41 articles
Per-token pricing, caching credits, batch discounts, and hidden costs across OpenAI, Anthropic, Google, and open-source LLM providers. Includes four real workload simulations and cost optimization strategies.
Gemini 3.1 Pro tops the LM Council April 2026 board on GPQA Diamond and ARC-AGI-2 at 50% lower cost — but Opus 4.7 still leads on coding. The honest task-by-task decision guide.
Moonshot's Kimi K2.6 hits ~74% SWE-Bench Pro at $0.30 per typical run — 17-25x cheaper than Opus 4.7. Real benchmarks, where it falls short, and the two-tier routing pattern teams use in production.
Zhipu AI's GLM-5.1 beat Claude Opus 4.6 on SWE-Bench Pro at 7x lower API cost. Where the headline holds (batch coding, cost-sensitive loops) and where Opus still wins (subjective quality, agentic tool use, latency).
DeepSeek V4's 1T-parameter MoE architecture, the Engram learned-memory layer behind its 1M-token context window, real benchmarks vs Claude Opus 4.7 and GPT-5.4, API pricing, and the honest case for when to pick V4.
Aider for CLI / git-native, Continue for IDE-native BYO-model, OpenHands for autonomous multi-step tasks. Real SWE-Bench scores with Qwen 3.5 32B local.
Benchmarks and architecture comparison of Ollama, vLLM, and llama.cpp. Tokens/sec at 7B through 70B, quantization trade-offs, concurrent throughput, VRAM requirements, and a clear decision framework for local dev, production, and edge.
Q8 KV cache halves VRAM with under 0.1% perplexity cost. Q4 K-cache is OK, Q4 V-cache hurts. Asymmetric Q4-K + Q8-V is the magic combo.
RTX 5090 unlocks Qwen 3.5 32B at Q5_K_M with 16K context. NVFP4 native gives 60-80% inference speedup over RTX 4090. Real benchmarks and build guide.
New articles delivered to your inbox. No spam.