Claude Sonnet 5 has appeared in weekend leaks with an internal date string of February 3, 2026, though it remains unclear whether that reflects a public debut or an internal milestone. The timing would place the model announcement in the same week as Super Bowl LX on February 8, a period when AI labs have increasingly synchronized marketing efforts with major consumer events.
What leaked
The leaks describe a new Anthropic model labeled Sonnet 5, with a variant specifically called the non-thinking build. Visual examples and short demos circulated by testers show a mix of creative and engineering outputs — from an ASCII world map to UI-centric rendering and a styled chess layout — credited to vvirtr.
Early hands-on testing
Hands-on testing reported by TestingCatalog indicates that the non-thinking Sonnet 5 variant already appears competitive on math relative to current frontier models, while showing stronger coding output in several workflows compared with Claude Opus 4.5. A recurring highlight in those write-ups was structured visual generation: the ASCII world map prompt reportedly produced a level of completeness and detail the tester had not seen before. Additional examples included newsletter landing page generation and UI rendering tasks.
Technical notes and positioning
The observed build was listed with a 128k context window, though the release version’s context size was noted as likely to differ. The leaks also suggest Sonnet 5 is aimed at a faster, lower-cost tier than Opus, positioning it as a general-purpose model for developers and everyday users rather than a premium research flagship.
If pricing and latency align with those signals, Sonnet 5 would continue Anthropic’s recent pattern of shipping a broadly usable “workhorse” model intended for everyday developer workflows and mainstream use, alongside OpenAI and Google competing for the same audience.
What’s visible and what isn’t
Available material comes from early testing and leaked builds; public availability, official specs, pricing, and final context-window decisions were not included in the leaked reports. The demos and performance notes are limited to the examples circulated by testers and the outputs shared by vvirtr.
Original reporting: https://www.testingcatalog.com/anthropic-is-about-to-drop-sonnet-5-during-super-bowl-week/?