Claude Sonnet 5 & Opus 4.8 for Coding Agents: Benchmarks Breakdown

Anthropic’s current frontier lineup — Opus 4.8 (shipped May 28, 2026) and Sonnet 5 (shipped June 30, 2026) — reports different strengths on different benchmarks:

• Opus 4.8: 69.2% on SWE-bench Pro, the leading score among the two on agentic coding
• Sonnet 5: 63.2% on SWE-bench Pro, up from Sonnet 4.6’s 58.1%
• Sonnet 5: 81.2% on OSWorld-Verified, a computer-use/tool-driving benchmark
• Sonnet 5: 1,618 on GDPval-AA v2 (knowledge work), essentially matching Opus 4.8’s 1,615

This post breaks down what each number measures and what it means for a coding agent you actually run in production on Paperclip.

TL;DR: If your agent is a deep, multi-file autonomous coder, Opus 4.8’s SWE-bench Pro lead (69.2% vs 63.2%) is the number that matters most. If your agent mixes coding with tool use, browsing, or general knowledge work, Sonnet 5 gets you most of the quality at roughly a third of the price.

SWE-bench Pro — the agent-relevant coding number

What it measures: SWE-bench Pro evaluates models on real-world software engineering tasks — reading a codebase, understanding an issue, and submitting a patch that passes hidden tests. It’s the closest widely-used public benchmark to a real coding agent’s day-to-day work.

Scores: Opus 4.8 at 69.2%, Sonnet 5 at 63.2%, Sonnet 4.6 at 58.1%.

What this means in your agent:

• Opus 4.8 completes more hard, multi-file tasks without giving up mid-run
• Sonnet 5’s jump over Sonnet 4.6 (+5.1 pp) means agents that were previously escalating to Opus for “safety margin” alone may now do fine on Sonnet 5
• The remaining ~6-point gap between Sonnet 5 and Opus 4.8 concentrates on the hardest tasks in most teams’ reports — simple tasks show little difference

OSWorld-Verified — the computer-use number

What it measures: OSWorld-Verified evaluates a model’s ability to drive a real computer environment — clicking, typing, navigating applications and browsers to complete a task, not just writing code in a sandbox.

Sonnet 5 score: 81.2%. Anthropic doesn’t report an Opus 4.8 number on this specific benchmark, and Sonnet 5 is explicitly positioned as the more agentic, tool-driving model of the two.

What this means in your agent:

• Browser-automation agents, form-filling agents, and any agent that operates a UI rather than an API should default to Sonnet 5
• This is one of the clearest signals that “bigger model = better agent” no longer holds uniformly — Sonnet 5 is the stronger pick for this task shape specifically

GDPval-AA v2 — the knowledge-work number

What it measures: GDPval-AA v2 evaluates general knowledge-work tasks — the kind of judgment-heavy, non-coding work agents do in support, research, and operations roles.

Scores: Sonnet 5 at 1,618, Opus 4.8 at 1,615 — essentially a tie.

What this means in your agent:

• For support, research-summary, and operations agents that aren’t doing deep coding, Opus 4.8 offers no real quality advantage over Sonnet 5 — only a much higher price
• This is the benchmark that should drive most “which model do I default new agents to” decisions

What drives the differences

Opus 4.8’s coding edge comes from targeted improvements to code review: Anthropic reports it’s roughly 4× less likely than its predecessor to let a flaw pass unremarked. That directly reduces the “agent opened a PR but it’s subtly broken” failure mode.

Sonnet 5’s tool-use edge comes from being built specifically to plan and drive tools — browsers, terminals, multi-step workflows — at a level that previously required an Opus-class model.

Sonnet 5’s updated tokenizer is a separate, easy-to-miss factor: the same input can map to 1.0–1.35× more tokens than on Sonnet 4.6. When comparing cost-per-task across model generations, measure actual token counts rather than assuming the sticker price scales linearly.

What the benchmarks don’t measure

Reliability under production prompts. Benchmarks use clean, researcher-tuned prompts. Your production prompts may carry baggage compensating for older model quirks — worth re-testing simplified prompts now.

Tool ecosystem quality. A stronger model with weak tools (bad code search, flaky test runner) still produces mediocre agent results. The model score is a ceiling, not a guarantee.

Real-world repo scale. SWE-bench Pro tasks are real but curated. Your production repo may have far more history and unfamiliar conventions than any benchmark set — expect benchmark deltas to be directional, not exact.

How to measure this in your own agent

Don’t trust vendor benchmarks blindly. Run your own comparison over 48–72 hours:

1. Fix a frozen task set — 50-100 recent tasks covering your real difficulty distribution.
2. Run the set on your current model (baseline) — record pass/fail, tokens, latency, tool calls.
3. Run the set on Sonnet 5 and, separately, Opus 4.8 — same recording.
4. Compute the delta on pass rate, tokens per successful task, and tool call count.
5. Decide per agent. Coding-heavy agents may justify Opus 4.8’s price; most others won’t.

On Paperclip, this loop is a dashboard action: clone the agent, swap the model, replay tasks, compare metrics — no infra work.

Practical recommendation

1. Today: Default new agents to Sonnet 5 unless they’re deep, multi-file coding agents.
2. This week: Re-run your golden task set against Sonnet 5 for any agent currently on Opus 4.7 or 4.6 to see if it can move down in cost without a quality hit.
3. Keep on Opus 4.8: agents where SWE-bench-Pro-style correctness is the bottleneck, or that need Dynamic Workflows’ large-scale parallel subagent fan-out.

FAQ

Does Opus 4.8 replace Sonnet 5, or vice versa? Neither — they’re different tools. Opus 4.8 leads on the hardest autonomous coding; Sonnet 5 leads on cost/performance and computer-use/tool-driving tasks.

Is Opus 4.8 always better for coding? For the hardest multi-file tasks, usually yes. For routine coding tasks, Sonnet 5 is frequently indistinguishable in output quality at a third of the price.

How does this lineup compare to OpenAI and Google’s current models? See our full comparison of Sonnet 5, Opus 4.8, GPT-5.6, and Gemini 3.1 Pro.

Can I run these self-hosted? No, both are API-only. If self-hosting matters, open-weight alternatives trail significantly on agent benchmarks but give you full data control.

---

Related: Deploy Sonnet 5 & Opus 4.8 on Paperclip → · Migrate to the current Claude lineup → · Sonnet 5 & Opus 4.8 vs GPT-5.6 vs Gemini →