The USD 300 card limit, the OTAF form, RFCD accounts, why crypto top-ups are off the table, and what card fees really cost on small top-ups. Plus what changes when you pay an AI gateway in taka.
Eight low-cost models you can run a coding agent on, with official list prices, context limits and the catches: DeepSeek's peak hours, GPT-6 Luna's tool-calling rule, MiniMax's 512K tier and Gemini's January price change.
Claude Sonnet 5.5, GPT-5.6 Sol, Grok 4.7, Kimi K3, GLM-5.3, Qwen 3.8 Max, DeepSeek V4 Pro, MiMo V2.6 Pro and MiniMax M3 compared on list price, context, tools and vision, plus a simple start-cheap, escalate-on-failure routing pattern.
Million-token context lets an agent hold a whole repository and a long session. It doesn't make each turn cheaper or faster, and some providers raise the price above 200K, 272K or 512K tokens. Here's the arithmetic.
How prompt caching is priced at DeepSeek, Anthropic, OpenAI, Moonshot, Alibaba and others, why coding agents benefit more than chat apps, what quietly breaks cache hits, and what you can and can't control through a gateway.
A practical method for estimating what a coding agent will cost per month: measure tokens per session, split input from output, account for caching, multiply by working days. Worked through with DeepSeek V4.1 Flash and Claude Sonnet 5.5 list prices.
Claude Code only speaks the Anthropic Messages API, so any gateway that speaks it can put a different model behind the CLI. Here is how the environment variables fit together, what Anthropic doesn't support, and which settings keep a non-Claude model usable.
OpenClaw and Hermes Agent keep working after you close the laptop, which is the point and also the risk. How to connect both to a pay-per-token gateway, cap what they can spend, and pick cheap, fast models for the jobs they actually do.
The Tokens CLI writes one key into Claude Code, Codex CLI, OpenCode and Crush in a single run, backs up every file it touches, and can show you the plan first. What each flag does, which files change, and how to wire up Cline, Kilo, Continue and Zed by hand.
Agent loops re-send their whole context every turn, so a stuck one gets expensive fast. How to layer per-key monthly caps, model allow-lists, usage windows and alerts, poll your remaining budget from a script, and recognise the errors when a limit trips.
Three request formats, one job. Where the system prompt goes, how tool calls and their results are shaped, what the streaming events look like, why output limits and usage fields differ, how a gateway serves the same model in more than one format, and which coding agents speak which.
A field guide to the errors an AI gateway returns, using Tokens' real error codes: which ones mean fix your setup, which mean top up, which mean wait. Plus a retry function that honours Retry-After and never retries a 4xx, and what to put in a support ticket.
