Claude Sonnet 5: Performance Close to Opus 4.8 but at a Lower Price
On June 30, Anthropic introduced Claude Sonnet 5, the mid-range model that quickly becomes the default for the Free and Pro plans of Claude and is also available for Max, Team, and Enterprise subscriptions. The central point of the announcement is the cost-performance relationship: according to Anthropic, the performance of Sonnet 5 is close to that of Opus 4.8, the flagship model, but at significantly lower prices, with a clear improvement over its predecessor Sonnet 4.6 in reasoning, tool usage, coding, and knowledge work.
The model is already usable in Claude Code, on the Claude apps, and via the API with the identifier claude-sonnet-5. On the pricing front, Anthropic offers an introductory rate of $2 per million input tokens and $10 per million output tokens, valid until August 31, 2026; starting September 1, the rates will switch to standard prices of $3 and $15 per million. This is a less straightforward detail than it seems, as Sonnet 5 adopts an updated tokenizer: the same text can translate into up to 35% more tokens compared to the previous model, depending on the type of content, and the introductory price is calibrated to make the transition roughly cost-neutral.
The Most Agentic Sonnet Yet
Anthropic positions Sonnet 5 as the most agentic Sonnet model ever created: capable of crafting plans, using tools like browsers and terminals, and operating independently at a level that, until a few months ago, required larger and more expensive models. Testers who had early access described it as significantly more autonomous than its predecessors, capable of completing complex tasks where earlier versions would stop, and monitoring its own output without explicit prompts.
On the BrowseComp and OSWorld-Verified benchmarks, at the highest effort levels, Sonnet 5 can match Opus 4.8 on some tasks, while at intermediate effort, it offers significantly better cost efficiency. Developers can adjust this effort level to balance cost and performance between Sonnet 5 and Opus 4.8. To accommodate higher token consumption at upper levels, Anthropic has increased the rate limits on Chat, Cowork, Claude Code, and Claude Platform.
Security and Limits on Cybersecurity
Compared to Sonnet 4.6, the new model shows lower rates of undesirable behaviors: fewer hallucinations, less sycophancy, and better resistance to prompt injection attacks. However, on a specific front, Anthropic notes a deliberate setback: Sonnet 5 has not been trained for cybersecurity tasks and performs significantly worse than Opus 4.8 in developing exploits. In a test conducted in collaboration with Mozilla on vulnerabilities in Firefox 147, already patched in Firefox 148, the model never produced a working exploit. The model has still been released with the active cyber safeguards by default, as stated in the official announcement from Anthropic, which on the same day also corrected a graph on BrowseComp that was based on a simplified methodology that underestimated the model's performance.
The Cyber Verification Program for Sonnet 5 is available on Claude Platform, AWS, and Microsoft Foundry, with its arrival on Google Vertex announced as imminent. For those considering the switch, the practical calculation remains tied to the introductory window: until August 31, the price gap with the standard September rates is the most concrete leverage to test the model on their workloads.