Claude Models 2026: Opus 4.8 vs Sonnet 4.6 vs Haiku

Last updated: June 2026 · Author: Ignacy Kwiecień, founder & editor-in-chief, DecodeTheFuture.org

The Claude model lineup at a glance (2026)

Anthropic ships Claude in tiers, each tuned for a different point on the intelligence-versus-cost curve. These are the current-generation models and their June 2026 availability status — older versions like Opus 4.6 and Opus 4.5 are still available but are no longer the default choice.

The names follow a simple pattern: Haiku is small and fast, Sonnet is the balanced middle, Opus is the heavyweight, and Fable sits above all of them on paper as Anthropic’s most capable model. The important availability distinction is that Fable 5 is suspended, so the practical top of the line is Opus 4.8. The number (4.8, 4.6, 4.5) is the generation — higher is newer and generally smarter at the same tier.

Which Claude model should you use?

Skip the spec sheet if you just want the decision. The honest rule of thumb for 2026:

The single most common mistake is reaching for the biggest model by reflex. Opus 4.8 and Fable 5 are remarkable, but running classification or routine summarization on them is paying flagship prices for work Haiku does fine. The cost gap is large: Haiku 4.5 at $1/$5 is 5× cheaper on input and output than Opus 4.8, and 10× cheaper than Fable 5. Match the model to the difficulty of the task, not to the prestige of the name.

Claude Opus 4.8 — the flagship

Claude Opus 4.8 is the model to reach for when the task is genuinely hard. It is Anthropic’s most capable Opus-tier model — state-of-the-art on long-horizon agentic execution, complex coding, knowledge work and memory, with a notably clearer, warmer writing voice than the previous generation. It carries a 1M-token context window at standard pricing (no long-context premium) and up to 128K output tokens. It is the default model in agentic coding tools, and the one we benchmarked in our Cursor vs Claude Code comparison. For a deeper look at what changed in this release, see our Claude Opus 4.8 explainer.

Claude Sonnet 4.6 — the everyday default

Sonnet 4.6 is the model most teams should run by default. It hits the best balance of speed, cost and intelligence in the line: $3/$15 per million tokens, a 1M context window, and 64K max output. For chat interfaces, summarization, extraction, routing and the large majority of production app features, Sonnet 4.6 is fast enough to feel responsive and smart enough that you rarely notice the ceiling. Reserve Opus for the cases where Sonnet visibly struggles — that keeps your blended cost down without hurting quality where it counts.

Claude Haiku 4.5 — fastest and cheapest

Haiku 4.5 exists for volume and speed. At $1/$5 per million tokens with a 200K context window, it is the right tool for classification, content moderation, simple extraction, routing layers and anything you run at high throughput where latency and cost dominate. It is not the model for deep reasoning — but using Opus or Fable for tasks Haiku handles is one of the most common ways teams overspend on LLM bills.

Claude Fable 5 — the most capable model, currently suspended

Fable 5 is technically Anthropic’s most capable model — built for the most demanding reasoning and longest-horizon agentic work, at $10/$50 per million tokens, with thinking always on and a 30-day data-retention requirement. But there is a critical caveat that overrides everything else: Fable 5 was launched on 9 June 2026 and suspended worldwide three days later, on 12 June 2026, under a US export-control directive that blocks access for foreign nationals. In practice that means most users cannot run it today, and you should not plan production work around it until the status changes. We cover the suspension, its ties to the March 2026 supply-chain-risk designation, and the restricted Claude Mythos 5 sibling (Project Glasswing only) in Claude Fable 5 & Mythos 5 suspended and What is Claude Mythos. The upshot: for any model you can actually deploy right now, Opus 4.8 is the real ceiling.

Pricing compared — what you actually pay

All Claude API pricing is per million tokens, split into input (what you send) and output (what the model generates). Output is always more expensive than input.

Two levers cut the bill regardless of model. Prompt caching charges cached input at roughly one-tenth the normal rate, so any large, stable prefix (a system prompt, a long document) you reuse across requests should be cached. And tiering your traffic — Haiku for the easy 80%, Sonnet for the middle, Opus only for the hard tail — almost always beats running everything on one big model.

How thinking and effort work across the line

One practical thing developers should know: the modern Claude models (Opus 4.6 and up, Sonnet 4.6, Fable 5) use adaptive thinking rather than a fixed “thinking budget.” You turn it on and the model decides how much to reason per request. You then tune the trade-off with an effort parameter (low / medium / high / xhigh / max) — higher effort means deeper reasoning and more tokens. On Fable 5, Opus 4.8 and 4.7 the old fixed-budget thinking and the sampling parameters (temperature, top_p) have been removed entirely; steer behaviour with prompting and effort instead.

# Same call, three models — swap the ID to move up or down a tier.
import anthropic
client = anthropic.Anthropic()

resp = client.messages.create(
    model="claude-opus-4-8",            # or "claude-sonnet-4-6" / "claude-haiku-4-5"
    max_tokens=16000,
    thinking={"type": "adaptive"},      # model decides how much to reason
    output_config={"effort": "high"},   # low | medium | high | xhigh | max
    messages=[{"role": "user", "content": "Plan a migration from REST to gRPC."}],
)
print(resp.content[0].text)

# Cheap, high-volume work? Drop to Haiku and lower the effort:
#   model="claude-haiku-4-5", output_config={"effort": "low"}