What's the price difference between Claude Sonnet 5 and Opus 4.8?

Sonnet 5 is $2 input / $10 output per 1M tokens (introductory through Aug 31, 2026), then $3/$15. Opus 4.8 is $15 input / $75 output. Sonnet 5 is roughly 5-7x cheaper at introductory pricing and 5x cheaper at standard pricing.

How does Sonnet 5 compare to Opus 4.8 on benchmarks?

On Anthropic's agentic coding benchmark, Sonnet 5 scores 63.2% vs Opus 4.8's 69.2% — a 6-point gap. On knowledge work benchmarks, Sonnet 5 matches or slightly exceeds Opus 4.8 in some categories. The gap is meaningful on the hardest tasks but not on the median task.

When should I still pick Opus 4.8 in July 2026?

Pick Opus 4.8 when the task is genuinely hard (architecture design, complex debugging, long autonomous runs where a small quality drop compounds), when the output tokens per request are low enough that cost doesn't matter, or when you're already on a fixed Claude Max plan and Opus doesn't burn extra credits.

Is Sonnet 5 a drop-in replacement for Sonnet 4.6?

Yes. Sonnet 5 uses an updated tokenizer for better throughput and improves on Sonnet 4.6 across coding, reasoning, and agent benchmarks. Anthropic positions it explicitly as the default for anyone previously on Sonnet 4.6 or Sonnet 4.7.

Quick Answer

Claude Sonnet 5 vs Opus 4.8: Which One to Pick (July 2026)

Published: July 1, 2026

Claude Sonnet 5 vs Opus 4.8: Which One to Pick (July 2026)

Anthropic’s June 30, 2026 release of Claude Sonnet 5 changes the default answer to “which Claude should I use?” For the first time, the mid-tier model gets close enough to the flagship that Opus is no longer the automatic pick for serious work — but it’s not obsolete either. Here’s when each one earns its price tag.

Last verified: July 1, 2026

The numbers

Metric	Claude Sonnet 5	Claude Opus 4.8
Released	June 30, 2026	January 2026
Input price (per 1M tokens)	$2 (intro) / $3 (after Sep 1)	$15
Output price (per 1M tokens)	$10 (intro) / $15 (after Sep 1)	$75
Agentic coding benchmark	63.2%	69.2%
Positioning	Balanced mid-tier	Flagship reasoning
Availability	Free, Pro, Max, Team, Enterprise, API, Bedrock, Copilot	Pro, Max, Team, Enterprise, API

The core tradeoff

Sonnet 5 gets you ~91% of Opus 4.8’s agentic coding score at ~1/5 to 1/7 the cost. That ratio is what makes this release important. Historically, Opus was 3-5x more expensive but delivered a real quality lift; Sonnet 5 compresses the quality gap without much movement on the price gap.

If you were paying Opus prices to squeeze out the last few benchmark points, the math has changed.

When Sonnet 5 is the right default

Pick Sonnet 5 for:

Everyday coding automation — Cursor, Claude Code, Copilot, in-house agents. The 6-point benchmark gap disappears in noise for most tasks.
High-volume agent runs — background workers, batch processing, anything where you’re spending real money on tokens
Any Sonnet 4.6 / 4.7 workload — drop-in upgrade with better throughput via the updated tokenizer
Cost-sensitive startups — introductory $2/$10 pricing through Aug 31 is aggressive; lock in the workflow before standard pricing kicks in
New projects — start on Sonnet 5, promote specific hard subtasks to Opus if benchmarks say so

When Opus 4.8 still earns its 5x premium

Pick Opus 4.8 for:

Architecture and design work — the 6-point benchmark gap tends to concentrate on the hardest tasks (complex system design, refactoring large codebases, deep debugging). If a mistake here costs a day of engineer time, Opus’s premium is trivial.
Long autonomous runs — small per-step quality drops compound over 20+ tool calls. Opus loses that compounding problem less often.
Fixed-price Claude Max plans — if you’re on Max and Opus doesn’t burn extra credits, use Opus by default. It costs you nothing extra.
Low-output-volume, high-value tasks — legal analysis, medical review, research synthesis, where output tokens are limited but stakes are high
When the task is at the edge of what LLMs can do — the top of the capability curve is still Opus

A pragmatic routing strategy

Instead of picking one, route based on task complexity:

Task classifier (Haiku or Sonnet 5) →
  Simple / medium tasks → Sonnet 5
  Hard / high-stakes tasks → Opus 4.8
  Debug / retry loop → escalate one tier up

This is the same pattern GPT-5.6 formalizes with its reasoning_effort parameter — dial the cost/accuracy tradeoff per request rather than per app.

What changed with Sonnet 5 specifically

Beyond the benchmark bump, Sonnet 5 ships three practical improvements:

Updated tokenizer — better throughput per dollar; some prompts now use fewer tokens than in Sonnet 4.6
Improved tool use — noticeably better at driving browsers and terminals autonomously, closing more of the agent-completeness gap
Wider distribution — available on Free plans, GitHub Copilot, and AWS Bedrock at launch (Opus is still gated to Pro+ tiers)

That last point matters for infrastructure decisions: if your product needs a “good default model for free users,” Sonnet 5 is now the answer.

Migration checklist (July 2026)

✅ Currently on Sonnet 4.6 / 4.7: switch to Sonnet 5. It’s strictly better on price and benchmarks.
✅ Currently on Opus 4.8 for everything: benchmark your specific workload on Sonnet 5. If quality holds, save 80%+.
✅ Currently on Opus 4.8 for hard tasks only: no change. Opus still owns the top of the curve.
✅ Currently on GPT-5.5 or Gemini 3.5 Flash: Sonnet 5’s intro pricing makes it competitive; A/B test on your workload.

The bottom line

Sonnet 5 is now the default Claude model for most work. Opus 4.8 is now a specialist tool for hard problems. That’s the inverse of the March 2026 answer, where Opus was the default and Sonnet was the cost-saver. Anthropic just moved the price/performance frontier, and the pragmatic play is to route accordingly.

Last verified: July 1, 2026. Sources: Anthropic Claude Sonnet 5 announcement, DataCamp Sonnet 5 review, llm-stats.com benchmarks, AWS Bedrock announcement.