The Pulse
Sber’s Cowork Bench Lists 475 Tasks; GitHub Repository Lists 496
CNews reported that Sber’s Cowork Bench covers 475 office tasks, while a public GitHub repository with the same name describes 496. Neither page explains the difference, and the available report attributes its details to Sber representative
AI.info Team · Written with AI ·
CNews reported that Sber’s Cowork Bench covers 475 tasks; a public GitHub repository with the same name describes 496. The CNews report says the details were supplied by Sber representatives. The GitHub project does not identify Sber as its publisher, and neither page explains the difference in counts. The figures should therefore be treated as two separate descriptions, not as evidence that one count is wrong.
Sber’s test measures completed work
Cowork Bench evaluates whether an agent can finish an office workflow, rather than whether a model can produce a convincing answer. According to CNews’s report of the information provided by Sber, the suite spans 12 business functions and 26 corporate systems, including work with spreadsheets, documents, email, calendars, data storage and knowledge bases. Tasks can require several steps across those tools, such as preparing an analysis and sending its results onward.
One example in the report asks an agent to compare budget plans with actual figures across four departments, calculate and explain deviations, prepare a report and presentation, revise a forecast and send emails. Another asks it to organize a team business trip by selecting employees, finding tickets, calculating costs and adding meetings to calendars. Those scenarios test whether an agent can coordinate the pieces of a job, not just answer a question about them.
Andrey Belevtsev, Sber’s senior vice president and head of Technology Development, described the benchmark’s intended use:
«Cowork Bench — первый бенчмарк в таком классе, он помогает оценить, насколько хорошо агенты справляются с реальными рабочими процессами и какую их часть могут забрать на себя.»
Belevtsev also said the benchmark is open for developers and companies to use. The report describes GigaCowork as a product of Salyut for Business, a company within Sber’s ecosystem.
GigaCowork completed 75.2% in Sber’s comparison
Sber representatives told CNews that GigaCowork developers compared nine agent software setups using one open model, the same tasks and data, and the same checking procedure. The setups included Goose, Claude Code, DeepSeek, Pi, Codex CLI, OpenHands, OpenCode, Strands baseline and GigaCowork. CNews reported completion rates ranging from 62% to 80%, with GigaCowork completing 75.2% of the tasks, putting it among the stronger results.
The comparison is useful because it isolates differences in the software surrounding a model, according to the account supplied by Sber. However, CNews’s report does not name the open model or provide a full score table or task-by-task breakdown. The reported percentages show the range Sber observed, but do not establish how each system would perform with another model or a different set of tasks.
The 496-task repository leaves a version question
The GitHub repository presents Cowork Bench as a multi-tool benchmark with 496 end-to-end tasks, roughly 120 tools available per task, and deterministic evaluation through programmatic checks and SQL side-effect verification. Its README says the evaluation runs locally using Docker Compose and a seeded PostgreSQL database, without external API calls during evaluation. It also says task prompts are in Russian and describes coverage of Russian-specific domains such as railways, the Moscow Exchange and 1C human resources.
The repository credits Toolathlon as the project it builds on, while describing its own work as an expansion with Russian-localized tasks and services. Those implementation details differ from the 475-task count and corporate-system framing in CNews’s report of Sber’s announcement. Until Sber or the repository maintainers explain the relationship, the available pages give two different published figures: 475 in CNews’s October 16 report and 496 in the GitHub README.