Blog

Notes from the fleet

Cost breakdowns, benchmarks, and architecture teardowns — how we think about serving open models cheaply, and what the numbers say.

2026-10-24·Featured·11 min readWhat a coding agent really burns: 1.6B tokens of receiptsA dated receipt for one real coding-agent workload: 1.6B prompt tokens all-time, a 224M-token peak day, 93% of prompt volume served from cache — and the arithmetic that turns that into dollars per task, per day and per month.coding-agentstoken-economicscachereceiptspricingRead the article →

2026-10-24·13 min readClaude Code alternatives (2026): priced, sourced, honestEvery real Claude Code alternative in 2026 — open-source harnesses (Aider, Cline, OpenHands, opencode), local stacks, and metered APIs — priced per task at the same workload, with the runaii row disclosed and the row where we lose stated plainly.claude-codecoding-agentsalternativespricingcomparison2026-10-24·18 min readAI coding plans compared (2026): Claude, ChatGPT, GLM, Qwen, Kimi, OpenCode, Subconscious and Token-MaxAn AI coding plan comparison built on arithmetic instead of vendor list prices: what one agent task costs on the meter, what a 224M-token day costs, and where each flat plan's break-even actually lands — Claude, ChatGPT, GLM, Qwen, Kimi, OpenCode Go, Subconscious and our own Token-Max, every price checked 2026-10-24.coding-planspricingcomparisoncoding-agentstoken-max2026-10-14·15 min readHow to run Qwen3.8 27B in 2026: the two-minute API path, the honest local guide, and the break-even mathHow to run Qwen3.8 27B, every path priced and verified on 2026-10-14: an OpenAI-compatible API you can wire in two minutes (no GPU), the local Ollama install with real VRAM requirements per quant (18GB at 4-bit, 56GB at BF16), other providers, and the arithmetic that decides which one you should pick.qwentutoriallocal-inferenceopenai-compatiblepricing2026-10-14·18 min readPrompt caching pricing comparison (2026): the column every provider buriesA prompt caching pricing comparison that leads with the cached-read column: cache read and write prices per 1M tokens for OpenAI, Anthropic, Google Gemini, DeepSeek, Kimi, OpenRouter, and runaii.cloud — TTLs included, every rate verified 2026-10-14.prompt-cachingpricingcomparisonllm-api2026-10-13·16 min readGLM 5.3 Flash API alternatives — every lane, pricedHunting a GLM 5.3 Flash API alternative? Every option priced and verified: z.ai's coding plan, the OpenRouter listing at $0.036/1M input, sibling models, DeepSeek V4.1-Flash, Kimi K3, and one honest row for runaii.cloud — cached-input column included.glmpricingcomparisonopenai-compatible2026-10-09·13 min read890 bytes per token: DeepSeek V4.1-Flash KV cacheDeepSeek V4.1-Flash stores its KV cache in about 890 bytes per token — roughly 3.9x less than the previous generation. We tear down the published architecture behind that number: the causal encoder-decoder split, cross-layer KV sharing, FP4 quantization, and the sparse indexer that makes it all retrievable.deepseekkv-cachearchitectureresearch