What this video shows
The video compares Claude Code with self-hosted GLM 5.3 through Revolut’s reported deployment for about 2,000 engineers. Revolut says that a lower-cost default removed strict token budgets, so engineers used agents for requirements, tickets and documentation as well as code.
The decision lesson is broader than one model: measure accepted work, lead time and total operating cost on your workload. Keep stronger or multimodal models available for tasks where the cheaper default falls short.
Check the underlying Revolut conference talk for the company-reported deployment numbers and serving constraints. Compare current rates in Anthropic’s Claude Code cost guide and Z.AI’s pricing table. Use the Vals Terminal-Bench 4.0 leaderboard as one fixed-harness signal, then test your own repositories.
What you will learn
- A lower-cost default can increase use across the full engineering workflow.
- A public benchmark measures a fixed task set, not your team’s accepted output or lead time.
- Self-hosting adds GPU capacity, serving, monitoring, security and staffing costs.
- A router can reserve stronger or multimodal models for tasks where they earn the extra cost.
How to apply this safely
- Collect two to four weeks of task-level cost, accepted output and lead-time data from the current tool.
- Choose a representative evaluation set with coding, repository search, documentation and failed-tool cases.
- Price hosted tokens and dedicated capacity using the same traffic mix, cache rate and concurrency.
- Run both models behind the same permissions, tools and verification checks.
- Route tasks by measured quality and cost, then keep a rollback path for every model change.
Important limitations
- Revolut supplied the productivity and cost figures, and the public material does not establish a controlled causal comparison.
- The comparison reflects Revolut’s scale, traffic and infrastructure team. A smaller team may not keep dedicated accelerators busy enough to match its economics.
- Published token rates omit engineering labor, observability, capacity headroom, incident response and idle hardware.
- Terminal-Bench 4.0 represents command-line tasks in one harness and cannot predict every repository or workflow.
Sources to check
- Self-Hosted AI vs Claude Code: When It Pays Off | Devsplainers Devsplainers video published November 6, 2026.
- How Revolut Cut AI Coding Costs 90% With Self-Hosted Open Models David Abitbol’s AI Engineer talk published October 29, 2026.
- Manage Claude Code costs effectively Anthropic documentation covering usage, budgets and enterprise averages.
- GLM model pricing Z.AI’s current hosted token pricing.
- Terminal-Bench 4.0 Vals AI fixed-harness coding benchmark and methodology.
Continue learning on Learnetto
Reduce coding agent cost
Measure cache use, retries, model routing and accepted work.
Switch LLM models safely
Evaluate a replacement with repeatable checks and rollback thresholds.
Build an AI agent test set
Turn production tasks and failures into an evaluation suite.