Claude 5

Anthropic shipped the Claude 5 generation across mid-2026, with Claude Sonnet 5 arriving on June 30 and Claude Opus 5 on July 24, alongside Claude Fable 5 at the top of the range. The structure of the family reflected how the market had segmented: a scalable default for volume work, a high-capability tier for the hardest problems, and a middle model positioned to come close to frontier intelligence at roughly half the cost.

Opus 5 was described as thoughtful and proactive, reaching state of the art on coding and knowledge-work evaluations including Frontier-Bench and GDPval-AA while delivering markedly better performance than its predecessor at the same price. Sonnet 5 became the default model on the free and paid consumer tiers, launched at an introductory two dollars per million input tokens and ten per million output tokens before moving to standard pricing.

The generation arrived into a release cadence that had become genuinely unusual. Between February and April 2026 the three leading laboratories collectively shipped seven frontier models, a new state of the art roughly every eleven days. What had been three or four significant releases per year in 2024 had become three or four per month.

Claude 5 marked the maturity of a metric that barely existed two years earlier. Benchmarks like GDPval-AA measure performance on real occupational tasks across dozens of industries rather than on academic questions, and evaluations of this kind became the ones labs led with. The field had moved from asking whether a model knows things to asking whether it can complete the work someone is actually paid to do.

The release rhythm around it carried its own lesson, and it is the one most visible from inside an engineering organization. When the frontier moves every few weeks, no system built tightly around the quirks of a particular model retains its value for long. The durable investment shifted toward the surrounding harness: evaluation suites that can be pointed at a new model in an afternoon, observability that survives a swap, and permission layers that do not care which model sits behind them. Capability became a commodity delivered on a schedule, and the ability to absorb it became the scarce skill.