Guides, comparisons, and insights on LLM routing and AI API cost optimization.
Claude Code attaches experimental anthropic-beta headers that third-party endpoints do not implement, so requests are rejected before the model sees them — and the 400 usually blames a parameter instead of the header. How to confirm it and where to strip them.
A 400 on the step after a tool call: duplicate tool_result blocks for one tool_use id, an orphaned result, or 'value did not match any expected variant'. Why translation layers break the tool-call contract, and how to find the offending turn.
Claude Code fails with a 400 unknown-parameter error the moment you point it at LiteLLM, Bedrock, Vertex or any OpenAI-compatible proxy. Why the client sends fields the upstream has never heard of, and the three places the fix can live.
Claude returns a 400 when your prompt plus max_tokens exceeds the model's context window. Why long agent sessions hit it suddenly, how the budget actually adds up, and five fixes ranked by how much conversation you keep.
Pointing Cursor or Claude Code at a local Ollama or LM Studio endpoint and getting connection refused, ECONNREFUSED, or a silent hang? The five real causes — wrong port, missing /v1 suffix, loopback binding, the app not running, and cloud-side IDE features that can never reach localhost.
OpenAI's 401 invalid_api_key is rarely a wrong key — it is usually a stale environment variable, a trailing newline, a revoked key, or the wrong provider's key reaching an OpenAI-compatible endpoint. How to find which one in under a minute.
The OpenAI model_not_found 404 usually is not a typo — it is access, org, or endpoint mismatch. Why a model works for someone else but not your key, the four real causes, and how to list the models your key can actually see.
Anthropic's HTTP 529 overloaded_error means the API is under load — not that you did anything wrong. How it differs from 429 rate limits, why it comes in bursts, correct retry strategy, and how to keep coding agents productive through it.
Why the Anthropic credit balance error appears even for Claude Pro/Max subscribers, how prepaid API credits work, the Claude Code case where a login/API-key mixup triggers it, and how to fix and prevent it.
Gemini's RESOURCE_EXHAUSTED 429 covers several different limits: free-tier per-minute and per-day caps, paid-tier rate limits, and project quota. How to tell which one you hit, fixes for each, and the free-tier math that surprises people.
The OpenAI 429 insufficient_quota error is a billing problem, not a rate limit. What the error actually means, why it hits accounts with a paid subscription, and how to fix it in the API dashboard, in code, and in coding agents.
The context_length_exceeded 400 error means your messages plus max_tokens overflowed the model's context window. Why coding agents hit it mid-session, how the math actually works, and five fixes ranked by how much history you keep.
A factual guide to Claude Code Router — the open-source local gateway that routes Claude Code, Codex and other coding agents across OpenAI, Anthropic, Gemini, OpenRouter, DeepSeek, Kimi and more — with install commands, the feature list from its README, and an honest comparison against a hosted router like CodeRouter.
A maintained reference of real LLM API errors, organized by exact error message: DeepSeek reasoning_content 400s, Claude Code 401s, Cursor base-URL failures, Gemini endpoint 404s and more — each with cause and working fixes.
Seeing 'cc switch local proxy failed while handling codex endpoint /responses ... upstream_status: http 400; cause: the reasoning_content in the thinking mode must be passed back to the api'? What the error chain means, why proxies trigger it with DeepSeek thinking models, and three working fixes.
Your agent logs show 'deepseek returned tool calls without replayable thinking content; continuing with degraded reasoning'. What the warning actually tells you, whether degraded reasoning matters, and how to restore full thinking round-trips.
Using langchain-deepseek with deepseek-reasoner and tool calling? The 'reasoning_content in the thinking mode must be passed back' 400 appears when chain state drops the thinking block. Version guidance, message-history handling, and a router-side fix.
Which models should an agent router actually route between? The 2026 lineup by tier — frontier reasoning, workhorse implementation, and cheap bulk models — with real per-token pricing, what each tier should handle, and how to audit which model served your request.
What an agent router API actually is — an OpenAI-compatible endpoint that sits between your coding agent and model providers. Setup examples for Cursor, Claude Code, and raw HTTP, plus the API features that separate good routers from thin proxies.
How to override the base URL to use a Google AI Studio (Gemini) API key with OpenAI-compatible tools — the official compatibility endpoint, correct model names, common 404/401 errors, and when a router is the cleaner fix.
Agent routers sit between your coding agent and your LLM provider — which means they see your prompts, your code, and sometimes your API keys. Here's what a router can actually access, the six real risks, and a legitimacy checklist for evaluating any routing service.
Moonshot's Kimi K3 (2.8T params, 1M context, native vision) just beat Claude Opus 4.8 and GPT-5.5 on coding and agentic benchmarks — at $3/$15 per million tokens. Here's a router operator's honest read: real pricing math with cache economics, phase-by-phase fit for Cursor / Claude Code / Aider workloads, and what changes for your API bill.
Step-by-step guide to configuring Cursor's Override OpenAI Base URL for custom models — covers OpenAI-compatible providers, Anthropic via router, DeepSeek, model settings, and troubleshooting. Everything you need to route Cursor through your own endpoint.
Pointing Claude Code at a proxy or router via ANTHROPIC_BASE_URL and getting 401 authentication_error / invalid x-api-key? The cause is almost always which env var you used. Here's the header-level difference and the exact working setup.
Setting a custom OpenAI base URL in Cursor can break Anthropic/Claude models with 422 errors, and several Cursor features ignore the override entirely. The exact failure modes, working configurations, and a router setup that avoids the trap.
DeepSeek V4 returns HTTP 400 when tool calls meet thinking mode and your client strips reasoning_content. Why it happens, three working fixes (client patch, router-side fix, disable thinking), and how to verify.
Looking for an agent router alternative? Compare OpenRouter, LiteLLM, RouteLLM, Portkey, and CodeRouter — features, pricing, best practices, and how phase-aware routing cuts coding costs 70-90%.
Our smart-routing product had a blind spot: 80%+ of traffic was bypassing the router because client defaults hardcoded Opus. After 12 hours of fixes, real data: per-request cost down 36%, Opus share of spend from 80% to 45%, honest savings from 60% to 91%, ~$90K/month saved. Here's the story.
我们的智能路由产品发现了一个尴尬的盲区:大部分流量根本没在路由。修完之后 12 小时真实数据:每请求成本降 36%,Opus 花费占比从 80% 降到 45%,真实 savings 从 60% 升到 91%。这是过程和数据。
中文圈 LLM 中转站的本质是 OpenAI / Anthropic API 转售,大多数把所有请求都发给 Opus 或 GPT-5。我们用 5000 次真实路由的生产数据告诉你:为什么 80% 的编码 token 应该用 V4-Flash / Sonnet 而不是 Opus,以及任务感知路由和中转站的本质区别在哪里。
GitHub Copilot just raised prices 6x. Claude Code heavy users burn $50-200/day. Here's why AI coding bills spiral — and the 5 concrete steps to cut them by 60-90% without losing productivity.
OpenRouter is great for model access, but it won't cut your coding bill. Here are 7 alternatives that actually reduce what you spend on Claude Code, Cursor, Aider, and Copilot — with real pricing and tradeoffs.
Chinese LLMs ship with a 'thinking mode' that breaks the OpenAI/Anthropic API contract on turn 2. Real production errors from Claude Code, Cursor, and Aider — plus the one-line fix per provider, a cross-provider audit checklist, and why this trap will keep biting API gateways through 2026.
Most LLM routers pick by cost or capability. Phase-aware routing detects which phase of coding you're in (plan / implement / debug / test / refactor / docs / small-edit) and routes each call independently. After 30 days in production: 78% cost savings, 7 different models touched per Claude Code session, no measurable quality regression. The detection algorithm, the model assignments, the pitfalls, and the data.
Four major coding models dropped in one week (April 20–23, 2026): GPT-5.5, GPT-5.4, DeepSeek V4 Pro/Flash, and Kimi K2.6. One table, one decision tree, one verdict per task. Skip the marketing — here's what actually changed for coding agents and which model to pick for which job.
DeepSeek V4 ships in two tiers. V4 Pro at $1.74/$3.48 scores 81% on SWE-bench Verified — near-Opus territory. V4 Flash at $0.14/$0.28 is 12× cheaper, still 1M context, still strong on implementation. Here's the decision matrix for coding agents, plus why pinning just one wastes money.
OpenAI's GPT-5.5 tops the Artificial Analysis Intelligence Index at 60, beating Opus 4.7 on Terminal-Bench 2.0 (82.7% vs 75.1%) — but Opus still wins real-world SWE-Bench Pro (64.3% vs 58.6%). Here's how to pick per-task, and why paying for both via phase-aware routing is cheaper than committing to one.
Moonshot's Kimi K2.6 (April 2026) scores 58.6% on SWE-Bench Pro — tied with GPT-5.5 — at $0.60/$4.00 per 1M. It's open-weight, 256K context, purpose-built for long-horizon agentic coding (300 sub-agents, 4000 steps). Full review, benchmarks, and when it's the right pick over Claude Opus, GPT-5.5, and DeepSeek V4.
Cut Claude Code's bill 50–80% by routing every call through a phase-aware proxy. Includes the AUTH_TOKEN-not-API_KEY trap, the experimental-betas fix for v2.1.x, and a complete settings.json template that just works.
Aider's architect + editor split is brilliant — but both modes default to Opus. Here's how to combine Aider's --architect flag with phase-aware routing for 80%+ cost reduction without touching your workflow.
OpenRouter and CodeRouter sound similar — both are 'routers'. But they solve different problems. OpenRouter gives you multi-model access; CodeRouter reduces your coding agent bill by picking the cheapest capable model per request automatically.
Cursor Pro burns tokens fast when you hit fast-request limits. Here's how phase-aware API routing cuts your real monthly coding spend without switching away from the Cursor IDE.
Head-to-head: DeepSeek V3 at $0.28/$0.42 vs. Claude Sonnet 4.6 at $3/$15 per 1M. On coding tasks, when does the 15× cost difference show up in output quality — and when doesn't it?
GitHub Copilot's $10/month is cheap but locks you into their model choices. For power users who hit Copilot's rate limits, phase-aware routing via a Custom Model endpoint delivers more context + cheaper per-token + model diversity.
Most LLM routers pick one model and stick with it. Phase-aware routing detects which *phase* of coding you're in — planning, implementing, debugging, testing — and picks the cheapest capable model per phase. Here's how it works in <10ms.