Posts

Showing posts from October, 2026

Claude Haiku 5.5 降价之后,Agent 的成本为什么仍不能只看 token 单价

Image
北京时间 10 月 8 日清晨,AI 开发者看到的一条新消息,是 Anthropic 在当地 10 月 7 日发布 Claude Haiku 5.5。这次更新值得关注,但理由不是又多了一组跑分。它同时改变了小模型的价格、可调推理投入,以及大小模型之间的分工方式。 官方公告 给出的定位很具体:摘要、上下文压缩、分类、数据库查询,以及编码系统里的子代理。 我的判断是:这次发布更值得被看作一次 Agent 架构的成本重算,而不是一次“廉价模型取代旗舰”的宣告。模型调用变便宜之后,团队必须回答的新问题是:哪些任务可以下放,谁检查结果,失败时怎样升级,以及一次真正完成的工作到底花了多少钱。 先拆开三个容易混淆的降价数字 Anthropic 同时给出了“单价降低 90%”“单价降低 50%”和“平均运行成本降低约 75%”。这三个数字不是互相矛盾,而是口径不同。 根据发布公告,Haiku 5.5 对不超过 100,000 tokens 的 prompt,输入和输出分别为每百万 tokens 0.10 美元与 0.50 美元;超过这个门槛,则分别为 0.50 美元与 2.50 美元。Haiku 4.5 的对应单价为 1 美元与 5 美元。因此,短 prompt 档的单价降低 90%,长 prompt 档降低 50%。 价格及脚注 公告列出的计费项目 Prompt ≤100k tokens Prompt >100k tokens 输入,每百万 tokens $0.10 $0.50 输出,每百万 tokens $0.50 $2.50 缓存读取,每百万 tokens $0.01 $0.05 缓存写入,每百万 tokens $0.125 $0.625 这张表引用的是厂商发布页,不是对所有云平台账单的承诺。具体部署仍需核对供应商当前价格、缓存规则及其他费用。 “平均降低约 75%”则是厂商对完成任务成本的估算。公告脚注称,上一代约 90% 的请求处于较低价格档,同时新 tokenizer 会让同一份工作消耗略多的 tokens。这意味着不能拿每百万 tokens 的价格直接推导每项业务的降幅,更不能把请求数量占比当作支出占比。 还有另一项同期变化:Sonnet 5.5 的缓存读取单价从每百万 tokens 0.20 美元降到 0.10 美元。Anthropic 据此...

How to Use NanoBanana MCP in OpenCode for Multi-Image Editing

Image
Multi-image editing becomes much more useful when it can stay inside the same terminal workflow where you are already writing docs, adjusting product copy, or preparing release assets. What you can do The NanoBanana MCP server is designed for image generation and image editing workflows inside OpenCode. The documented strengths are practical rather than abstract: product background replacement, virtual try-on, subject-consistent edits for people or objects, and editing one image with another image as a reference. In an OpenCode session, that means you can ask for tasks such as: placing a product image into a scene while keeping lighting consistent; changing a model image to wear a referenced garment while preserving face and pose; generating a small illustration for a 404 page or documentation page; using several source images as references without leaving the terminal. How it works OpenCode reads MCP servers from an opencode.json file. Ace Data Cloud exposes Nan...

How to Add NanoBanana Image Editing to Claude Code with MCP

Image
When you are already working in a terminal, switching to a browser just to generate or edit a product image breaks flow. A practical alternative is to connect an image MCP server to Claude Code, then ask for image work in the same place where you read code, edit files, and run commands. This guide walks through the NanoBanana MCP setup for Claude Code using the public Ace Data Cloud documentation. The goal is not to replace your design process; it is to give your coding agent a callable image tool for common builder tasks: README visuals, landing page illustrations, product mockups, and multi-image composition. What you can do NanoBanana is useful when the task needs image understanding, not only fresh image generation. The documentation describes it as a tool that can understand relationships across images—for example, placing an item from one image into the scene of another and adjusting angle and lighting so the result looks natural. In Claude Code, the NanoBanana MCP server ex...

Claude Code и поисковый MCP: воспроизводимый рабочий процесс для исследования и отладки

Когда разработчик разбирает сбой на удалённой машине, изучает незнакомый пакет или готовит архитектурное решение, самый дорогой ресурс — не сам поисковый запрос, а потеря контекста. Логи остаются в терминале, фрагменты конфигурации — в репозитории, а найденные материалы — в браузере. Связка Claude Code и поискового MCP помогает удержать эту цепочку в одной рабочей сессии: агент получает задачу, формулирует запросы, возвращает ссылки и превращает результаты в проверяемые действия. Эта статья описывает инженерный подход к такой интеграции: как выбрать область конфигурации, как не раскрыть секрет, как проверять соединение и как добавить обычный API-вызов в скрипт автоматизации. Она опирается на руководство Ace Data Cloud для Claude Code и Google Search MCP. Сначала создайте приложение и API Token в консоли приложений ; обзор платформы доступен на русской странице Ace Data Cloud , а актуальные инструкции — в каталоге документации . Что именно добавляет MCP в рабочий процесс Claude Cod...

How to Use NanoBanana MCP in Claude Code for Image Editing Workflows

Image
When you are already working inside a terminal agent, the slow part of an image workflow is often not the model call itself. It is switching tools, copying URLs, explaining context again, and then bringing the result back into your project. NanoBanana MCP gives Claude Code a direct image generation and editing toolchain from the same terminal session, so you can ask for a visual asset, compose several source images, or clean up an image without leaving the coding flow. What you can do The NanoBanana MCP documentation describes a focused setup for Claude Code: connect the remote MCP server once, then call two tools from natural language: nanobanana_generate_image for text-to-image generation. nanobanana_edit_image for image editing, including workflows that pass multiple image URLs. That makes it useful for small but common builder tasks: creating a 404-page illustration, placing an object from one image into the scene of another image, or removing unwanted text from a sc...

How to Generate Veo Videos from Claude Desktop, VS Code, or Cursor with MCP

Image
Video generation becomes much easier to prototype when it sits inside the tools where builders already write prompts, inspect assets, and iterate on ideas. The Veo MCP Server from Ace Data Cloud lets an MCP-capable client such as Claude Desktop, VS Code, or Cursor call Veo video tools through a standard interface, so you can move from a written brief to a generated video task without leaving your working environment. What you can do The documented Veo MCP setup exposes a small, practical tool surface for video workflows: veo_text_to_video for generating video from text prompts. veo_image_to_video for generating video based on images. veo_get_1080p for upgrading an already generated video to 1080p. veo_get_task for checking a single task status. veo_get_tasks_batch for checking multiple task statuses. The same document also describes support for multiple Veo models, including veo3 , veo2 , and veo31-fast-ingredients , plus output options such as 4K, 1080p, GIF...

A Practical Guide to Pay-Per-Call AI APIs with the Ace Data Cloud SDK and X402

Image
If you are building a CLI, backend worker, or agent that calls AI APIs on behalf of different users, API keys are often the awkward part: who owns the balance, who rotates the token, and how do you charge only for one call? The paymentHandler hook in the Ace Data Cloud SDK gives you another path: let the first request receive 402 Payment Required , sign an X402 payment envelope locally, then retry the same API call with a PAYMENT-SIGNATURE header. What you can do The documented flow is useful when you want business code to stay close to a normal SDK call while payment happens per request. In practice, you can: Call client.openai.chat.completions.create(...) without passing an apiToken . Inject a TypeScript or Python paymentHandler that signs the payment challenge after a 402 . Use EVM networks such as base or skale , or use solana where the Solana signer path is supported. Choose preferScheme: 'upto' / prefer_scheme="upto" for metered chat-st...

How to Add Image Understanding to Claude Code with NanoBanana MCP

Image
Sometimes the slowest part of a developer workflow is not writing code. It is leaving the terminal to make a README illustration, clean up a screenshot, compose two product images, or generate a small visual for a 404 page. The NanoBanana MCP server gives Claude Code a practical way to do those image tasks from the same command-line session where you are already working. What you can do NanoBanana is useful when the image task depends on understanding what is inside one or more images, not just generating a pretty picture from text. According to the Ace Data Cloud documentation, NanoBanana can work with Google Gemini-style image understanding and supports both image generation and image editing through Claude Code. In practice, that means you can ask Claude Code to: generate an illustration from a text prompt with nanobanana_generate_image ; edit an existing image with nanobanana_edit_image ; pass multiple image URLs to nanobanana_edit_image and describe how the image...

MCP в Claude Code: надёжный workflow для ссылок и документации

В инженерной практике ссылки становятся частью интерфейса: они попадают в README, описания pull request, отчёты о релизе, тикеты и сообщения команде. Однако постоянные URL репозиториев, документации и страниц с параметрами быстро делают текст плохо читаемым. Вместо отдельного ручного шага для каждой ссылки можно включить сокращение ссылок в привычную цепочку работы агента в терминале: он анализирует изменения, подготавливает текст и вызывает нужный инструмент. Ниже — практический подход к организации такого процесса в Claude Code с Ace Data Cloud. Цель не в том, чтобы автоматизировать всё без контроля, а в том, чтобы получить воспроизводимый поток: хранить секреты безопасно, выбрать область действия конфигурации, проверять соединение и вызывать обработку одной или многих ссылок по явному заданию. Что даёт MCP в терминальной разработке Claude Code хорошо работает с кодовой базой, файлами и командами. MCP добавляет к этому внешние действия как вызываемые инструменты. Для задачи со с...

A Practical Guide to Building Version-Safe Music Workflows with the Suno Studio Projects API

Image
If you have ever tried to turn generated music into an editable product feature, you know the hard part is not only creating audio. The hard part is keeping project state, track edits, replacements, and final exports coordinated without overwriting someone else's work. The Suno Studio Projects API solves that problem with a single project-oriented endpoint. Instead of treating every music generation as a one-off file, you work with a project, retrieve its editable state, save complete versions, add or generate tracks, commit selected candidates, and render the final song when the mix is ready. What you can do The project API is centered on one endpoint: POST /suno/projects Authorization: Bearer YOUR_API_KEY Content-Type: application/json The operation is selected with an action field. The documented actions cover the full lifecycle of a multi-track project: create : create an empty project. retrieve : read the project and its complete editable state . save : save ...

Getting Started with Short Links in Gemini CLI Using ShortURL MCP

Image
Long URLs are awkward in chat, email, release notes, and terminal output. If your day already runs through Gemini CLI, the most practical fix is not another browser tab—it is a small MCP tool that can shorten links directly inside the conversation. This guide walks through connecting the ShortURL MCP server to Gemini CLI, what the configuration means, and how to use it in a repeatable builder workflow. The goal is simple: let the assistant shorten one link or a batch of links while you stay in the terminal. What you can do The ShortURL MCP integration documented by Ace Data Cloud exposes one focused capability: generating short links from long URLs. The documented tool is shorturl_create , described as a tool that shortens long links. That narrow scope is useful. Instead of asking a general-purpose agent to invent a link workflow, you give Gemini CLI a concrete MCP server with a concrete tool. In practice, that means you can ask for tasks such as: Shorten a single produc...

How to Build Image Generation and Editing Workflows with the Nano Banana Images API

Image
Most image workflows start simple—one prompt in, one image out—but real products quickly need more: editing an existing asset, combining multiple references, tracking task IDs, and receiving results without holding a connection open. The Nano Banana Images API is useful when you want both text-to-image generation and image editing behind a single HTTP endpoint. This guide walks through the practical shape of that workflow using only the fields and behavior documented for the API. What you can do The API supports two actions through the same endpoint, POST /nano-banana/images on the base URL https://api.acedata.cloud : action: "generate" creates images from a text prompt . action: "edit" edits one or more existing images supplied through image_urls , guided by a text prompt . The same request shape also gives you production-oriented hooks: count for requesting multiple images, model for choosing the image model, callback_url for asynchronous deliv...

How to Use NanoBanana MCP in Claude Code for Image Editing Workflows

Image
When a development workflow needs visual assets, the slowest part is often not the image model itself—it is leaving the terminal, uploading references, copying outputs back, and trying to keep the result aligned with the project you are already editing. NanoBanana MCP gives Claude Code a terminal-native way to generate images, edit existing images, and compose multiple image inputs through a remote MCP server. The useful part is not that it adds another image tool; it lets a coding agent call that tool while it still has your project context. What you can do The NanoBanana MCP document describes two exposed tools: nanobanana_generate_image for text-to-image generation. nanobanana_edit_image for image editing, including multiple image inputs. That makes it practical for everyday builder tasks: create a page illustration for a 404 screen, remove unwanted text from a screenshot, or combine an object from one image with the environment of another. The document specifically ...

Veo MCP в Claude Code: инженерный поток создания видео без выхода из терминала

Команда разработчиков нередко готовит короткие видеодемонстрации одновременно с кодом: для README, релизной заметки, тестового стенда или презентации функции. Узким местом здесь обычно становится не сам сценарий, а разрыв процесса: контекст проекта остаётся в терминале, а визуальный материал приходится вести отдельным потоком. MCP позволяет сделать работу более связной: агент читает задачу, вызывает инструмент генерации и возвращает результат в тот же рабочий диалог. Ниже — практический способ организовать такой поток с Claude Code, Veo и Ace Data Cloud. Это не инструкция по монтажу, а схема интеграции, в которой конфигурация, проверка доступа, формирование запроса и контроль результата остаются воспроизводимыми для команды. Что именно подключается Veo — модель Google для генерации видео. В задачах на ролики полезны текстовый сценарий, исходные изображения, продолжение клипа и получение версии 1080p. В вариантах veo3 и veo31 описание может включать звук: модель создаёт визуальную ...

How to Add X402 Pay-Per-Call Payments to the Ace Data Cloud SDK

Image
If you are building a CLI, agent, or developer tool, prepaid API tokens are not always the cleanest fit: sometimes you want the caller to pay for exactly the request they make, without your backend holding their platform token. What you can do The SDK + X402 Payment Hook flow lets an Ace Data Cloud SDK client make a normal SDK call, receive a 402 Payment Required challenge, sign a payment envelope locally, and retry the same request with a PAYMENT-SIGNATURE header. The business code still calls client.openai.chat.completions.create(...) ; the payment path sits underneath the SDK through a paymentHandler . This is useful when you want to build: agent tools that can pay per call from a wallet, CLI utilities where the operator owns the payment key, browser flows where a connected wallet signs the request, backend jobs that should avoid storing an Ace Data Cloud API token. The documented SDK path supports TypeScript through @acedatacloud/sdk and @acedatacloud/x402-c...