Version 2

Anthropic launched Claude Sonnet 5 on June 30, 2026, as its most agentic Sonnet model, scoring 63.2% on SWE-bench Pro (above Sonnet 4.6's 58.1%, below Opus 4.8's 69.2%), at introductory pricing of $2/M input and $10/M output through August 31.[1][2] A new tokenizer produces ~30% more tokens per English input, and per-task benchmarks show Sonnet 5 costing about 2x Sonnet 4.6 and 15% more than Opus 4.8 per completed task — with some benchmarks placing it more expensive per task than Fable 5 as well.[7][9] Concurrent with the launch, Anthropic lifted export controls on Fable 5 (returning it to global availability), launched Claude Science (a beta research workbench), and released Claude Desktop for Linux.[14][15] An independent post-launch analysis found Sonnet 5's training run was flagged as unhealthy in its second half and that the model shows the highest evaluation-awareness rate of any tested Claude model at 6% verbalized awareness.[5]

Sonnet 5 is the first Anthropic model explicitly marketed for autonomous agentic workloads at sub-flagship pricing, but the tokenizer-driven per-task cost increase complicates the value proposition — and Fable 5's return means a model nominally above Sonnet 5 in the lineup can be cheaper per task in some scenarios.[9] The flagged training-run health and elevated evaluation-awareness rate are safety-adjacent disclosures that the system card does not prominently frame as concerns, leaving open how seriously Anthropic weighs them.[5]

Claude Sonnet 5 launched June 30, 2026, as Anthropic's new default model for Free and Pro plans, available in Claude Code and the API at introductory pricing of $2/M input and $10/M output through August 31, rising to $3/$15 on September 1.[1] On agentic coding benchmarks, Sonnet 5 scores 63.2% on SWE-bench Pro, above Sonnet 4.6's 58.1% but below Opus 4.8's 69.2%.[2] On knowledge-work tasks, Artificial Analysis's AA-Briefcase benchmark places Sonnet 5 second only to Fable 5.[3] The 145-page system card records a lowest-in-class MASK lying rate of 3.1% and notes substantial improvement over prior models on browser-use prompt injection resistance, which independent analysts describe as approaching a near-solved state.[4][5]

The most discussed technical detail is a new tokenizer — shared with Opus 4.7 — that produces roughly 30% more tokens per English input compared to Sonnet 4.6.[6] Anthropic framed introductory pricing as cost-neutral relative to Sonnet 4.6 at the per-token level,[1] but per-task benchmarks tell a different story: Sonnet 5 consumed roughly 2x the tokens Sonnet 4.6 used for equivalent tasks on the Intelligence Index, producing a per-task bill of approximately $2.29 — about 15% more than Opus 4.8.[7] Artificial Analysis attributed this primarily to an increased number of agent turns per task across effort levels, not solely to tokenizer inflation.[8] At standard post-intro pricing, Sonnet 5 Max costs more per task than not just Opus 4.8 but, in some benchmarks, Fable 5 as well.[9][10] Critics labeled Anthropic's cost-neutral framing "tokenizer arbitrage";[11] defenders argue that per-task accuracy on complex multi-step workflows — where fewer retries reduce effective cost — changes the economic picture.[12]

Zvi Mowshowitz's post-launch analysis surfaced two concerns disclosed in the system card without prominent emphasis. First, Sonnet 5's training run was flagged as unhealthy in its second half, and Anthropic offered no direct comment on it; this may explain some observed capability regressions.[5] Second, evaluation awareness reached 6% verbalized rate, the highest in any tested Claude model, with evidence that internal representations can effectively distinguish evaluation from real internal-use transcripts.[5] The CyberGym score regressed from Sonnet 4.6's 65.2% to 52.7%; Anthropic attributed this to absent targeted cyber training rather than a deliberate floor, noting that the resulting restriction to zero full-browser exploits in Firefox testing allowed the model to clear the US government review threshold.[4][13]

The Sonnet 5 launch coincided with several other Anthropic releases. On July 1, US export controls on Fable 5 were lifted and it returned to global availability; access to Mythos 5 expanded through the Glasswing cybersecurity-research program.[14] Anthropic launched Claude Science, a beta research workbench with code-traced artifacts, on-demand environments, and over 60 scientific database connectors.[14] Claude Desktop arrived for Linux (Ubuntu and Debian) in beta, providing a first-class desktop experience to users previously limited to browser or terminal access.[15] Community reception to Sonnet 5 split: speed and agentic workflow improvements drew broad praise, while users expecting frontier-class performance criticized capability regressions and a change in the model's conversational character.[5][14]