Claude Sonnet 5 matches Opus 4.8 on key knowledge tasks

Anthropic's pricing model just became harder to justify. Claude Sonnet 5, released this week, scores 1,618 on the GDPval-AA v2 knowledge work benchmark, edging past the larger Opus 4.8 on that specific test. The Decoder reports the new mid-tier model beats its predecessor, Sonnet 4.6, across all benchmarks. When a cheaper model outperforms the premium one on the tasks enterprises actually care about, the premium tier needs a new argument.

That argument is getting thinner. Opus was always sold on capability headroom: the model you reach for when the work is complex, the stakes are high, and cost is secondary. Knowledge-work benchmarks are precisely the domain that made Opus worth the price difference. Sonnet 5 closing that gap is not a minor incremental update; it is a structural shift in how enterprises should think about model selection.

What the benchmark actually measures

GDPval-AA v2 is designed to evaluate performance on the kind of analytical, multi-step reasoning tasks that senior professionals in finance, policy, and research actually perform. A score of 1,618 on that test is meaningful because it measures output quality on work that resembles real procurement decisions, policy briefs, and financial analysis rather than code completion or trivia retrieval. For a multilateral institution or a financial services firm deploying AI at scale, the relevant question is never raw benchmark position in the abstract; it is which model produces defensible, accurate outputs on the specific task type at the lowest cost.

Sonnet 5 now answers that question differently than it did six months ago.

The cybersecurity signal is deliberate

Anthropic's decision to publish Sonnet 5's low score on cybersecurity tasks alongside its knowledge-work performance is not accidental transparency. The US government has blocked certain frontier models on national security grounds, and Anthropic is positioning Sonnet 5 explicitly below that threshold. Publishing the low score is a regulatory strategy dressed as a benchmark disclosure. For enterprise and institutional buyers operating under procurement rules or government-adjacent compliance frameworks, this matters: the model is designed to be approvable, not just capable.

For UN system agencies, development finance institutions, and major industrial groups with government contracts, the compliance positioning may matter as much as the benchmark score. A model that scores well on knowledge work and explicitly sits below the flagged cybersecurity capability tier is, by design, easier to get through a procurement committee.

The consequence for LLM-driven brand visibility

None of this is abstract for brands whose authority depends on how AI answers questions about their domain. The models organisations deploy to generate, summarise, and retrieve information are now cheaper and more capable than they were. That increases the volume of AI-generated answers, which increases the importance of being cited in the training and retrieval pipelines those models draw on.

When Anthropic compresses the capability gap between its tiers, enterprises upgrade their mid-tier deployments. More sophisticated models run more queries. The citation surface expands. Brands that have treated AI visibility as a long-term concern just had their timeline shortened.

Sonnet 5 at Sonnet prices doing Opus work means more enterprises will run more capable models on more tasks. Every query those models field is a moment where a brand is either cited or absent. The cost barrier that kept some organisations on lighter, less authoritative models just dropped.