Anthropic Launches Claude Sonnet 5: Agentic Performance, New Tokenizer, and Per-Task Cost Surprises

Anthropic launched Claude Sonnet 5 on June 30, 2026, positioning it as its most agentic Sonnet model to date, with benchmark scores approaching but not matching Opus 4.8 — 63.2% on SWE-bench Pro versus Opus 4.8's 69.2% and Sonnet 4.6's 58.1%.[3] The model ships at introductory pricing of $2/M input and $10/M output through August 31, 2026, rising to $3/$15 on September 1.[8] A new tokenizer shared with Opus 4.7 produces roughly 30% more tokens for the same English input as Sonnet 4.6,[7] and per-task benchmarks show Sonnet 5 costing approximately 2x Sonnet 4.6 and 15% more than Opus 4.8 to complete equivalent tasks.[8] Sonnet 5 is now the default model on Free and Pro Claude plans and available in Claude Code and the API.[2]

Sonnet 5 is the first model Anthropic is explicitly marketing for autonomous agentic workloads at sub-flagship pricing, but the tokenizer-driven per-task cost increase complicates the value proposition for API users who expected a straightforward cost-neutral upgrade from Sonnet 4.6.[7][8] How enterprises respond to the token-efficiency tradeoff will determine which tier of Anthropic's model lineup sees adoption in production agent pipelines.

Claude Sonnet 5 launched June 30, 2026, after several weeks of social media speculation — including a period where the model appeared in the claude.ai model picker before any official announcement[1] — and Anthropic positioned it as its most capable agentic Sonnet to date.[2] On agentic coding, Sonnet 5 scores 63.2% on SWE-bench Pro, above Sonnet 4.6's 58.1% but below Opus 4.8's 69.2%.[3] On agentic search benchmarks the improvement over Sonnet 4.6 is more pronounced, with commentary suggesting the older model is no longer competitive for that use case.[4] On GDPval, a knowledge-work benchmark, Sonnet 5 ties Opus 4.8 at approximately 1618 versus 1615.[5] The 145-page system card also records a lowest-in-class MASK lying rate of 3.1% — meaning Sonnet 5 is less likely than other tested models to lie under pressure — and notes the model occasionally declined to be helpful when requests conflicted with its stated welfare preferences.[6]

The single most discussed technical detail is a new tokenizer, shared with Opus 4.7, that produces roughly 30% more tokens for the same English text compared to Sonnet 4.6.[7] Anthropic set introductory pricing at $2/M input and $10/M output to be roughly cost-neutral relative to Sonnet 4.6 given this inflation,[2] but that neutrality holds only at the per-token level. Per-task benchmarks tell a different story: on the Intelligence Index, Sonnet 5 consumed roughly 2x the tokens Sonnet 4.6 used to complete equivalent tasks, producing a per-task bill of approximately $2.29 — about 15% more than Opus 4.8 and well above Sonnet 4.6.[8] Sonnet 5 Max, at standard post-intro pricing, is projected to cost more per task than Opus 4.8 Max across several benchmarks.[9] Adaptive thinking is enabled by default in Sonnet 5, which analysts identify as a main driver of higher token counts.[7] Critics labeled Anthropic's cost-neutral framing "tokenizer arbitrage," arguing the introductory discount was designed to obscure a structural price increase.[10]

Not all cost comparisons work against Sonnet 5. An informal test using three physics-coding demos showed Sonnet 5 completing the same tasks as GPT 5.5 using 15,047 tokens at $0.15, versus GPT 5.5's 31,152 tokens at $0.94 — roughly 6x cheaper for equivalent output quality.[11] Some analysts argue per-task cost comparisons are inherently workload-dependent: if Sonnet 5's higher per-call accuracy reduces retries on complex multi-step tasks, the effective cost on longer workflows may be lower than single-task benchmark snapshots suggest.[12]

On safety, Sonnet 5's most notable regression is on CyberGym, where it scores 52.7% versus Sonnet 4.6's 65.2%.[6] Anthropic states this reflects the absence of targeted cyber training rather than a deliberate capability floor — Sonnet 5's cyber performance derives from general reasoning ability.[13] The system card notes Sonnet 5 produced zero full browser exploits in Firefox testing, versus Mythos 5's 88.4%, and attributes this gap to deliberate capability controls that allowed the model to be released without triggering US government review process.[7]