Claude Sonnet 5 lands just below the compute threshold the Trump administration uses to flag "frontier" AI models for safety review and export controls, posting near flagship scores on autonomous, multi step coding tasks, at lower cost.
Anthropic's newest AI model was built to do something specific: stay just under the line Washington draws for its most powerful systems. Claude Sonnet 5, released this week, posts near-flagship performance on coding tasks while sitting below the compute threshold the Trump administration uses to flag "frontier" models for mandatory safety review and export controls. The design choice is now openly part of Anthropic's launch positioning.
Sonnet 5 is the mid-tier entry in Anthropic's Claude family, between the smaller Haiku and the larger Opus line. Anthropic describes it as "the most agentic Sonnet model yet", meaning it can plan, use tools like browsers and terminals, and run multi-step tasks with less hand-holding than its predecessors. That capability used to live almost exclusively in frontier-tier models, including Anthropic's own flagship Opus 4.8.
On SWE-Bench Pro, a benchmark that asks models to resolve real GitHub issues end to end, GovInfoSecurity reports Sonnet 5 scored 63.2%, within roughly six percentage points of the 69.2% the publication attributes to Opus 4.8. On Terminal-Bench 2.1, which measures terminal-based agentic work, Sonnet 5 hit 80.4%. Both numbers come from GovInfoSecurity's analysis and should be treated as such until Anthropic's system card confirms them independently.
That gap matters because the regulatory threshold does. The administration's executive order on AI defines a "frontier" model largely by training compute, and models above that line face voluntary safety commitments, federal review, and potential export restrictions. Sonnet 5 lands days after the U.S. government partially lifted an export ban on a frontier-class model that crossed the threshold, according to GovInfoSecurity's reporting. That timing is the regulatory hook for the "without frontier model scrutiny" framing in the publication's analysis.
Anthropic is not framing this as arbitrage. Its launch materials position Sonnet 5 against Opus 4.8 in terms of safety, cost, and reduced regulatory exposure. But the gap between capability and threshold is shrinking. Sonnet 4.6, the previous mid-tier release, scored noticeably lower on agentic benchmarks, per Anthropic's own characterization. The trajectory of the mid-tier line is upward; the threshold itself is fixed. At some point, "small enough to avoid scrutiny" and "capable enough to matter" become harder to keep separate.
Distribution is immediate. AWS has confirmed Sonnet 5 is available on its Bedrock platform, the dominant enterprise channel for Anthropic's models. That means the regulatory positioning has customer consequences from day one: enterprise buyers can adopt near-flagship agentic capability without the procurement friction that attaches to models above the frontier threshold.
Sonnet 5 is therefore a policy story as much as a product story. The threshold line drawn in an executive order now shapes which models get built and how they are positioned, not just how they are reviewed after the fact. Two things will tell whether the trade-off is durable: whether Anthropic's full system card confirms the cited benchmarks independently, and whether the next mid-tier release stays under the threshold or crosses it. A pattern of release-under-threshold, capability-near-flagship would confirm the policy line is a permanent road map input. A future jump above the threshold would suggest the strategy only works as long as the gap holds.