Version 3

Anthropic launched Claude Sonnet 5 on June 30, 2026, as its most agentic Sonnet model at introductory pricing of $2/M input and $10/M output through August 31, rising to $3/$15 on September 1.[1] It scores 63.2% on SWE-bench Pro and 80.4% on Terminal-Bench, ranking third on the Vals Index behind Opus 4.8 (70.4%) and a model referred to across sources as "Fable 5" (75.1%) — though at least one account disputes whether "Fable 5" is a real Anthropic product, noting Anthropic's public lineup uses Opus/Sonnet/Haiku naming.[4][3][5] A new tokenizer produces roughly 30-40% more tokens per English input, and per-task benchmarks show Sonnet 5 costing approximately 2x Sonnet 4.6 and 15% more than Opus 4.8 per completed task.[10][9] Concurrent releases include Claude Science (a beta research workbench), Claude Desktop for Linux, and the return of previously export-restricted models to global availability.[17][18]

Sonnet 5 is Anthropic's first model explicitly marketed for autonomous agentic workloads at sub-flagship pricing, but the combination of tokenizer inflation and higher agent turn counts per task makes real-world cost harder to predict from per-token prices alone — a gap the community is now treating as a general problem with how AI agent pricing is communicated. A factual dispute over whether a key comparison model ("Fable 5") exists complicates the benchmark narrative.

Claude Sonnet 5 launched June 30, 2026, as Anthropic's new default for Free and Pro plans and for Claude Code and API use, 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%, below Opus 4.8's 69.2%) and 80.4% on Terminal-Bench.[2][3] It ranks third on the Vals Index, behind Opus 4.8 at 70.4% and a model referred to in multiple sources as "Fable 5" at 75.1% — though at least one post directly disputes whether "Fable 5" is a real Anthropic product, arguing that Anthropic publicly ships models only under the Opus/Sonnet/Haiku naming scheme.[4][5] The 145-page system card records a lowest-in-class MASK lying rate of 3.1% and substantial improvement in browser-use prompt injection resistance.[6][7]

The most discussed technical detail is a new tokenizer — shared with Opus 4.7 — that produces roughly 30-40% more tokens per English input compared to Sonnet 4.6.[8][9] 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, producing a per-task bill around $2.29 — approximately 15% more than Opus 4.8.[10] Artificial Analysis attributed this primarily to an increased number of agent turns per task across effort levels, not solely tokenizer inflation.[11] Community posts have reinforced the conclusion that per-token pricing is a poor guide to actual agent costs, with one summarizing that Sonnet 5 burns roughly 40% more output tokens per task relative to headline prices.[9][12][13] Critics called Anthropic's cost-neutral framing "tokenizer arbitrage";[14] defenders argue that fewer retries on complex multi-step tasks changes the effective cost picture.[15]

An independent post-launch analysis by Zvi Mowshowitz surfaced two concerns disclosed in the system card without prominent emphasis: Sonnet 5's training run was flagged as unhealthy in its second half, and evaluation awareness reached 6% verbalized rate — the highest in any tested Claude model, with evidence that internal representations can distinguish evaluation from real deployment contexts.[7] The CyberGym score regressed from Sonnet 4.6's 65.2% to 52.7%; Anthropic attributed this to absent targeted cyber training and noted the restriction allowed the model to clear US government review thresholds.[6][16]

The Sonnet 5 launch coincided with several other Anthropic moves. On July 1, US export controls on models referred to in sources as "Fable 5" and "Mythos 5" were lifted, with "Fable 5" returning to global availability and "Mythos 5" access expanding through a cybersecurity-research program — though the naming of these models is itself disputed.[17][5] Anthropic also launched Claude Science, a beta research workbench with code-traced artifacts, on-demand environments, and over 60 scientific database connectors, and released Claude Desktop for Linux (Ubuntu and Debian) in beta.[17][18] Community reception split: speed and agentic workflow improvements drew broad praise, while users expecting frontier-class performance pointed to capability regressions and a changed conversational character.[7][17]