Integration guides and usage notes for ROIBest AI. Base URL is https://ai.roibest.com.
Pointing Claude Code at ROIBest AI takes two environment variables. What actually trips people up is picking the wrong protocol group for the key, and adding /v1 to ANTHROPIC_BASE_URL — one curl call catches both.
A working Claude API curl example is one POST to /v1/messages with three headers and a JSON body. Copy-paste commands for streaming, images, tools, token counting, model listing and debugging, based on Anthropic's docs.
Use the OpenAI SDK with Claude through Anthropic's official compat layer: base_url, model IDs, and which parameters are supported, capped, or silently ignored.
How to route LiteLLM to Claude: the SDK-versus-proxy split, the three backend paths and what each needs, which OpenAI parameters translate cleanly and which do not, streaming and tool-call pitfalls, and an error table mapping symptoms to causes.
The OAI compatible provider in VS Code is deprecated. Custom Endpoint replaces it with three API types, a new JSON config file, and one URL convention that trips up most setups.
The phrase covers two unrelated setups: an LLM gateway set with ANTHROPIC_BASE_URL and a credential, and a corporate HTTP proxy set with HTTPS_PROXY. Here are the exact variables for each, where to put them, and how to confirm which one is live.
A Claude API error 400 means the key worked but the request broke a rule. Nine real causes, from prefill and sampling params to tool_result order and images.
In Anthropic's ecosystem, OAuth is the Claude account sign-in used by Claude Code and other native apps. Here is how it differs from a Console API key, what claude setup-token creates, and what Anthropic's terms say about third-party tools.
The base URL is whatever the SDK appends /chat/completions to — usually your provider host plus /v1, with no trailing slash. The four ways it is set wrong, what each client expects, one curl that verifies it, and a table mapping each error back to its cause.
"OpenAI-compatible" describes request and response shape, not that every parameter is honoured. This guide maps the four layers where compatibility breaks, the parameters most often silently ignored, the streaming and tool-calling differences that break clients, and a paired-request method for telling which layer you are in.
Two mechanisms for two jobs: output_config.format shapes the response, strict on the tool definition shapes tool arguments. Why the prefill trick now returns 400, why forced tool_choice stopped working, and why you must parse tool inputs instead of matching them.
The context window is total input per request, not memory — and the Models API calls it max_input_tokens, not context_window. What actually consumes it, why tool results dominate agent loops, how caching interacts with it, and when to compact versus clear.
Five different problems wear the same costume, and four of them are not the key. What 401, 403, 404, 400 and 429 each actually mean — plus the credential-resolution order that makes a valid key look ignored.
The Batches API runs Messages API requests asynchronously at half price. The three decisions that make or break the integration: design for 24 hours not one, size against 256 MB not 100,000 requests, and key results by custom_id not position.
Configuration arrives through environment variables and settings files at four scopes that merge key by key rather than replacing each other. Which layer wins, what belongs in each, why MCP is separate, and how to verify a change actually took effect.
A complete first mile for the Claude API: what you need before the first call, what every part of a Messages request does, how to read the response, and the four mistakes that eat a beginner's first afternoon.
Response time is not one number. Time to first token, tokens per second and total completion time behave differently, respond to different fixes, and are routinely confused. What drives each, how to measure them without fooling yourself, and which levers actually move.
Benchmark tables go stale. At the integration layer the two APIs diverge in four concrete places: where the system prompt lives, required max_tokens, strict message alternation, and how caching is configured - plus the parameters compatibility layers drop silently.
A practical guide to Claude API streaming: the server-sent event transport, the full event sequence from message_start to message_stop, how to accumulate deltas by block index, why tool-call JSON fragments must be parsed once, and the failure modes worth handling.
Claude Code talks to an endpoint, not a vendor. How the base-URL swap works, which capabilities survive it, a five-check acceptance pass, and when swapping costs more than it saves.
Claude Code itself is not region-locked — the constraint is the API account behind it. Three routes resolve that: a first-party Anthropic account, a protocol-compatible provider, or a self-hosted gateway. Here is what each requires and who each one fits.
Four open-source LLM gateways cover most real deployments: LiteLLM, Portkey, Envoy AI Gateway, and the Kong/APISIX AI plugins. Here is what separates them, the five questions that actually decide your pick, and what self-hosting really costs.
An OpenAI API proxy forwards OpenAI-format requests to a host you choose, so your client keeps working after a one-line base URL change. Here is what it buys you, when a gateway fits better, and the five behaviors to verify before routing production traffic.
A claude.ai subscription does not include API access. Here is the full path — console organization, payment, verification, workspace-scoped keys, rate-limit tiers — plus the cloud and gateway routes and what each changes about your integration.
An LLM API gateway puts one endpoint and one credential in front of multiple model providers, then handles routing, fallback, per-key quotas, usage accounting, and logging. How it differs from a proxy, when you do not need one, and the seven things to verify before choosing.
An OpenAI-compatible API copies the Chat Completions request and response shape, so existing clients work after a base-URL change. What the compatible surface covers, where it breaks, and how to verify an endpoint in three curl calls.
A Claude API proxy forwards your requests to Anthropic and returns the response — solving network reachability, shared billing, and OpenAI-compatible tooling. How it works, proxy vs relay vs gateway, and five checks before you switch.
Codex speaks the OpenAI protocol, so the key, the endpoint and the config format all differ from Claude Code: an OpenAI-compatible key, a base_url that does include /v1, and settings in config.toml rather than environment variables.
The official SDKs need no code changes — only a different base URL and key. Here are all the endpoints the gateway exposes, plus the most common source of 404s: the OpenAI SDK wants /v1, the Anthropic SDK does not.
Manual setup is not hard, it is just easy to mistype. The import flow generates config that matches the key's own protocol group, taking the two easiest mistakes — group mismatch and the /v1 suffix — off your hands.
Claude API cost optimization starts with the usage object: find which token bucket dominates your bill, then apply the lever that targets it, from effort and caching to tool overhead, images and batching.
Mainland China, Hong Kong and Macau are not on Anthropic's supported regions list; Taiwan is. Here is what the policy actually says, why it is an eligibility question rather than a technical one, and how the cloud routes differ.
Anthropic's usage tier ladder in one place: how organizations are placed on Start, Build, Scale or Custom, what each tier's spend cap and per-model limits are on the official page, where to check your tier in the Console, and what to do when it is not enough.
The Anthropic API billing page lives in the Console, not in Claude.ai. Where to find it, what usage, cost, credit balance and invoices each actually count, why workspaces decide your breakdown, and the four settings that change next month's bill.
Published per-million-token rates are the starting point of a price comparison, not the substance of one. Input/output asymmetry, incomparable tokenizers, reasoning billed as output, cache discounts worth an order of magnitude, and the only metric that decides the winner.
Prompt caching is a prefix match, and it fails without an error. The breakpoint rules, the per-model minimum that swallows short prompts, when the 1-hour TTL costs more than it saves, and the usage fields that prove a cache is real.
Rotation as a quarterly calendar ritual causes outages without reducing much risk. The version that works is narrower: keys scoped so rotation is cheap, an overlap window instead of a swap, verification by traffic before revoking, and a rehearsed path for the unplanned rotation.
Requests, input tokens, and output tokens are three separate limits measured on separate axes. How to tell which one is binding, what fixes each, retry behaviour worth shipping, and the extra limit layer a gateway adds.
Getting Claude API credits is a payment-layer problem, separate from pricing and from access. Three routes exist — a first-party account, a cloud marketplace, or a compatible gateway. Here is why cards get declined, how prepaid balances behave, and what to check before you top up.
Claude API and Anthropic API pricing are the same list: per-million-token rates with output billed at 5x input. Current rates by model, how prompt caching and batching cut the bill, why thinking tokens count as output, and the five levers that actually reduce spend.
The console records usage per request. This is what each column means, and where to look first when a call costs more than the last one — most of the variance comes down to model, output length and cache hits.
One account can issue many keys, each with its own quota and expiry. This is not just access control — it is the cheapest way to turn 'who spent what' from one opaque bill into readable data.
