

The pattern worth internalizing: each tier eventually inherits the capabilities of the tier above it from a few releases back. Claude Haiku 4.5 performs near where Claude Sonnet 4 did; Claude Sonnet 5 now trades wins with Claude Opus 4.8 on knowledge work — the latest in a run of Opus releases stretching back to Claude 3. If your workload runs fine on today's Sonnet, there's a good chance next year's Haiku handles it for a third of the price. We re-test our own pipelines against the tier below every time a new release lands, and the downgrade succeeds more often than not.

Claude Sonnet 4.6 ($3/$15) remains available for teams pinned to it in production, but Claude Sonnet 5 costs the same at standard rates and less during the introductory window, so new projects have no reason to start on 4.6.

Claude API pricing in full

API pricing is per million tokens (MTok), input and output priced separately. Here's the complete matrix including both discount levers:

The two levers matter more than the base rates for real workloads:

• Batch processing (−50%) applies when you submit jobs asynchronously and accept results within 24 hours. Any pipeline that isn't interactive — nightly summarization, bulk classification, report generation — should run through the batch API by default. Half price for changing an API endpoint is the easiest cost win in this ecosystem.
• Prompt caching (−90% on cached input) applies when requests share a long common prefix — a system prompt, a codebase snapshot, a document under discussion. Agents and chat apps re-send that prefix on every turn, so caching routinely cuts effective input cost by 60–80% in our own logs.

Stacked together, a cached, batched Sonnet 5 workload at introductory rates pays a tenth or less of naive Opus list price. When someone tells us Claude is "too expensive," the fix is almost always routing and caching, not switching vendors.

Context windows: up to 1M tokens

Claude Opus 4.8 and Claude Sonnet 5 support a 1M-token context window — roughly 750,000 words, or a mid-sized codebase plus its documentation in a single request. Claude Haiku 4.5 caps at a still-generous 200K tokens, plenty for most single-document and pipeline work.

Practically, this changed how we choose models. Through 2024–2025, context capacity was a reason to buy the bigger model; now it's rarely the constraint, and the differentiator is reasoning reliability over long contexts, where Opus still degrades least. Remember that input tokens are billed per request — filling a 1M window with Opus costs $5 per call before output. Prompt caching exists precisely so you don't pay that repeatedly for the same material.

How to pick a Claude model by task

The decision comes down to one question: what happens if the model is slightly wrong? If mistakes are cheap to catch, buy speed and volume. If they compound invisibly, buy reasoning.

Our standing rule: start on Sonnet, escalate to Opus only when Sonnet measurably fails, and downgrade to Haiku wherever outputs are mechanically verifiable. Most teams that run this experiment stay on Sonnet for far more than they expected.

The Sonnet-or-Opus call is the one with real money and real quality on the line, so we've given it a full head-to-head with benchmark data and worked cost scenarios — see Claude Sonnet vs Opus: which model should you actually use?

Claude model names decoded

Anthropic's API model strings look cryptic until you see the pattern. Take claude-sonnet-4-5-20250929: it reads family-tier-version-snapshot. claude is the family, sonnet is the tier (haiku, sonnet, or opus), 4-5 is the version number with dashes standing in for dots — so, Sonnet 4.5 — and 20250929 is the snapshot date, September 29, 2025.

Three practical rules follow:

1. Pin snapshots in production. The dated string always returns the exact same model. Aliases without a date resolve to the newest snapshot and can change behavior under you without warning.
2. Version numbers only compare within a tier. Claude Haiku 4.5 is not "almost Sonnet 4.6" — the tiers are different model sizes that happen to share a versioning scheme.
3. The marketing name maps directly. "Claude Sonnet 5" in a press release is claude-sonnet-5-<date> in the API. If a tool's model picker shows a string you don't recognize, decode it segment by segment rather than guessing.

Models inside the products: claude.ai, API, and Claude Code

The same models surface differently depending on where you meet them, and the pricing logic flips between products:

On subscriptions the strategy inverts from the API: you're not paying per token, so use the best model your usage limits allow. Pro users burn through caps quickly on Opus, so Sonnet stretches a session much further; Max 20x subscribers can comfortably default to Opus for interactive work.

Claude Code — Anthropic's agentic coding tool, available as a terminal CLI, VS Code and JetBrains extensions, a desktop app, and on the web at claude.ai/code — makes the tier system tangible: you can switch models mid-session, so escalation from Sonnet to Opus costs one command. How that compares to routing models inside other editors is exactly what we cover in Cursor vs Claude Code. On the knowledge-work side, Claude Cowork runs the same models against files and documents instead of codebases.

Upgrade cadence: how to stay current

Anthropic shipped Claude Opus 4.8 in May 2026 and Claude Sonnet 5 five weeks later — a tier refresh roughly every one to two quarters is the recent rhythm. That pace rewards a process, not vigilance:

1. Keep a small eval set. Ten to twenty real tasks from your own workload, with known-good outputs. When a release lands, run it before reading anyone's benchmark commentary — including ours.
2. Exploit introductory pricing windows. Claude Sonnet 5's $2/$10 intro rate runs until August 31, 2026. Anthropic has used launch discounts to accelerate migration; treat them as a scheduled deadline for your own testing.
3. Test the tier below on every release. The lineup's consistent pattern is capability flowing downhill. Yesterday's Sonnet workload is often today's Haiku workload at a third of the cost.
4. Watch deprecation notices. Older snapshots retire on published schedules. Pinned production models buy you stability, not immortality — calendar the sunset dates.

We maintain deep dives for each individual model and a full guide to Claude's subscription plans as separate cluster articles; this hub will link to each as it publishes.

Claude pricing at a glance

For the full breakdown of every plan, see our how much Claude costs guide.