Claude Sonnet 5 Review: Opus Power at Half Price
The Short Version
Claude Sonnet 5 went live on June 30, 2026. In one sentence: this is the first Sonnet that makes you question whether you need Opus at all.
After running benchmarks and real-world tasks, Sonnet 5’s agentic task completion rate is within striking distance of Opus 4.8 — and in some cases matches it. At 40% of the cost.
If you’re still on Sonnet 4.6, the upgrade is a no-brainer. If you’ve been debating whether to pay for Opus, Sonnet 5 might save you the expense.
What Actually Changed
From “It Can Start” to “It Can Finish”
Previous Sonnet models had a well-known problem: they’d quit halfway through complex tasks. Ask it to refactor a module spanning 5 files, and it would finish 3 before “summarizing” and waiting for another prompt.
Sonnet 5 largely fixes this. Early access partners consistently reported the same thing — tasks that used to stall now run to completion.
A concrete example: one engineer asked Sonnet 5 to investigate a bug. Unprompted, it wrote a reproducing test, implemented the fix, stashed the fix to confirm the bug returned, then restored the fix and committed. Single pass, no human intervention.
That kind of self-verification behavior was essentially absent in Sonnet 4.6.
Coding Performance Jumps a Tier
Here are the numbers from Anthropic’s official release:
| Benchmark | Sonnet 4.6 | Sonnet 5 | Opus 4.8 |
|---|---|---|---|
| SWE-bench Verified | 65.3% | 72.1% | 75.6% |
| Terminal-bench | 43.2% | 55.7% | 59.1% |
| Humanity’s Last Exam (tools) | 46.8% | 58.3% | 66.1% |
| OSWorld-Verified | 78.5% | 88.2% | 91.4% |
| BrowseComp | 39.0% | 55.8% | 60.5% |
Key takeaways:
- SWE-bench Verified jumped from 65.3% to 72.1%, closing in on Opus 4.8’s 75.6%. The gap is now single digits.
- BrowseComp (agentic search) saw the largest gain — from 39% to 55.8%, nearly 1.5x improvement.
- OSWorld-Verified hit 88.2%, showing Sonnet 5 is mature for browser automation and desktop tasks.
Agentic Capability: The Real Selling Point
Sonnet 5 is positioned as “the most agentic Sonnet yet.” This isn’t marketing talk. The improvements are real:
The model now handles:
- Multi-step planning with sequential execution
- Correct tool selection (browser, terminal, APIs) with fallback strategies on failure
- Sustained context coherence across long tasks without “forgetting” the goal
- Proactive output verification without being asked
Anthropic provides cost-performance curves at different effort levels. The short version:
- At medium effort, Sonnet 5 delivers far better value than both Sonnet 4.6 and Opus 4.8
- At high effort, Sonnet 5 matches Opus 4.8 on certain tasks
- You control this via the API’s
effortparameter — use low effort for simple tasks to save money, crank it up for complex work
Bottom line: one model, no switching. Covers Q&A to complex agents.
Pricing: The Launch Window Matters
| Introductory (through Aug 31) | Standard | Opus 4.8 | |
|---|---|---|---|
| Input (per 1M tokens) | $2 | $3 | $5 |
| Output (per 1M tokens) | $10 | $15 | $25 |
Important details:
- Introductory pricing expires August 31 — roughly 67% of the standard price. If you’re planning to migrate, start testing now.
- Sonnet 5 uses a new tokenizer that consumes ~1.0-1.35x more tokens for the same text. The introductory pricing is designed to make this transition roughly cost-neutral.
- Combined with prompt caching (up to 90% savings) and batch processing (50% savings), real-world costs can drop significantly.
Compared to Opus 4.8, Sonnet 5’s standard pricing is 60% on both input and output. Given the performance gap has narrowed to single-digit percentages, Sonnet 5 offers better ROI for most workloads.
Safety: Measurably Better Than Before
This isn’t boilerplate — there’s data behind it:
- Lower hallucination rate than Sonnet 4.6
- Lower sycophancy than Sonnet 4.6 — won’t fabricate answers to please you
- Improved agentic safety — better at refusing malicious requests and resisting prompt injection
- Overall alignment score better than Sonnet 4.6, though slightly behind Opus 4.8
One thing to note: Sonnet 5 ships with cyber safeguards enabled by default, same as Opus 4.7 and 4.8. It won’t help you write exploit code. If you have legitimate security testing needs, you’ll need to apply through Anthropic’s Cyber Verification Program or use Opus 4.8.
Who Should Upgrade
Upgrade Now
- Teams using Sonnet 4.6 for agents/automation: The improvement in multi-step task completion is decisive.
- Individual developers using Claude Code: Sonnet 5 is already the default on Free and Pro plans — you might already be using it.
- Anyone considering Opus but hesitant on price: Try Sonnet 5 at high effort first. It might be enough.
Wait and See
- Teams already on Opus 4.8 and satisfied: Opus still leads in reasoning depth and the hardest tasks. No reason to downgrade.
- Cybersecurity testing workflows: Sonnet 5’s safeguards are restrictive. Opus 4.8 is better suited.
No Action Needed
- Casual chat / lightweight task users: Sonnet 4.6 or even Haiku already covers you.
Where to Use It
Sonnet 5 is available broadly:
- Claude.ai — Web, iOS, Android. Default model for Free and Pro users
- Claude Code — CLI development tool
- API — Model name:
claude-sonnet-5 - Cloud platforms — AWS (Amazon Bedrock), Google Cloud (Vertex AI), Microsoft Foundry
How It Stacks Up Against Competitors
The mid-2026 AI model landscape is crowded. Here’s where Sonnet 5 sits:
| Model | Position | Input $/1M | Output $/1M | Agentic Strength |
|---|---|---|---|---|
| Claude Sonnet 5 | Mid-high versatile | $2-3 | $10-15 | Strong |
| Claude Opus 4.8 | Flagship reasoning | $5 | $25 | Strongest |
| GPT-5.5 | Flagship coding | $5 | $30 | Strong |
| DeepSeek V4 Pro | Budget reasoning | $0.44 | $0.87 | Medium |
| GLM-5.2 | Bilingual coding | $0.93 | $3 | Medium |
Sonnet 5’s position is clear: it’s not the cheapest, and it’s not the strongest, but it hits the best balance between “good enough” and “affordable.” For teams that need reliable agentic capability without paying flagship prices, Sonnet 5 is one of the most sensible choices in mid-2026.
Bottom Line
Anthropic’s release cadence keeps accelerating — Sonnet 3.5, 3.6, 3.7, 4, 4.5, 4.6, now 5. The version number jump signals this is not a minor iteration.
Sonnet 5’s core narrative is “agentic democratization.” Multi-step autonomous tasks that previously required Opus-class models now work at the Sonnet price point. This has real implications for the cost structure of AI application development.
Introductory pricing runs through August 31. If you’re building any form of AI agent, start testing your migration now.