Claude Sonnet 5 Review: Price, Benchmarks, and Verdict

This Claude Sonnet 5 review examines Anthropic's new general-purpose model without pretending to provide hands-on testing. Released on June 30, 2026, Sonnet 5 is now the default model for Claude Free and Pro users and is available through Claude Code, Anthropic's API, AWS, Google Cloud, and Microsoft Foundry.

The appeal is straightforward: stronger coding and agent performance, a one-million-token context window, and lower API pricing than Opus-class models. The less obvious details matter just as much. Most launch benchmarks were run by Anthropic, the tokenizer can turn the same text into more billable tokens, and several API behaviors changed. Here is what the current evidence supports.

Claude Sonnet 5 Review: Quick Verdict

Claude Sonnet 5 looks like a strong default for coding, research, document analysis, and agent workflows. It improved substantially over Sonnet 4.6 on Anthropic's evaluations, with introductory API pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026.

It is not simply a cheaper Opus 5. Anthropic positions Sonnet 5 as the balance of speed and intelligence, while Opus 5 is intended for the most complex agentic coding and enterprise work. The published evidence supports starting with Sonnet 5 and escalating difficult tasks only when an Opus model produces meaningfully better accepted outcomes.

The main caution is evaluation design. Results depend on the model, effort, harness, tools, token budget, and grading method. Validate Sonnet 5 on representative work before assuming a leaderboard gain will reduce cost or review time.

What Is Claude Sonnet 5?

Claude Sonnet 5 is Anthropic's latest Sonnet-tier model and a direct successor to Sonnet 4.6. It accepts text and image inputs and produces text output. The API model ID is claude-sonnet-5.

Anthropic describes Sonnet 5 as a capability upgrade for coding and agentic work. It can plan, call provided tools, inspect results, and continue through longer tasks. These capabilities depend on the product or agent harness; the API model receives no filesystem, browser, or business-system access by default.

Sonnet 5 is proprietary; Anthropic has not disclosed its parameter count or released model weights.

Claude Sonnet 5 Features and API Changes

One-Million-Token Context and 128K Output

Sonnet 5 supports a one-million-token context window by default and a maximum output of 128,000 tokens. The large window is useful for broad codebase context, long document collections, extended conversations, and agent traces.

Long-context work still needs retrieval, clear source boundaries, compaction, and checks that the model used the right evidence. Cost and latency also rise when an application repeatedly sends a large context.

Adaptive Thinking and Effort Controls

Adaptive thinking is enabled by default. Developers can use the effort parameter to balance compute, latency, and performance. Manual extended-thinking budgets are no longer supported.

Sonnet 5 also rejects non-default temperature, top_p, and top_k values. Migrating teams should remove those settings.

A New Tokenizer Changes Effective Cost

Anthropic says the new tokenizer produces about 30% more tokens for the same text on average, with the exact increase depending on content. This changes token counts, context capacity in practical text terms, output limits, and per-request cost.

Anthropic says the introductory discount makes the initial transition roughly cost-neutral. Recount representative prompts and measure total tokens and accepted output rather than comparing list prices alone.

Claude Sonnet 5 Benchmarks: What the Results Show

Anthropic's system card reports strong gains on coding and agent tasks. Sonnet 5 scored 85.2% on SWE-bench Verified, 63.2% on the harder SWE-bench Pro, and 80.4% on Terminal-Bench 2.1 at extra-high effort. The same Terminal-Bench setup gave Sonnet 4.6 a 67% score.

Cursor independently reported 61.2% for Sonnet 5 on CursorBench, compared with 49% for Sonnet 4.6 and 63.8% for Opus 4.8. That result supports the idea that Sonnet 5 narrows the gap with an Opus model in a production coding harness, although it does not prove parity on every repository.

The model also scored 84.7% as a single agent on BrowseComp, which tests difficult web research, and 81.2% on OSWorld-Verified, which tests computer tasks in a live Ubuntu environment. Anthropic changed parts of its OSWorld setup and reran Sonnet 4.6, so older published comparisons should not be mixed with the updated numbers.

These results have important qualifications:

• Most scores were produced or compiled by Anthropic.
• High effort can use more tokens and take longer than default settings.
• Agent benchmarks include tools, scaffolding, retries, and graders.
• A pass does not measure maintainability, security, or review time.

Artificial Analysis offers a useful independent signal. It places max-effort Sonnet 5 among leading models and measures above-average output speed, but also reports unusually high output-token use. Compare cost per accepted task rather than token price or benchmark score alone.

Claude Sonnet 5 Pricing and Real Workload Cost

Anthropic's introductory API pricing runs through August 31, 2026:

• Input: $2 per million tokens
• Output: $10 per million tokens
• Five-minute prompt-cache write: $2.50 per million tokens
• Prompt-cache read: $0.20 per million tokens

Standard pricing from September 1 is scheduled to be $3 per million input tokens and $15 per million output tokens. Opus 5 currently costs $5 per million input tokens and $25 per million output tokens.

The cheapest model per token is not always cheapest per completed task. Verbose output, retries, and human correction can erase a list-price advantage.

For a fair pilot, track total tokens, elapsed time, tool calls, retries, reviewer minutes, and accepted-task rate. Recheck after the promotional price ends.

Claude Sonnet 5 vs Opus 5 and Sonnet 4.6

Choose Sonnet 5 over Sonnet 4.6 when a workflow benefits from stronger coding, tool use, research, or computer interaction and can accommodate the migration changes. The list price after the introductory period matches Sonnet 4.6, but the tokenizer means equivalent text may not have equivalent cost.

Choose Opus 5 when the task is unusually difficult, failure is costly, or internal tests show a consistent quality advantage that justifies the higher price. SD's Claude Opus 5 review covers that model's position in more detail.

For coding teams, model selection should happen inside the agent workflow. Sonnet 5 may be the economical default for issue investigation, test writing, routine feature work, and review preparation. Opus can be reserved for architecture, stubborn debugging, large migrations, or tasks where Sonnet repeatedly stalls. The same workflow-first principle appears in SD's Claude Code vs Codex comparison.

Who Should Use Claude Sonnet 5?

Sonnet 5 suits high-volume agents, Claude Code users upgrading from Sonnet 4.6, and teams analyzing large code or document collections. It may also fit research and operations workflows that need browsing, tools, and auditable drafts.

It is less compelling when a smaller model meets the quality threshold, prompts depend on custom sampling parameters, or higher token counts break the budget. Consequential tasks still need qualified review and tightly scoped tools.

Safety and Limitations

Anthropic reports that Sonnet 5 improved over Sonnet 4.6 on prompt-injection robustness, hallucination, sycophancy, and several agentic safety evaluations. In a one-week Gray Swan bug bounty, 0.19% of unique attacks succeeded against Sonnet 5, tied with Opus 4.8 and lower than the rate reported for Sonnet 4.6.

That is encouraging, not a guarantee. The system card also notes over-refusal in some benign cases, occasional hallucinations, and sensitivity to suspected prompt injection. Teams still need least privilege, isolated execution, approval gates, source checking, and audit logs. SD's AI agent security guide explains those controls.

How to Evaluate Claude Sonnet 5

Build a test set from resolved bugs, code reviews, research questions, documents, and safe tool-use scenarios. Include cases where the model should decline or acknowledge missing evidence.

Run Sonnet 4.6, Sonnet 5, and the relevant Opus model with the same instructions, tools, data, and success criteria. Evaluate correctness, task completion, regression rate, unsupported claims, security behavior, latency, token use, reviewer effort, and total cost. Test more than one effort level because Sonnet 5's value proposition depends partly on selecting the right compute for each task.

Move in stages: read-only analysis, draft changes, sandboxed execution, and then narrowly approved actions. Only representative results should justify broader autonomy.

Conclusion

This Claude Sonnet 5 review finds a credible upgrade for coding, research, computer use, and high-volume agent workflows. The strongest evidence is the consistent improvement over Sonnet 4.6 across Anthropic's evaluations, supported by independent Cursor and Artificial Analysis signals.

The practical caveats are equally important: most launch benchmarks are vendor-run, high effort can be expensive, the tokenizer increases token counts for equivalent text, and API migration requires more than changing the model ID. Start with Sonnet 5 as an economical default, reserve Opus for tasks where it earns its premium, and decide from accepted outcomes on representative work.