Claude Sonnet 5: Anthropic has released one of the most powerful AI models for work and programming

On June 30, 2026, Anthropic released Claude Sonnet 5—a new mid-tier model that has narrowed the performance gap with the flagship Opus while maintaining a lower cost. Developers describe it as the most "agentic" model in the Sonnet lineup: it excels at programming, planning, and executing complex tasks using external tools. Let’s examine what has changed and why this release could be one of the most significant for the AI market.

What it is and why it matters

Anthropic positions Sonnet 5 as the most agentic model in its lineup to date. It can independently create plans, utilize tools such as a browser and terminal, and operate autonomously at a level that, until recently, required larger and more expensive models.

Previously, Sonnet models versions 3.5, 3.6, and 3.7 were the first to demonstrate serious skills in coding and tool usage. However, in recent generations, the primary progress in agentic capabilities went into Opus. Sonnet 5 closes this gap: its performance is close to Opus 4.8, but it costs less.

Who it is for

The model is suitable for developers and teams using Claude Code. The agent now completes multi-step pull requests to a verified result without constant supervision.

It is useful for product teams building AI agents to automate CRM, ticket processing, and workflows. A real-world case is cited where the model updated account tiers in Salesforce and sent announcements to clients entirely autonomously, from start to finish.

For companies with high query volumes, Sonnet 5 is interesting as a way to achieve a good price-to-performance ratio where absolute maximum capability is not required. For such tasks, Opus is available.

For enterprise, it is now the default standard model. It is available on Free and Pro plans, as well as in Max, Team, Enterprise, Claude Code, and via API.

Technical innovations

New effort levels have been introduced: low, medium, high, max, and x-high. These allow users to adjust the depth of reasoning according to the specific task and budget.

At the maximum level, the model approaches Opus 4.8 on certain tasks. However, this may cost more than Opus at a comparable accuracy level, so effort has become a full-fledged lever for cost management, not just a switch for reasoning mode.

The tokenizer has changed, similar to Opus 4.7. This improves text processing quality, but the same input now converts into more tokens, approximately 1.0 to 1.35 times more depending on the content type. The input price has been set so that the transition from Sonnet 4.6 is roughly cost-neutral; this should be considered when recalculating budgets and limits.

The context window consists of one million tokens according to various sources, with increased request limits at higher effort levels.

The model works significantly better with terminals and legacy code, including race conditions and hidden tests. According to partners like Cursor, Warp, and Lovable, it automatically writes reproducing tests, fixes bugs, and reverts fixes to ensure the bug has truly disappeared, without an explicit request to do so.

Pricing

The introductory price, valid until August 31, 2026, is $2 per million input tokens and $10 per million output tokens. After this date, the price will increase to $3 and $15, respectively.

For comparison, Opus 4.8 costs $5 for input and $25 for output. The model is available via API under the name claude-sonnet-5.

Benchmarks

Anthropic’s official post compares Sonnet 5 with Sonnet 4.6 and Opus 4.8 using a graph, without exact figures in the text. Independent reviews, relying on the model card, generally agree with this picture.

On SWE-bench Pro for agentic coding, Sonnet 4.6 shows about 58%, Sonnet 5 about 63%, and Opus 4.8 about 69%. The increase is noticeable, but Opus is still ahead.

On Humanity's Last Exam using tools, Sonnet 4.6 scores about 46.8% after grader recalculation, Sonnet 5 about 57.4%, and Opus 4.8 about 57.9%. Here, Sonnet 5 almost catches up to Opus.

On OSWorld-Verified, a computer usage test, Sonnet 4.6 scores about 78.5% after methodology recalculation, while Sonnet 5 shows a significant increase beyond that. On one of the knowledge-based benchmarks, Sonnet 5 even slightly surpasses Opus 4.8.

A caveat is necessary. Some secondary sources, particularly certain SEO blogs, cite contradictory figures, with some reporting 92 percent on SWE-bench Verified and 96% on GPQA, which conflicts with all other materials and appears to be unverified clickbait. The most reliable source for exact figures is the model card on Anthropic’s website.

It is also important that Anthropic retrospectively recalculated Sonnet 4.6 results on HLE due to a change in the grader and on OSWorld-Verified due to a change in methodology. Old posts about version 4.6 and new figures are not always directly comparable.

Safety

The model has a lower frequency of undesirable behaviors, including hallucinations, sycophancy, and assistance with abuse, compared to Sonnet 4.6. It also shows better resistance to prompt injection attacks in agentic scenarios.

Regarding cybersecurity, the model was not specifically trained on these topics and shows significantly weaker results in exploit development than Opus 4.8 and Mythos 5. In a Mozilla test to create exploits for Firefox 147, neither Sonnet 5 nor Sonnet 4.6 could create a working exploit, though Sonnet 5 had a slightly higher rate of partial success.

Due to this, the model was launched with real-time cybersecurity threat protections enabled by default. These are the same protections as those in Opus 4.7 and 4.8, although less strict than those in Fable 5.

How to choose between models

Haiku 4.5 remains the choice for simple, mass-market, and inexpensive tasks. Sonnet 5 has become the new default for most production scenarios involving agentic work, including coding, automation, and tool usage, where a balance of price and quality is important.

Opus 4.8 remains the best choice where maximum accuracy on complex tasks is needed or where cybersecurity requirements are lower. Fable 5 and Mythos 5 occupy the top of the lineup, with more limited availability and additional restrictions on biology, cybersecurity, and LLM development topics for Fable 5.