Esc to close · ⌘K / Ctrl-K opens search anywhere
Notes and how-tos from BharatRouter.
Part 2: the same three models inside a real coding agent, priced per task through one neutral gateway. Kimi K3 lost single-turn on cost-per-correct — but wins the agent at ₹10.87/task, because caching is the whole game. Cache-accurate four-component cost; deterministic grading; 16/16 solved.
A reproducible quality × cost benchmark of GPT-5.6 Sol, Claude Fable 5 and Kimi K3 through one gateway — accuracy AND ₹-per-correct-answer across 5 axes, deterministic grading, Wilson CIs. Sol wins both; open-weight Kimi beats Fable. And the twist: Kimi has the cheapest tokens but not the cheapest answers.
Moonshot's Kimi K3 (2.8T MoE, 1M context, always-on reasoning) in brcode, OpenCode, Claude Code and Codex CLI — one BharatRouter key, your Moonshot key as BYOK. Copy-paste configs, a pricing comparison across providers, and sample code on GitHub.
Warp is the terminal a TUI coding agent deserves. Launch configs for agent + shell panes, workflows for brcode commands, and one rule that keeps Warp AI and your agent out of each other's way.
We measured every model×provider route through the production gateway — the cheapest, snappiest and fastest-streaming models are three different models. How to pick per use case, with real numbers and a live picker at /compare.
BharatRouter Code installs a governed coding agent on GLM-4.7 in one shell command — pick a Codex-style (OpenCode) or Claude-Code-style (the real Claude CLI) front-end, both routed and metered through your own BharatRouter key.
Point Claude Code at BharatRouter and drive it with an open GLM model — one br- key in your client, GLM set once as platform BYOK. The gateway translates the Anthropic wire to GLM and back, streaming and reasoning included.
API keys were built for humans. Agents are short-lived, numerous and act on someone's behalf — so they deserve an identity, not a shared password. How an agent authenticates to BharatRouter with a verifiable, short-lived identity token, with zero setup by default.
A reproducible benchmark — open GLM vs Claude Opus 4.8 and GPT-5.5 across 14 execution-checked coding tasks, 100 runs each with Wilson 95% CIs. On Baseten, GLM matches frontier accuracy, runs faster, and costs ~a tenth. Every script and result on GitHub.
We benchmarked the same open GLM family across three hosts through one BharatRouter key — speed, cost, and how to route across all three for reliability with failover and the optimize knob.
The methodology behind our GLM benchmarks: execution-checked correctness (MBPP, SWE-bench), accuracy + latency + cost reported together (HELM), and Wilson 95% confidence intervals on every pass-rate. Interim results today; a broader, contamination-controlled benchmark coming.
Stand up a governed, open-source coding agent on GLM with a single shell command — installs OpenCode, prompts and verifies your keys, and runs a first query so you see it working immediately.