Model Comparison

Opus vs Sonnet vs Haiku: Choosing a Claude Tier

Three tiers, three prices, three very different jobs. Here is the verified data on Opus 4.8, Sonnet 5, and Haiku 4.5, plus the cost math and routing patterns that make the choice mechanical instead of vibes-based.

By Kylian Migot · Updated July 2026 · 9 min read

Quick answer

Start with Claude Opus 4.8 for serious agentic work, it is Anthropic's own recommended starting point for complex agentic coding. Route routine and high-volume work to Claude Sonnet 5, intro-priced through August 31. Drop to Claude Haiku 4.5 for subagents and mechanical edits. Above all three sits Claude Fable 5, the frontier escalation path at twice Opus's price.
Claude Opus 4.8 price
$5 in / $25 out per MTok
Claude Sonnet 5 price
$3 in / $15 out per MTok (Intro pricing $2/$10 through Aug 31, 2026)
Claude Haiku 4.5 price
$1 in / $5 out per MTok
Opus headline
SWE-bench Verified 88.6% (swebench.com leaderboard)
01

The Three Tiers, Side by Side

The Claude lineup in July 2026 is a clean ladder: Claude Haiku 4.5 (October 2025) at the bottom, Claude Sonnet 5 (June 30, 2026) as the new mid-tier default, Claude Opus 4.8 (May 28, 2026) as the flagship, and Claude Fable 5 above all of them in the Mythos-class frontier tier. This page is about the three tiers you will actually route most work between:

Claude Opus 4.8Claude Sonnet 5Claude Haiku 4.5
Price (per MTok)$5 in / $25 out$3 in / $15 out$1 in / $5 out
Price noteFast mode $10/$50Intro pricing $2/$10 through Aug 31, 2026
ReleasedMay 28, 2026June 30, 2026October 2025
Context window1M tokens1M tokens (default)200k tokens
TierFlagshipWorkhorseFast/cheap
SWE-bench Verified88.6% (swebench.com leaderboard)Not in our verified setNot in our verified set
SWE-bench Pro69.2% (Scale leaderboard)Not in our verified setNot in our verified set
Claude Code access/model opus (recommended starting point for complex agentic coding)/model sonnet/model haiku
02

What Each Tier Is Actually For

Claude Opus 4.8: the recommended starting point

The flagship workhorse for complex agentic coding, Anthropic's own recommended starting point. Anthropic's guidance is unusually direct here: Opus is the recommended starting point for complex agentic coding, and the 88.6% SWE-bench Verified / 69.2% SWE-bench Pro scores back it up. Full breakdown on the Opus 4.8 page.

Strengths

  • Strongest price-to-capability ratio for hard coding tasks
  • Effort defaults to high, thorough on multi-file changes
  • Half Fable 5's price with most of the capability

Trade-offs

  • Fable 5 measurably ahead on the hardest benchmarks
  • Slower than Sonnet on quick iterations

Best for

  • Day-to-day feature implementation on real codebases
  • Complex agentic coding sessions
  • Code review passes on big diffs

Claude Sonnet 5: routine work and volume

The new mid-tier default: 1M context standard, new tokenizer, replaces Sonnet 4.6. The intro pricing (Intro pricing $2/$10 through Aug 31, 2026) makes it the cheapest serious coding model right now, and the 1M-token default context removes the old mid-tier penalty of not fitting the repo. Full breakdown on the Sonnet 5 page.

Strengths

  • 1M context at mid-tier pricing
  • Fast enough for tight edit-test loops
  • Intro pricing makes it the cheapest serious coding model right now

Trade-offs

  • Noticeably behind Opus/Fable on hard multi-step reasoning
  • New tokenizer changes token-count budgeting vs Sonnet 4.6

Best for

  • Routine feature work, tests, and refactors
  • High-volume parallel agent fleets where cost dominates
  • Fast iteration loops with human review

Claude Haiku 4.5: subagents and mechanical work

The fast/cheap tier: sub-agents, classification, quick mechanical edits. It is the oldest model in the current lineup (October 2025) and the only one still on a 200k tokens window, but for fan-out work where you spawn ten cheap subagents, nothing else in the family comes close on cost. Full breakdown on the Haiku 4.5 page.

Strengths

  • Cheapest Claude tier by far
  • Low latency, good for sub-agent fan-outs

Trade-offs

  • Oldest model in the current lineup (Oct 2025)
  • 200k context, smallest window in the family
  • Not built for complex autonomous coding

Best for

  • Cheap sub-agents inside larger workflows
  • Lint-fix style mechanical changes
  • Summarization and triage

Ship your first agent today

Download AIDEN free and point it at your existing Claude Code or Codex setup. No credit card, running in minutes.

Download AIDEN free

Free to start · macOS 12+ · No credit card required

03

Cost per Story: The Actual Numbers

Per-MTok prices are abstract; stories are not. Take a realistic agentic story, 300k input tokens of repo context, spec, and tool results, 60k output tokens of code and reasoning, and the tier choice turns into dollars:

300k in / 60k outClaude Opus 4.8Claude Sonnet 5Claude Haiku 4.5
Input cost$1.50$0.90$0.30
Output cost$1.50$0.90$0.30
Story total$3.00$1.80$0.60

So: $3.00, $1.80, $0.60 per story at standard pricing, and Sonnet is even cheaper than that until August 31 under its intro pricing (Intro pricing $2/$10 through Aug 31, 2026). For scale, the same story on Claude Fable 5 at $10/$50 per MTok runs about $6.00. The gaps look dramatic in percentage terms and small in absolute terms, which is exactly why the right question is not “which is cheapest” but “which tier fails least on this story”: one failed $1.80 Sonnet run plus a $3.00 Opus retry costs more than starting on Opus. The full ladder, including Fable 5, is mapped in the best Claude model for coding.

04

The Opusplan Pattern: Plan High, Implement Cheap

The pattern the community converged on is plan with Opus, implement with Sonnet, popularized by Claude Code's opusplan model setting, which uses Opus in plan mode and hands execution to Sonnet. The logic: architectural mistakes are the expensive kind, so spend the premium tokens where mistakes compound, then let the workhorse type out the plan.

The math holds up. On our 300k-in/60k-out story, suppose planning takes 50k input and 10k output tokens and implementation takes the remaining 250k/50k. All-Opus costs $3.00; Opus-plans-Sonnet-implements costs about $2.00, a third off, while keeping the flagship on the decisions that matter. On a hard story you flip it and stay on Opus throughout; on a routine one you skip Opus entirely and let Sonnet plan too.

05

Fleet Economics: Mixing Tiers Across Parallel Agents

The tier question changes shape once you run agents in parallel. A common setup is a Sonnet fleet with an Opus lead: the one genuinely hard story goes to Opus, the four routine ones run on Sonnet simultaneously. Five stories at our worked rates: $10.20 for the mixed fleet versus $15.00 for all-Opus, roughly a third cheaper for the same wall-clock time, because the stories that did not need the flagship did not pay for it.

The catch is that per-story routing is tedious to do by hand across terminal tabs. That is the part AIDEN automates: it runs your existing Claude Code CLI on a kanban board where each story is a card with its own git branch and its own model assignment, so Opus takes the migration story while Sonnet clears the backlog next to it. The full lineup, including the GPT-5.6 side, lives at AI models for coding.

FAQ

Which Claude tier should I use for coding?
Start with Claude Opus 4.8 ($5/$25 per MTok), Anthropic's own recommended starting point for complex agentic coding. Move routine feature work, tests, and refactors to Claude Sonnet 5 ($3/$15 per MTok, intro pricing $2/$10 through aug 31, 2026). Use Claude Haiku 4.5 ($1/$5 per MTok) for subagents, triage, and mechanical edits. Reserve Claude Fable 5 for the hardest cross-cutting work.
Is Opus 4.8 worth it over Sonnet 5?
For hard multi-step work, yes. Opus 4.8 posts 88.6% on SWE-bench Verified and 69.2% on SWE-bench Pro, while Sonnet 5 has no published benchmark scores in our verified set yet, and Anthropic positions it as noticeably behind Opus on hard multi-step reasoning. The cost gap on a realistic 300k-in/60k-out story is $3.00 on Opus versus $1.80 on Sonnet at standard pricing, small enough that one failed Sonnet run on a hard story erases the savings.
When is Haiku 4.5 enough?
When the task is mechanical or the model is a subagent inside a larger workflow: lint-fix style changes, summarization, triage, and cheap fan-outs. At $1/$5 per MTok it is the cheapest Claude tier by far, but it is the oldest model in the lineup (October 2025), has the smallest context window (200k tokens), and is not built for complex autonomous coding.
What is the opusplan pattern?
Plan with Opus, implement with Sonnet. You let Opus 4.8 do the expensive thinking, reading the repo, drafting the plan, making the architectural calls, then hand the approved plan to Sonnet 5 for the mechanical implementation. On our worked 300k-in/60k-out story, routing the planning tokens to Opus and the implementation tokens to Sonnet costs about $2.00 versus $3.00 for all-Opus, with most of the quality where it matters.
Where does Claude Fable 5 fit?
Above all three. Fable 5 is Anthropic's frontier Mythos-class tier at $10/$50 per MTok, twice Opus 4.8's price, with the best published SWE-bench Verified and SWE-bench Pro scores of any generally available model. It is the escalation path when Opus is not enough, not the daily driver: the same 300k-in/60k-out story costs about $6.00 on Fable 5 versus $3.00 on Opus.

Keep reading

Route every story to the right tier.

AIDEN runs your Claude Code CLI on a kanban board: Opus on the hard story, Sonnet on the routine ones, each on its own git branch. Free for one project.

macOS 12+ · Bring your own Claude Code or Codex · Your code stays local