model-routing

LegalOn 通过依据任务复杂度在 GPT-6 Astra、GPT-6 Luna 和 GPT-6.1 Sol 之间智能选择模型，并配合预算管控，将 Codex 成本削减一半，同时维持开发速度。

Cisco researchers evaluate semantic entropy as an escalation signal for routing queries from small to large LLMs across three benchmarks, showing it works on GSM8K but can be misleadingly matched by naive difficulty heuristics. The paper's main contribution is a checklist of evaluation pitfalls—question-only baselines, ambiguous success metrics, benchmark saturation, and hidden sampling costs—for assessing routing signals.

The author describes a method for deciding when an agent's repeated task can safely move from a frontier model to a cheaper one, using held-out past runs, human review, and one-click rollback, while seeking critique on the evaluation approach and disclosing they build Operant in this space.

一位用户分享如何在 Claude Code 中通过三层模型分工（Opus 5.5 负责规划与最终 diff，Sonnet 5.5 负责侦察/执行/文档，Fable 5.1 作为 advisor 仅在关键节点被咨询）大幅降低成本，并给出可直接粘贴的配置提示词，附带 Claude Code advisor 工具官方文档。

GitHub announced that Copilot will soon gain intelligent local model routing, automatically sending tasks to a local model when it is most suitable in order to save AI credits. The feature is described as the next step in GitHub's Project HydraFusion vision.

LangChain shared a video by Sydney Runkle on building a model router into a coding agent harness, demonstrating how to cut costs by 64% without reducing quality by matching tasks to appropriately-sized models.

OpenAI launched a B2B enterprise marketplace at DevDay with 32 partners, letting customers apply existing OpenAI commitments to third-party products including open models served by Baseten — a strategic move to become the hyperscaler layer for AI apps as open-weight models grow in volume.

This paper measures the per-token 'sufficient compute' of LLMs using a Mixture-of-Agents panel of fifteen models across Qwen, OLMo, and R1-distilled families, finding a 0.5B agent reproduces 92-95% of reference tokens while the costliest 10% of tokens account for 64-80% of FLOPs. The resulting per-token compute map improves model routing latency from 7.59s to 5.12s and enables speculative drafting with 32.6% fewer draft tokens than fixed-window baselines.

This paper presents a paired, self-audited evaluation of two System-1 decision models (open-weight Laya and hosted Jev) across 11 agent-harness decision points, finding Jev significantly more accurate on 9 of 11 tasks while auditing several analysis errors that had distorted prior deployment claims.

作者赞同判断：智能体中超过一半的所谓「推理」其实是路由决策，应把工具选择、路径分支等交给小模型以毫秒级速度完成，只在真正需要推理的步骤上花费昂贵的大模型 token。

A tweet discussing a testing trick that reveals whether a user is in the gray release (staged rollout) for Claude 5.5 — asking "Who is Tibo" makes the model leak that it's actually 5.5.

This arXiv paper introduces Contingent Exposure Routing (CER), a framework for assigning displaced financial decisions across AI model endpoints during outages to minimize market-impact risk, showing that shared backups create concentration floors and that randomized routing strategies reduce exposure risk by 3.3-10.5% in synthetic and replayed scenarios.

A step-by-step guide to building a Claude Code mod called run-ledger that integrates with Jev for model routing, cost estimation per run, and subagent status tracking. The post walks through connecting Jev, authoring the mod, testing hot reload and duplicate counting, and installing it as a persistent plugin.

Aster by AsterWise launches an intelligent model routing service that dynamically routes tasks across models, with aster-code for coding agents and aster-work for custom agents and workflows, offered on a paid, prepaid basis.

A tweet praising Factory's AI coding automation tool, highlighting its ability to handle low-leverage work and an auto-model router that automatically selects the best model among new releases.

LangChain shares a practical guide on building a model router inside the agent harness—covering task understanding, model selection, and outcome tracking—claiming a 64% reduction in median cost per thread with no quality loss.

LangChain shares how they built a model router inside the agent harness for their open source coding agent Open SWE, cutting median cost per thread by 64% with no measurable quality change by routing tasks to cheaper models where frontier intelligence isn't needed.

Onepin launches as a model routing and quality-control layer for voice generation, sitting on top of existing TTS providers to normalize scripts, select voices per line, and automatically detect and fix pronunciation or naturalness issues at scale.

OpenRouter released Pareto 26.10 Preview from TheUnbiasedCo, a multimodal composite model that internally runs several models, evaluates their answers, and returns the best one via a single API response at $0.80/M input and $3.20/M output with 1M context and tool calling.

Weave Router 2.0 is an open-source model router for coding agents that intelligently switches between LLMs, matching GPT-6 Astra's pass rates on Terminal Bench 4.0 and SWE Atlas while cutting costs to ~52% and speeding tasks 2.2–2.5x via an HMM-based architecture, larger training data, and smarter cache-eviction calculations.