Anthropic Launches Claude Sonnet 5, Betting Cheaper AI Can Still Win the Agentic Race

The company is calling Sonnet 5 its “most agentic Sonnet model yet,” capable of building plans, operating browsers and terminals, and running autonomously at a level that, until recently, required far larger and more expensive models. Anthropic says the new release substantially closes the performance gap with Opus 4.8 while undercutting it sharply on price.

Benchmarks show a narrowing gap

On SWE-bench Pro, a widely cited coding benchmark, Sonnet 5 scored 63.2%, well ahead of predecessor Sonnet 4.6’s 58.1% and edging toward Opus 4.8’s 69.2%. On Terminal-Bench 2.1, the model jumped to 80.4% from 67.0%, and tool-assisted reasoning climbed to 57.4% from 46.8%. In a knowledge-work benchmark, Sonnet 5 actually edged out Opus 4.8, scoring 1618 against the larger model’s 1615.

“Opus 4.8 is still the model of choice for higher accuracy on these tasks, but Sonnet 5 provides developers with lower-priced options that are of much higher quality than what was previously available,” Anthropic said in its announcement. The company added that users can now toggle effort levels between Sonnet 5 and Opus 4.8 to balance cost against performance.

Sonnet 5 is close to Opus 4.8 levels of intelligence, Source: Anthropic

Pricing undercuts the field

Sonnet 5 launches with introductory API pricing of $2 per million input tokens and $10 per million output tokens, running through August 31, 2026, after which rates rise to $3 and $15 respectively. That remains well below Opus 4.8’s $5/$25 pricing, and Anthropic says it also beats OpenAI’s GPT-5.5 and Google’s Gemini 3.1 Pro, though Google’s Gemini 3.5 Flash still comes in cheaper.

The model is live immediately across Free, Pro, Max, Team, and Enterprise plans, where it now serves as the default for Free and Pro users. It’s also available in Claude Code and via the Claude API as claude-sonnet-5.

Early testers point to follow-through

Several launch partners highlighted the model’s tendency to finish multi-step jobs rather than stall partway through. Daniel Shepard, a senior engineer at Zapier, said the company handed Sonnet 5 a two-part task — updating Salesforce account tiers and sending a launch announcement to enterprise contacts — “and it finished end to end. That used to stall halfway. For day-to-day automation, it’s a no-brainer.”

Cursor co-founder Sualeh Asif said the model keeps agents “on plan,” following coding conventions through clean multi-step changes “at an efficient cost.” Lovable co-founder Fabian Hedin praised its handling of unsafe requests, saying it “refuses unsafe requests cleanly and consistently.”

Safety posture improves, but gaps with Opus remain

Anthropic says Sonnet 5 shows a lower overall rate of undesirable behaviors than Sonnet 4.6, including reduced rates of hallucination and sycophancy, and is better at resisting prompt-injection hijack attempts. The company has enabled real-time cyber safeguards by default, similar to those used on Opus 4.7 and 4.8, though less restrictive than the controls applied to Fable 5.

On a Firefox exploit-development evaluation run with Mozilla, neither Sonnet 5 nor Sonnet 4.6 produced a working exploit, though Sonnet 5 posted a higher partial-success rate (13.2% versus 8.8%). Both remain far behind Opus 4.8’s 68.8% and Mythos 5’s 88.4% on the same test. Anthropic notes Sonnet 5 still shows “somewhat higher rates of misaligned behavior” than Opus 4.8 and the company’s restricted Mythos Preview model.

Context: an IPO push and a reshuffled model lineup

The launch lands against the backdrop of Anthropic’s confidential SEC filing on June 1 for an IPO reportedly targeting a valuation near $1 trillion. It also follows a turbulent stretch for the company’s top-tier lineup: Fable 5 and Mythos 5, Anthropic’s first Mythos-class models, launched June 9 before a Commerce Department export-control directive forced Anthropic to suspend access worldwide on June 12, citing concerns over a possible bypass of Fable 5’s cybersecurity safeguards.

Anthropic disputed the severity of the issue but complied, taking both models offline for every user globally since it couldn’t filter access by nationality in real time. On June 26, Commerce Secretary Howard Lutnick partially eased the order, allowing the unrestricted Mythos 5 to be used by a narrow set of vetted US critical-infrastructure organizations and government agencies; Fable 5, the safeguarded version built for general and commercial use, remains suspended for subscribers, API developers, and all international customers, with no restoration date announced. OpenAI has faced a parallel intervention: the White House asked the company to limit the initial rollout of its GPT-5.6 lineup — Sol, Terra, and Luna — to a small group of government-approved partners after officials judged Sol’s capabilities “Mythos-like,” with OpenAI saying it expects broader availability “in the coming weeks” as both companies work with Washington on a longer-term review framework for frontier models. Sonnet 5’s launch effectively continues business as usual for Anthropic’s mid-tier lineup while the company’s most powerful models remain caught in that unresolved regulatory standoff.