Anthropic shipped Claude Sonnet 5 on June 30, setting introductory pricing at $2 per million input tokens and $10 per million output tokens through August 31, 2026. That is less than half of what Opus-class models typically command, and it is the clearest signal yet that Anthropic wants its mid-tier model, not its flagship, to carry the bulk of agentic workloads.
The pitch is narrowing a gap rather than opening a new frontier. Anthropic says Sonnet 5’s agentic performance, spanning planning, tool use, coding, and knowledge work, now approaches Opus 4.8 on several evaluations, while improving substantially over its predecessor, Sonnet 4.6. The company’s own BrowseComp and OSWorld-Verified charts show Sonnet 5 covering a wider range of cost-performance tradeoffs than Opus 4.8, with its higher-effort setting matching Opus on some tasks.
Those benchmarks come from Anthropic’s own system card, not an independent lab. The company does note one hard limit voluntarily: Sonnet 5 has a much lower capability ceiling for cybersecurity tasks than current Opus models, and it launches with the same cyber safeguards used on Opus 4.7 and 4.8 rather than the stricter set applied to Fable 5.
Anthropic’s safety testing found Sonnet 5 shows an overall lower rate of undesirable behaviors than Sonnet 4.6, including fewer instances of hallucination and sycophancy, and better resistance to prompt injection attacks in agentic settings. On the company’s automated behavioral audit, which screens for misuse and deception, Sonnet 5 scored safer than Sonnet 4.6 but somewhat less safe than Opus 4.8 and the Claude Mythos Preview model. Neither Sonnet model could complete a full working exploit in Anthropic’s Firefox vulnerability test, though Sonnet 5 showed a slightly higher partial success rate than its predecessor, a change Anthropic attributes to general intelligence gains rather than targeted training.
The pricing structure carries a caveat worth flagging for anyone budgeting token spend. Sonnet 5 uses an updated tokenizer, the same kind of change Anthropic introduced with Opus 4.7, which can push the same input to 1.0 to 1.35 times more tokens depending on content type. Anthropic says the introductory price is set to make the transition roughly cost-neutral, but that assumes typical usage patterns; workloads heavy in the content types that tokenize less efficiently will see costs drift upward even at the discounted rate.
Sonnet 5 is now the default model for Free and Pro plans, and it’s available to Max, Team, and Enterprise users, in Claude Code, and via the Claude API as claude-sonnet-5. Standard pricing after August 31 rises to $3 per million input tokens and $15 per million output tokens, still well below typical Opus rates.
The customer quotes Anthropic published, from companies including Lovable and ClickHouse, describe agents finishing multi-step jobs (updating account tiers, sending outreach, tracing bugs to root cause) that previously stalled partway through. Those are vendor-selected testimonials in a launch post, not third-party benchmarking, and Anthropic has not disclosed adoption or retention data for how those same partners used Sonnet 4.6.
For teams currently running agentic workloads on Opus because Sonnet 4.6 could not finish complex jobs, the six-week window before standard pricing kicks in on September 1 is the moment to benchmark Sonnet 5 against your own task set and lock in the lower effort tier that meets your accuracy bar before the discount expires.
Reported by Anthropic on June 30, 2026.
