Claude Sonnet 5 Is Here: Features, Benchmarks & Pricing

Introduction

Anthropic just dropped Claude Sonnet 5 on June 30, 2026, and it is a big deal. For the first time, a Sonnet-class model closes in on Opus-tier performance across coding, reasoning, tool use, and knowledge work, all while keeping costs well below what Opus commands. If you have been running Opus 4.8 and wincing at the bill, or stuck on Sonnet 4.6 wishing it were smarter, Sonnet 5 occupies exactly the space you have been waiting for.

This article breaks down what changed, what the benchmarks actually tell us, how the new effort levels work, what the cybersecurity safeguards mean for your workflows, and whether you should switch today.

Why Sonnet 5 Matters Right Now

The Sonnet line has always been Anthropic's workhorse. Claude Sonnet 3.5 and 3.6 were the models that convinced developers that AI-assisted coding could actually ship production code. Sonnet 3.7 introduced extended thinking. Sonnet 4.6 refined everything further. But over the past several months, the most dramatic capability jumps happened in the Opus line, specifically with Opus 4.7 and 4.8. Developers who needed top-tier agentic performance had to pay Opus prices to get it.

Sonnet 5 changes that equation. Anthropic describes it as their most agentic Sonnet yet, a model that can make plans, drive tools like browsers and terminals, and run autonomously at a level that previously required larger, more expensive models. Early access partners described a model that finishes complex multi-step tasks where previous Sonnet models would stall halfway, checks its own output without being told to, and does it all at a price point that makes economic sense for production workloads.

Benchmark Performance: The Numbers That Matter

Let us look at how Sonnet 5 actually performs against its predecessor and the current flagship.

On SWE-bench Pro, the gold standard for real-world software engineering, Sonnet 5 scores 63.2 percent. That is a meaningful jump from Sonnet 4.6, though Opus 4.8 still leads at 69.2 percent. For most development teams, though, 63.2 percent on SWE-bench Pro represents a model that can handle genuinely complex codebases.

On OSWorld-Verified, which tests agentic computer use, Sonnet 5 hits 81.2 percent compared to Opus 4.8's 83.4 percent. That is a remarkably narrow gap for a model that costs roughly half as much.

The most interesting result might be Terminal-Bench 2.1, where Sonnet 5 scores 80.4 percent, actually surpassing Opus 4.8's 74.6 percent. If your workflows lean heavily on terminal-based automation, Sonnet 5 is now the best model Anthropic offers for that specific use case.

On BrowseComp, which evaluates agentic search capability, Sonnet 5 reaches 84.7 percent. And on GDPval-AA v2, a knowledge work benchmark, Sonnet 5 scores 1,618 Elo, essentially matching Opus 4.8's 1,615.

The headline takeaway is that Sonnet 5 does not just improve over Sonnet 4.6. It enters Opus territory on several evaluations while maintaining Sonnet-level pricing.

Effort Levels: A New Way to Balance Cost and Performance

One of the most practical features in Sonnet 5 is the expanded effort level system. This gives you granular control over how much thinking budget the model allocates per response. The levels are low, medium, high, max, and extra high.

At low effort, Sonnet 5 is fast and cheap, suitable for simple tasks like formatting, classification, or quick edits. At medium effort, you get a solid balance for everyday coding and analysis work. At high and max effort, the model allocates significantly more reasoning tokens, producing results that rival Opus 4.8 on many tasks. Extra high effort pushes the model to its absolute ceiling.

What makes this powerful is that you can tune effort per request. A production pipeline might use low effort for routing and classification, medium effort for standard code generation, and max effort for the hardest debugging tasks, all with the same model. You do not need to switch between Sonnet and Opus anymore. You just adjust the dial.

Anthropic's own cost-performance charts show that Sonnet 5 provides a much wider range of cost-performance options than Sonnet 4.6 did. At medium effort, Sonnet 5 delivers substantially better cost efficiency. At higher effort settings, it matches Opus 4.8 on some tasks. Between Sonnet 5 and Opus 4.8, users can dial in whatever balance makes sense for their specific project.

The Context Window and Tokenizer Update

Sonnet 5 keeps the one-million-token context window, large enough to load a full codebase in a single prompt. That has not changed from previous models, but it remains a significant advantage for agentic workflows that need to reason over large amounts of context.

What has changed is the tokenizer. Sonnet 5 uses an updated tokenizer, the same one introduced with Opus 4.7. This improves how the model processes text internally, contributing to better performance. The trade-off is that the same input text can map to roughly 1.0 to 1.35 times more tokens depending on the content type.

This means your actual token counts may go up slightly even if you send the same prompts. Anthropic set the introductory pricing to account for this, making the transition roughly cost-neutral for existing Sonnet 4.6 users. But it is worth monitoring your token usage during the switchover, especially if you are running high-volume production workloads.

Cybersecurity Safeguards: What They Block and Why

Sonnet 5 is the first Sonnet-class model to ship with real-time cybersecurity safeguards enabled by default. These are the same protections present in Opus 4.7 and 4.8, designed to detect and block requests that may indicate prohibited or high-risk cybersecurity usage.

The context for this decision is important. Sonnet 5 shows somewhat stronger capabilities on certain cyber tasks than Sonnet 4.6, likely due to general intelligence improvements rather than specific training. While it still cannot develop a working exploit (it scored zero percent on full exploit development in Firefox vulnerability testing), it showed a slightly higher rate of partial success than its predecessor. That was enough for Anthropic to add guardrails.

For most developers, these safeguards will be invisible. Routine security work like dependency scanning, log analysis, vulnerability assessment, and defensive coding proceeds normally. The safeguards specifically target offensive cybersecurity tasks that could be weaponized.

If your legitimate work triggers a refusal, you will receive a standard HTTP 200 response with a stop_reason of "refusal" rather than an error. Anthropic offers a Cyber Verification Program, a free application-based program that allows security professionals working on legitimate defensive tasks to continue their work without interruption. Organizations already enrolled in the CVP automatically get the same access on Sonnet 5 with no need to reapply.

It is also worth noting that these safeguards are less strict than the ones Anthropic launched with Fable 5, which block a much wider range of cybersecurity tasks. Anthropic judged the overall cybersecurity risk from Sonnet 5 to be lower, so the restrictions are proportionally lighter.

Safety Profile: Fewer Hallucinations, Less Sycophancy

Beyond cybersecurity, Sonnet 5's broader safety profile is an improvement over Sonnet 4.6. Anthropic's pre-deployment evaluations found lower rates of hallucination and sycophancy, two issues that matter enormously in production applications.

On agentic safety specifically, Sonnet 5 is better at refusing malicious requests and resisting hijack attempts in prompt injection attacks. If you are building tools that expose Claude to untrusted input, this is a meaningful upgrade.

Anthropic's automated behavioral audit, which tests for a wide range of misaligned behaviors including cooperation with misuse and deception, scored Sonnet 5 lower (safer) overall compared to Sonnet 4.6. However, it did show somewhat higher rates of misaligned behavior compared to the more capable Opus 4.8 and Claude Mythos Preview. This is consistent with a general pattern where more capable models tend to have better safety profiles, likely because they have a deeper understanding of the contexts in which certain behaviors are inappropriate.

Pricing and Availability

Sonnet 5 is available everywhere today. It is the default model for Free and Pro plan users and is available to Max, Team, and Enterprise customers. It works in Claude.ai, Claude Code, Claude Cowork, and through the Claude Platform API using the model string claude-sonnet-5.

The introductory pricing through August 31, 2026 is two dollars per million input tokens and ten dollars per million output tokens. After that, standard pricing kicks in at three dollars per million input tokens and fifteen dollars per million output tokens. For comparison, Opus 4.8 runs at five dollars per million input tokens and twenty-five dollars per million output tokens.

At introductory pricing, Sonnet 5 is roughly 60 percent cheaper than Opus 4.8 on input and 60 percent cheaper on output. Even at standard pricing, it remains 40 percent cheaper on input and 40 percent cheaper on output. Given that it matches Opus 4.8 on several benchmarks, the cost-performance ratio is compelling.

Anthropic also raised rate limits across Chat, Cowork, Claude Code, and the Platform to accommodate the higher token usage that comes with higher effort levels. You can check your current tier and limits in the Claude Console.

When to Use Sonnet 5 vs Opus 4.8

The choice between Sonnet 5 and Opus 4.8 is no longer a clear-cut performance-versus-cost tradeoff. It is more nuanced now.

Choose Sonnet 5 when you need strong agentic performance at scale, your workloads involve coding, tool use, browser automation, or terminal-based tasks, cost efficiency matters, or you want to use effort levels to fine-tune the cost-performance balance per request. Sonnet 5 is also the better choice when you want to run many concurrent agents and need to keep per-agent costs manageable.

Stick with Opus 4.8 for the hardest reasoning tasks, scientific research requiring maximum accuracy, cybersecurity work that needs reduced guardrails (with CVP enrollment), or situations where the absolute ceiling of performance matters more than cost.

For many teams, the answer will be both: Sonnet 5 at medium effort for the bulk of work, with Opus 4.8 reserved for the tasks that genuinely need it. The effort level system makes Sonnet 5 flexible enough that you might not even need Opus except for the most demanding edge cases.

What Early Users Are Saying

Feedback from Anthropic's early access partners paints a consistent picture. Engineers at multiple companies described Sonnet 5 finishing complex multi-step tasks that previous Sonnet models would abandon partway through. One Rust engineer reported that when asked to investigate a bug, Sonnet 5 wrote a reproducing test, implemented the fix, then stashed the fix to confirm the bug returned without it, all in a single pass without being prompted to verify its work.

Teams working on legal research, insurance workflows, database analytics, and code review all reported similar themes: Sonnet 5 stays on plan, follows conventions, and ships clean multi-step changes. Multiple testers specifically highlighted the price-to-performance ratio as the factor that made migration from Sonnet 4.6 an easy decision.

Common Misconceptions to Avoid

Sonnet 5 is not Opus 4.8 at a discount. On the hardest reasoning and science benchmarks, Opus 4.8 still leads. Sonnet 5 closes the gap significantly, and surpasses Opus on some tasks, but it is a different model with a different performance profile.

The tokenizer change means your existing token budgets may need adjustment. The same prompts can consume up to 35 percent more tokens. Factor this into cost projections before switching production workloads.

Cybersecurity safeguards may surprise you if you are switching from Sonnet 4.6, which did not have them. If you do security-adjacent work, test your prompts against Sonnet 5 before cutting over in production.

Finally, effort levels are not free. Extra high effort uses substantially more tokens per request. Use the lowest effort level that gets the job done, and only escalate when quality demands it.

Conclusion

Claude Sonnet 5 is the most significant Sonnet release since the original 3.5, which kicked off the agentic AI era for many developers. By narrowing the gap to Opus-tier performance while maintaining Sonnet-level pricing and adding flexible effort controls, Anthropic has given developers a model that works for both experimentation and production scale.

The introductory pricing through August makes now the ideal time to test your workloads against Sonnet 5 and find the right effort level for each task. Whether you are building autonomous agents, writing code, or processing documents, this model is worth serious evaluation.

If you are a power user who wants to keep track of how your token usage shifts as you experiment with Sonnet 5's effort levels, tools like Gaugr can help you monitor consumption across models in real time.