Claude Blog 采集 (2026-07-27)¶
共采集 2 篇文章
📋 文章索引¶
- Working with Claude Fable 5 in Claude Cowork - Jul 16, 2026 (评分: 9.5)
- Secure access to the Claude Platform with Workload Identity Federation - Jun 17, 2026 (评分: 9.0)
Working with Claude Fable 5 in Claude Cowork¶
来源: Claude Blog 发布日期: Jul 16, 2026 采集时间: 2026-07-27 价值评分: 9.5/10 正文字数: ~12656 字符
摘要¶
Anthropic's Claude Fable 5 can carry long, complex work on its own. Here's how to get the most out of it in Claude Cowork.
正文内容¶
Claude Fable 5 can carry long, complex work on its own. Here’s how to get the most out of it in Claude Cowork, starting with as little as an idea and getting to finished work.
Claude Fable 5 is Anthropic’s most capable generally available model , built for long-running, complex and asynchronous work. Claude Fable 5 is particularly effective carrying out multi-step workflows (such as conducting deep research that it incorporates in a first-draft memo, performing due diligence before generating board presentations, or going through a folder to redline multiple contracts, to name a few) on its own for extended periods of time, testing and evaluating its results as it goes. Maximizing the model’s capabilities requires a change in how you work with it. As models improve over time, we've refined our recommendations for getting more out of Claude, including prompting best practices , providing context, and building skills that capture repeatable processes like weekly team updates, sales call prep, or customer feedback analysis. Those practices still matter with Claude Fable 5, in fact, the model performs even better with them in place. Claude Fable 5 applies your context, preferences, and skills across entire tasks, even those that take days to complete, while previous models may have lost track over long stretches and needed reminding. Working with it resembles working with a highly capable colleague: you explain the situation, agree on what a strong final result looks like, and let your colleague work. Less of your time goes to checking each step, so more of it can go to deciding what the work should be. Our prompting guide for Claude Fable 5 provides a detailed list of capability improvements and recommended behavior and prompting changes. In this article, we cover how some of those apply to Claude Cowork, Anthropic’s agentic AI system for knowledge work. How Claude Fable 5 complements Claude Cowork Claude Cowork is built for creating finished work. Give it an objective, and it manages the rest, even when the task is large and complex. A big job gets broken into parts that run at the same time, each with its own subagent, a separate instance of Claude that takes one part of the job to complete and reports back. That is how Claude gets through large amounts of information quickly. Claude also manages large tasks by setting a plan at the start, and it can check its own results against it as it goes. Claude Fable 5 has a wide lead over our other models on long, complex tasks and Claude Cowork tasks are often exactly that, with dozens of steps, each building on the last. Take a project like building next year’s budget based on this year’s actuals. Claude gets access to and reads the spreadsheets, pulls the run rates, projects each line, reconciles the projections against targets, and writes the summary. If it misreads a run rate early, the error will flow into each projection after it. Claude Fable 5 plans the workflow before starting and checks results as it goes, so it can catch the misread number while the job runs and correct it. Decide when to use Claude Fable 5 Claude Fable 5 isn’t the default model in Claude Cowork; you need to select it. As of the time of publication, the default is Claude Sonnet 5, and it is the right choice for everyday tasks you yourself would handle in quick passes. Claude Opus is a dependable choice for deep work with a clear shape, where you know what the end result looks like. Claude Fable 5 is for the projects that feel the most complex or ambiguous, and may have been out of reach for prior models. It spends more time thinking and more of your usage limits, but that may be worth it on work that’s time-consuming or costly to get wrong. We recommend that you reserve Claude Fable 5 for your most important work, especially jobs that use multiple tools and require a series of judgment calls. You can further tune your choice with Claude’s effort setting. At higher effort, Claude Fable 5 plans more before it kicks off a job and checks in more throughout its run. Keep effort higher for complex or multi-step projects you expect Claude to complete from beginning to end. At lower effort, you’ll get a faster response, while still taking advantage of Claude Fable 5 intelligence. Consider using it for tasks that still need frontier judgement, but not deep explorations, such as agentic runs made of many easy steps or work where the result is easy for Claude to check. In our testing, Claude Fable 5 at lower effort often matched or exceeded the performance of earlier models at their highest effort levels. For a more detailed look under the hood, we explain how model choice and effort interact in Claude Code . It’s also worth noting that Claude Fable 5 comes with a new set of classifiers: separate AI systems that detect potential misuse in requests related to cybersecurity or to biology and chemistry. When they trigger, the response is automatically handled by Claude Opus 4.8 instead, and users are informed whenever this occurs. Opus 4.8 is a highly capable model in its own right, and the chat stays on Opus from there; start a new one to get back to Claude Fable 5. We tuned these safeguards conservatively so we could release a Mythos-class model for general use both safely and quickly, so they'll sometimes catch harmless requests, including phrases in Claude Cowork that only touch on related topics. We're working to reduce these false positives as we refine the safeguards. Start with as little as an idea When you kick off a task, you don't always know what you're trying to accomplish. That early stretch is where Claude Fable 5 can be a powerful thought partner, and in Claude Cowork it works from your skills, tools, and knowledge. For example, if you see a news announcement relevant to your industry or company, you can ask Claude: "Here's the announcement, tell me what changes for us, considering everything you know about my organization." You read the output and as you start to develop a strategy, Claude builds a plan to produce the assets that you need, starting in the same conversation in Claude Cowork where the brainstorming happened. Brainstorming in Claude Cowork gives the model your real material to think with: it can read the files you've shared and use the tools you've connected while you talk, so it can offer relevant suggestions. It notices gaps in an idea while the idea is still easy to change, and it shows you directions you hadn't considered. And when the task begins in the same conversation, Claude Fable 5 already carries the goal you settled on, the constraints you named, and the decisions you made along the way. For example, a data scientist at Anthropic came to Claude Cowork with an idea for a new analytics dashboard while the team was still figuring out what it should show. Because Claude Fable 5 could read the team's usage data during the conversation, it knew which problems take weeks to get noticed, and it ranked the metrics that would have caught them sooner. By the end of the conversation, the data scientist had a shortlist of metrics worth adding and a clickable prototype. Here are two prompts that can help you evolve your idea into an output: Ask Claude to interview you: “Before you start, ask me everything you need to know to get this right.” Ask for directions: “Here is roughly what I want. Give me three ways you could take it, with a quick sample of each.” Provide context with your constraints Context is the information Claude works from, and in Claude Cowork that means your prompt, whatever files and folders you’ve shared, and any tools you’ve connected, such as Asana, Hubspot, Jira, Slack, or others . When you give Claude Fable 5 a task in Claude Cowork, think of how you'd brief a colleague on a report. You wouldn't hand them a list of rules, but chances are you'd tell them who it's for, when it's needed, and what it has to accomplish, and trust them to make good decisions from there. Claude Fable 5 works the same way. Constraints are still useful: "keep it under two pages and use plain language" is a fine instruction. But a constraint only tells Claude what not to do. Context tells it what the work is for, so it can make the right call in situations your constraints didn't anticipate. Claude Fable 5 handles long, multi-step tasks well partly because of this context: when a question or decision point comes up while it works, it looks for the answer in the context you shared. It also uses that context to check its own work as it goes, so give it something concrete to judge against, such as an early draft of a report and the final version, and Claude will work out what your standards are based on what changed between the two. Context also gives Claude an understanding of your situation, so it can infer or find details you never specified in the prompt. Sharing additional context lets Claude search through a wider base of knowledge than you could fit in a prompt. One thing to note about chats with lots of context: in order to stay caught up, Claude reads the whole conversation again with every new message you send, so a long conversation may use more of your usage. It helps to start new tasks in a fresh conversation. Scheduled tasks count towards your limit too, so check yours occasionally and turn off any you no longer need. Delegate larger, more complex jobs to Claude Fable 5 You may be used to breaking a task into parts and prompting Claude for each one. Claude Fable 5 needs far fewer of those intermediate prompts, so you can delegate complete jobs in Claude Cowork. You should still write out the steps yourself when the process matters to you, for example if a metric has to be calculated the exact same way each time. But if the outcome is what matters, describe the goal and let Claude do the rest. A step-by-step prompt can also limit it to the steps you thought of; a goal gives it room to find a better path. Delegating in Claude Cowork means handing Claude a decision you would normally make yourself. Here are three ways you can do that: Delegate the approach: Give Claude the material and describe the outcome you want, for example, "Here is last quarter's customer feedback. Find out why cancellations rose and what we should change." There may be several reasonable ways through that folder, and the right one depends on what's in the feedback. That's a choice Claude Fable 5 makes well: it reads everything, picks an approach, and checks its conclusion against the feedback before bringing it to you. You can judge the answer without needing to specify the details of how it gets there. Delegate the procedure: A skill teaches Claude a procedure your team uses—how you build a report, format a deck, run an analysis. You don't need to say which skills to use or in what order. Say "put together the quarterly review the way we always do it," and Claude Fable 5 picks the right skills at the right moment. Delegate the timing: For work you want repeated, describe the outcome and Claude will set up the schedule and turn it into a recurring task: "I want to start every Monday knowing what changed in the pipeline and what needs a decision." Bring Claude Fable 5 harder work than you're used to giving AI, even work you assumed wasn't possible. Even if the work is messy or unclear, or could take hours or days to finish. Describe it and see whether the model can work at that level. Pick an outcome you're responsible for and say how you'd judge it, then let Claude propose the steps to get there. Some of this work may turn into a scheduled task that runs on its own. Claude can save the procedure as a skill and put it on a schedule, working through the tools you're already connected to. Review Claude’s thought process Part of what lets Claude Fable 5 carry long work is that it knows how to set and follow a plan well. In Claude Cowork, you can see that plan while Claude works: the panel beside the conversation lists what it intends to do, then the files it's reading and writing and the tools and skills it's using. That panel is your chance to catch problems and redirect early. A mistake you'd otherwise find in the finished output instead shows up as one wrong step in the plan. You can correct the plan in one sentence and Claude adjusts without starting over.
Explore more product news and best practices for teams building with Claude.
Transform how your organization operates with Claude
Product updates, how-tos, community spotlights, and more. Delivered monthly to your inbox.
Please provide your email address if you'd like to receive our monthly developer newsletter. You can unsubscribe at any time.
采集自 Claude Blog，由 collect_claude_blog.py 自动采集
Secure access to the Claude Platform with Workload Identity Federation¶
来源: Claude Blog 发布日期: Jun 17, 2026 采集时间: 2026-07-27 价值评分: 9.0/10 正文字数: ~2387 字符
摘要¶
WIF replaces static API keys with short-lived, scoped credentials issued at request time. WIF is compatible with any OIDC-compliant identity provider and covers all Claude API endpoints
正文内容¶
Workload Identity Federation (WIF) is now generally available on the Claude Platform. WIF is compatible with any OIDC-compliant identity provider and covers all Claude API endpoints, including when accessing the endpoints through our first-party SDKs and Claude Code. With WIF for workloads and ant auth login for interactive sessions, developers never have to handle a static API key when building with the Claude Platform. How Workload Identity Federation works WIF replaces static API keys with short-lived, scoped credentials issued at request time. Whether you're a two-person startup running GitHub Actions or an enterprise with detailed credential policies, you can now authenticate with the Claude Platform the same way you authenticate with the rest of your stack. With WIF, there are no static Anthropic credentials to create, rotate, or leak. Workloads authenticate with the identity they already have: an AWS IAM role, a GCP or Kubernetes service account, an Azure managed identity, a GitHub Actions token, Okta, or other OIDC-compliant providers. We're also introducing service accounts to the Claude Platform, so each workload can have its own identity, roles, and audit trail instead of a shared API key. First, a federation rule binds an external identity to a service account. Then, when a workload requests access, the Claude Platform verifies the workload's signed OIDC token, matches its claims against your federation rules, and issues a short-lived access token bounded by the service account's roles. Every exchange and request is recorded against that service account in your audit logs. Set up your first workload in minutes The Claude Console has a guided setup flow for configuring workload identities. The setup validates each step and finishes with a test command that confirms your workload can authenticate.
Getting started API keys work alongside WIF, so you can migrate one workload at a time. Read the setup guides for each identity provider, or open the Claude Console to connect your first workload.
Explore more product news and best practices for teams building with Claude.
Transform how your organization operates with Claude
Product updates, how-tos, community spotlights, and more. Delivered monthly to your inbox.
Please provide your email address if you'd like to receive our monthly developer newsletter. You can unsubscribe at any time.
采集自 Claude Blog，由 collect_claude_blog.py 自动采集