Code with Claude Tokyo 2026: Opening Keynote

Heat. Heat. Heat. Heat. Heat up here. Hey, hey, hey. Please welcome to the stage head of engineering for the cloud platform at Enthropic, Caitlyn Les. Sto anthropic engineering Caitlyn le Tokyo. This is the first time that we've brought Code with Claw to Japan and we're grateful to spend the next couple of days with all of you. This morning we'll be talking about our models, our platform, and our products. But before we get into it, I want to start by sharing that just a few hours ago, we released the fifth generation of Claude models, Claude Mythos 5 and Claude Fable 5. These are our two most capable models ever. Diane, our head of product for research, will join us shortly to share a lot more about why these models are so special. I lead engineering for the Claude platform. The platform gives developers the tools they need to build systems on top of Claude to harness its intelligence. This is the highest leverage way for us to help solve the world's most important problems. And this is why Anthropic is a platform company. developers all over the world, many in this room today, produce far more value on top of the platform than we could ever build on our own. So, let's start with what I'm seeing from our customers lately. There's an incredible volume of powerful application shipping right now, and a lot of it is coming from right here in Japan. Rakutin is one of our favorite customers to work with. Their team went from using cloud code to accelerate development to building on cloud managed agents to power custom internal agents across engineering, product, sales, and finance. One of their product managers coordinates teams of agents exactly the same way a leader manages teams of humans. Now, they're shipping major releases every two weeks instead of only once per quarter. And it's the same across Asia-Pacific. Another great customer is Canva, the Australian design platform that hundreds of millions of people use. Most of their users have never written a line of code. But with the help of Plaude, Canva code changes all of this. Within a design, you can just ask for something interactive like a map, a calculator, or a widget. And Claude built it into a working mini app ready to drop right into your page. All around the world, people are building new systems and applications powered by Claude that nobody could build before. And we're all feeling this shift. The landscape is changing faster than ever. Things that weren't possible yesterday have become possible today. We're on a mission to keep raising the ceiling of what's possible to build by making models that are increasingly capable. A couple of years ago, the frontier of model development was just the ability to draft a really simple commit message. And then one year ago, we were standing on stage at our first ever code with Claude event. Opus 4 was the headline. And it was mind-blowing at the time that Claude could build an entire feature on its own. And then six months ago, agents were able to run overnight to complete longrunning and autonomous tasks. Two months ago, Mythos read the entire OpenBSD source tree and found a 27year-old vulnerability that had slipped past human reviews and static analysis for almost three decades. And earlier today, we released Mythos 5 and Fable 5. Fable 5 is state-of-the-art on nearly all tested benchmarks of AI capability, showing exceptional performance in software engineering, knowledge work, scientific research, vision, and more. These jumps keep getting bigger, but the time intervals keep getting shorter. But even though these model capabilities are improving on this exponential, most business capabilities are still on a linear. So there's this growing gap between what AI can do and what it's actually doing for people. Closing that gap is our collective opportunity. And that is why we've built the cloud platform. It gives developers the tools they need to build more powerful agents. Year-over-year API volume is up nearly 17 times on the platform. We've been shipping a lot and there's a lot more on the way. Over the next couple of days, we'll share where we're headed and we'll learn more about what you're building. Our goal is to make our models, our platform, and our products work better for all of you. So this morning, we'll dive into each. First, Diane will talk about our models. She'll share more about our latest frontier models, Mythos 5 and Fable 5, and what's coming next. Next, Angela and I will walk you through how you can build and deploy agents at scale on the Cloud Platform using Cloud Managed agents. And today, we're excited to share that we're shipping brand new cloud managed agents features. You can now schedule deployments to have your agents run on whatever cadence that you need. And you can store environment variables in vaults so your agents can securely make authenticated API requests without giving them access to the keys. Finally, Cat will walk you through the latest in Cloud Code where the average developer is now spending 20 hours per week with Claude. She'll cover the latest features like the new agent view and dynamic workflows. And all of this comes back to you and what you're going to build. Because most people will never integrate with the Claude API themselves. They'll never open a terminal and type Claude. They'll experience AI through something that one of you built on the Claude platform like a saleserson walking into a high stakes meeting fully briefed by Slack agents or a lawyer getting a brief out the door faster than ever with Lagora. or a developer using any one of the world's best coding agents. This is why we're a platform company. Every day I'm amazed by the collective impact that you're all having building solutions to the world's most important problems. So, thank you for being here, for partnering with us, and for showing us what's possible. Aratu goas. Please welcome Diane from our research product team for your first deep look at our newest models. product management for research Diane Penn sto anthropic research product management Diane Penn sama Hi, good morning Tokyo. I'm Diane and I joined Anthropic in 2023 and I've been a part of every version of Claude since Claude 2. For those of you who are counting, that's bringing 21 versions of Claude across Haiku, Sonnet, Opus, and now Fable, and Mythos to end users and developers like you. Our most recent launch happened just a few hours ago. We released Claude Fable 5 and Claude Mythos 5, the first generation of our fifth models. Fable 5 is the most capable model we've ever made generally available. It's based on the same foundations as my 5. These models are already accelerating our work at Enthropic. And for developers like you, Claude will be able to be stronger and your starting line will move forward. We talk a lot about the exponential at Enthropic, but what does this actually mean? For us, as model intelligence increase, we believe the value of use cases it creates increases exponentially. For example, agented coding that many of you are using today is far more valuable than simple autocomplete from just a few years ago. And we are already seeing this with Fable 5. Let's start with coding because we believe that's where most of you will feel the magic first. Fable 5 performs the highest on Sweepbench Pro. But the benchmark itself underscells the story. The longer and the more complicated and sophisticated the task, the farther the gap between Fable and every other model out there. Two things drive that gap. The first singleshot correctness. Give Fable 5 a complex wellsp specified problem and it will nail it on the first pass. Early testers have told us about single prompts that essentially create work that would have taken days or even weeks for a group of teams. The second is long horizon autonomy. Fable 5 can run for days on a single goal and stay coherent the entire way. It'll remember your specifications even on tasks that span millions of tokens. And it can dispatch sub agents, keep them on track far more dependably and with more cost consciousness than any other model we shipped. One more thing, Fable 5 isn't just good at writing code, it's even better at reading it. It's better at tree outing outages and better at digging through your repo history to figure out what's broken when, and also to proactively surface suggestions for making it better. Fable 5 is just as much of a step outside of coding. Start with a work that your organizations and companies actually run on. Whether that's financial analysis, documents, slides, spreadsheets, Fable 5 manages that work end to end. It'll stay follow instructions, stay on scope, and what it creates back for you will be professional grade because this is a model that we built for workflows at scale. It's also better where requests aren't clean. What we've seen is that when you give Fable 5 something messy, multi-threaded, and ask it to figure out what to do next, it could do that work and outshine other versions of Claude. Fable 5's version is also the best in the industry. It can read dense technical images, web applications, plots, diagrams, charts far more accurately than any other version of claude we've shipped before. Now, everything I've talked about is about the upside of a model of this capability. But intelligence of this magnitude cuts both ways. Two months ago, we launched Project Glass Wing and made Mythos preview available to a small group of partners because its capabilities in cyber security were strong enough to be potentially misused. Since then, we built a new safeguard system and that safeguard system allows that same intelligence to be shipped and created and used by everyone with Fable 5. This is a little of how it works. When a request first touches cyber security, biology or chemistry topics, Fable will route it to our next most capable model, Opus 4.8. The response is clearly labeled and your charged opus prices. We know this is not perfect yet. researchers doing legitimate work in these fields will sometimes hit a block and reroute. And we're continuing to work on that. But that's the type of trade-off that allows our most capable models to be in everyone's hands today as of this morning instead of months from now. And our customers are already finding value with Fable. Rakutin found that the highest effort levels Fable 5 reflects validates his work and for them that makes autonomous automated operations possible at scale. Cognition ran Fable 5 against their frontier coding eval Crunchier bench and it scored the highest of any model they've tested. They highlighted the long horizon reasoning and how it generalizes to unfamiliar tools, complex context right out of the box. And Jensen Spark told us that Fable 5 came out number one in their evals, winning head-to-head against every version of model they've tested. It's actually the strongest on the hardest tasks such as UI design and game coding. So what about Mythos 5? It's the same underlying model as Fable 5, but with the cyber and bios safeguards lifted. For Mythos preview, we showed what this model class can do, and Mythos 5 is a step beyond that. It's available today for our CL glasswing partners as part of Project Glasswing. Later this month, we'll begin enrolling more researchers from life sciences to access Mythos 5 because those same capabilities that make biology risky are actually the same ones that have the highest impact for real good with AI. So, as developers, what can you do with Fable 5? One metric I like to look at to make sense of all of this change is time horizon. which is how long a model can work autonomously before losing coherence on what to do next. With Fable 5, we've seen agents that are proactive and know what to do without being told. These agents can be responsible for higher level goals, responsibilities that require judgment, collaboration, and proactive taste. For example, instead of asking Claude to write a project update, you can now ask Fable to make sure the project stays on track for the whole week. Instead of asking Claude to produce a financial forecast, you can tell Fable to own, update, iterate, and improve on the forecast to keep it accurate. As you've seen today, the exponential keeps moving. We need to build for emerging capabilities, not just for today's current models. We expect future versions of Claude to be more capable than the one that even we are releasing today. And Fable 5 and Mythos 5 are just a glimpse into that future. So, as developers, how do you take advantage of all of this change ahead? First, you should design for the next version of Claude, not just the current one. What we've seen is that countless times, the developers who win are the ones whose architectures, harnesses, and product experiences are ready to absorb this next jump in intelligence. Claude is intelligent and resourceful. And as models become more intelligent, they can actually dig farther with more basic primitives such as a file system or sandbox computing environments and far bless with sophisticated or too complex harnesses. As models get more intelligent, you will also need to start making harder eval prototypes for experiences that may not work yet. This is how you actually know that the exponential is moving underneath you. When a task or a prototype or a product experience that wasn't quite always working well starts working, that's a signal to ship something magical that you couldn't have done before. And finally, as the pace continues to accelerate, the teams who win are the ones that get the most out of quad with model upgrades and treat them as business opportunities. You should make model upgrades easy. This means things like automated evals, testing processes, and making sure that you stay hands-on testing, pushing, creating new things with new versions of Claude. And this is how you'll know new capabilities will be enough to deliver new experiences for your customers. We're seeing the exponential continue, which means that Claude will get smarter and be able to pick up new capabilities for at scale. And you as developers are some of the first people to feel that. You're the ones who actually can experiment, build new products, and are the first to find opportunities for new markets that others don't see. And we can't wait to see what you'll build next with Fable 5. And now, Angela and Caitlyn are going to show you a bit of how the Claude platform can make this a reality come to life. Thank you so much. product for the Claude platform, Angela Jen. Stage Angela Jang. Last night, while we were all asleep, a product somewhere noticed that it was broken. This product read its own error reports. It found the bug and it wrote the fix and rolled it out to all the users. By the time the team actually woke up, the problem was gone and the change log was already written. Now, in this situation, there was no standup. There was no ticket. And this is what a native AI company looks like. This is not a company where people just use AI to do their work. But instead, it's a company where work itself runs on the substrate of AI and people decide what the outcome should be. Now, this story can be a reality if you have the right ingredients. And there's three things to that. The first is the harness, the second is the context, and the third is the infrastructure. These three are what turn raw intelligence into true business outcomes. And that's exactly what the Claude platform provides. We give you agentic harnesses, context management tools, and production-grade infrastructure that allows you to become truly AI native. Recently, we packaged all these together through Claude manage agents, our new product offering. Fable 5 is the best model for building longunning agents, and manage agents is purpose-built for Claude. This means when you put these two together, you're able to actually get better agent outcomes with significantly less effort. Okay, let's talk about the harness. So, the harness is what gives claude models the ability to actually do work. It includes tools, an environment, and the permission to act. With a harness, the model doesn't just tell you about the fix. It actually makes it happen. You don't want AI that just gives you helpful suggestions. You want AI that can actually do the work. With cloud manage agents, we offer an agentic harness that separates the brain from the hands. The brain decides what to do and then we spin up sandboxes or hands to execute the work as needed. Our harness also has an iterative based ability to operate on an outcome. You specify what that outcome is and then a managed agent is able to iterate until it actually achieves it for you. Second, let's talk about context. Now, a model is only as good as the context that you actually give it. We offer a 1 million context window so that you can have agents that consume high amounts of content without any degradation. But that's not all. We also offer our agents memory so they can actually remember what they've been doing. We also give our agents the ability to read and write skills of their own so they can fill in the knowledge gaps that they're missing. And lastly, we give our agents the ability to dream so they can inspect over their own previous trajectories and identify how to self-improve. And lastly, let's talk about infrastructure. Now, if you build really longunning autonomous agents, this requires extreme scale and reliability. This is one of the hardest parts to get right. Cloud manage agents automatically spins up and down sandboxes and has the ability to generate multiple agentic fleets when needed to help you optimize for reliable and persistent agents that just get the work done without all the hard work. Building self-improving companies is something that can happen today. And we've worked with so many companies that have built agentic systems on cloud managed agents. And they've been able to do that 10 times faster because they didn't have to roll their own harness or manage their own context or honestly build up all the infrastructure that they need. Notion actually used manage agents to power agent orchestration directly within their product. As a result, their users can delegate complex longunning work to claw directly inside their workspace. And Notion isn't alone. ASA used managed agents to build AI teammates. These are collaborative AI agents that work alongside humans inside Asa projects. And these agents can take on tasks and complete deliverables. Now, I'd love to welcome Caitlyn back to the stage to show what an AI native company can look like if you build on Cloud Manage agents. Back in February, Claude became the official thinking partner of the Atlassian Williams Formula 1 racing team. Competing in F1 racing requires a great driver who knows the track really well. But it also takes a team of engineers and researchers to build a rocket ship on four wheels that can go almost 400 km hour. To show you an example of Claude managed agents in action, we worked with a fictional racing team called Shankiro Racing. Um, and we helped them build a dashboard to analyze their car. Let's see it in action. Here we have the dashboard that we built for the Shankiro racing team. And you can see on the side we've got four research projects. Each of these research projects is backed by a claw managed agent. So we have aerodynamics, we have tire temperature, we have our power unit, and we have driver safety. And the goal for each of these projects is for our agents to be able to figure out what needs to change in order to better optimize our car. And in this we've used the cloud managed agents feature called outcomes which means that when an agent is doing its work we can use a rubric to define what good looks like and the agent can iterate until it does its work incredibly well. So we could go ahead and kick off any one of these agents and that will start a run of the agent in this case the driver safety. And what's cool about this dashboard being fully powered by AI is we can go ahead and ask a question like how much does a car currently weigh? and how is this weight distributed and the dashboard will go ahead and refresh with this information. So now we know our 800 kgs are approximately broken down between our power unit, our battery and other components. So if I were a developer working to build this dashboard, what I would use is the claw developer console as my home base for developer productivity, observability, and all the tooling that I need in order to create these managed agents. And so you can see we have all of our sessions from our past runs of our agents. And if we check one of these out, we can open it up and see we get really rich observability for everything the agent is doing. We can see its responses. We can see its tool calls and ultimately when it finished its work. Starting today, you can even schedule these agents to run at any cadence that you want. So we can come in here and click create deployment. And maybe we want to run a nightly safety check. And we'll do this using our safety agent. And so here we can decide when should the agent trigger. And we can choose a schedule such as every morning at 9:00 a.m. But it's not enough for us to have agents that are really easy to build and to run. What's also really important is for these agents to be really powerful. And so we've done this in a couple of ways like Angela mentioned earlier. First, we have memory. With memory, our agents can choose as they're working to write their findings or write their learnings to a file system. Then when the agents run in the future, they can look back into this file system on things like decisions or findings in order to do a better job the next time around. On top of this, we've built a feature called dreaming. And so if memory is real time deciding to write to a file system, dreaming allows the agent to look back over all of its past sessions and update its memory and its skills so that it does even better next time around. So if we come back to Shankiro's dashboard, we can hit this button that says dream. And this will kick off a session for a new agent to look back on all of the sessions for the old agents and write its findings to memory and to skills. So everything you just saw like outcomes, schedule deployments and dreaming are available today on the cloud platform. Developers everywhere can start building with Fable 5 and managed agents today. Now Cat will talk about how cloud code is making it even more fun to ship as a developer. product for Claude Code, Cat Woo. code. Caitlyn and Angela just showed you how to build production agents on the Cloud Platform. With Cloud Code, we're bring that same leverage to your work as a developer. Not agents you ship to customers, but agents that ship code for you. First, I want to thank all of the developers in the room today here and watching online. Thank you for trusting Quad Code back when sonnet 3.7 was our frontier model and when our product was rough around the edges. Your support is what makes the team so excited to come in every day and make the product even Let's back up to why Quad Code exists. The mission of Quad Code is to bridge the difference between an idea and a shipped product. The way that we enable this is we build tools that elicit the frontier intelligence from our models. And we make these tools accessible to every builder. And we we don't think of ourselves as having a finished road map to share with you. We think of ourselves more like mountaineers climbing alongside you in terrain that none of us has fully mapped yet, figuring out what works together as we go. And we're growing with you with increasing AI capabilities and helping you navigate new challenges that emerge. I remember just last year I would give Cloud Code a task and I would review every single edit that it made, giving it detailed instructions about what it should do instead, walking through on every little detail of these simple tasks. Now, many of us are using auto mode to delegate permissions to Claude and only checking in after Cloud Code has already tested its changes and put up a PR that's ready for our review. Cloud Code started in the CLI and this is still the place for Power users who want a minimal text interface and the most control and customizations. Then we added the IDE for users who want the same powerful agents but want to follow along with all of the code changes. Then we heard from many of you that you're running multiple quad codes in parallel which you've affectionately called multi-clotting. And we've added two new interfaces to help make that easier. The one that I use most frequently is Claude Code on Cloud Desktop. It's designed for people who want this full screen graphical interface with built-in previews, a sidebar control plane, and the ability to render images and rich outputs. We've built desktop to be a single view across all your local and cloud sessions with visual indicators of which agents are working and which ones need your input. Next is our newest surface, Claude agents view in the CLI. For those of you who want a control plane without having to ever leave the terminal, you can see what's waiting on you, what's running, and what's already done. You can reply in line to unblock or jump in and out of any session without losing your place. The VS code IDE extension and the cloud code on cloud desktop app are built on the cloud agent SDK. The same one that many of you are building on today. Many enterprises have now adopted these cloud code tools wallto-wall. At Enthropic, engineers on average ship 8x more code than they did in past years, even as the size of our engineering team has grown substantially. Together with you all, we're excited to discover and redefine what the future of engineering looks like by embracing new challenges that come and by building automations powered by Quad to attack each one. Here's some of the feedback that we've heard from our users and the products that we've built to address each one with the help of folks in this community. We heard from you that you want to spend less time on code review. So, we shipped a code review product that deploys a team of agents to catch critical bugs in your PRs. Thousands of companies use this every day and this includes all internal anthropic teams. We heard from you that you really want to code on the go. So, we launched remote control and quad code on iOS and Android. Now, you're no longer walking around with this laptop that's half open or stuck at your desk. You can now go to a park, touch grass, and still get your coding tasks done. We and we heard from you that you want to run cloud code on new tickets and so we built routines. You can configure this once and it'll run on a schedule, a web hook or an API call to kick off cloud code on the right task. So the work that used to require a human to manually kick off, routines can take care of for you. and we heard from you that you're landing so much code that your security teams are having a hard time keeping up. So, we built cloud security. It scans your codebase overnight. It flags a range of vulnerabilities, including the ones that it believes are most critical, and lets you kick off a cloud code session to tackle each one. And then finally, we heard that you want better ways to run ambitious large-scale tasks across your codebase like major refactors and migrations. So we launched dynamic workflows. This lets you kick off cloud code to run in parallel across tens or hundreds of agents in a deterministic structure to get your most ambitious tasks done. Each of these primitives composed together so that we can more easily adapt to the future of engineering. Everything I've covered is something that an individual developer can pick up today. But it's especially exciting to see how a range of companies are adopting this at the scale of entire engineering orgs. For instance, Spotify uses Quad to migrate thousands of repos. The team built a background agent on the quad agent SDK that reads a migration plan in plain English and then kicks off a fleet of agents that opens PRs. They're merging over a thousand PRs a month into production and they've cut migration time down by 90%. Another example is Merkari, Japan's consumer toconsumer marketplace. At Merkari, the entire engineering team runs on quad code and they've measured that engineering output is up 90% year-over-year using the tool. We see this across the industry. Millions of developers are getting more done and at higher quality than before. Now, let's see what this actually looks like in practice. I'm excited to show you one of my favorite new features in Cloud Code. For this demo, let's imagine we're Kaizen operations and we create apps for engineering teams. Right now, our marketing website is in English only, but we want to launch to 13 additional markets. So, we need to localize our site before the launch. Let's start with how we would normally do this with one quad code agent. We'll run this prompt asking Quad to convert our website to Japanese and we'll speed things up as it works. Quad explores a codebase, creates a language picker for the website and translates all the text to Japanese. So quads finishes work and let's check out the results. Here we can see a full marketing site and view the new Japanese translation from Quad. It took Quad about three minutes to complete this one translation. And we have 12 more to go. If we did this one at a time sequentially, it would take us almost an hour and lots of manual prompting with Claude. So instead, we'll use a dynamic workflow which will let Claude create a repeatable process and run each of the new translations at the same time in parallel. We'll prompt Quad to use a workflow for our translation to these 12 new languages. Quad creates a workflow and once it's ready, we can open it in the side pane and watch as it works. This is a powerful feature of the call desktop app. You can see all 12 translation agents running at the same time. After the translation agents complete, it's going to create 12 more agents to verify its work across these new This work would have previously required us to run 12 separate tasks and it can now be done in just one prompt. And we can also save this workflow as JavaScript code and reuse it for future translations. Let's check out the output. Quad has now created 12 new versions of the website, all localized with one action. We can see them all in the drop down and check out some of the different languages to see its work. This is just one example of dynamic workflows. We can use them for large-scale migrations, codebased audits, or doing performance optimizations. Any large job that requires running many agents at once in its terministic structure. Everything you just saw is available today, including dynamic workflows and cloud code in the Cloud Desktop app. And our newest model, Quad Fable 5, is available to all Quad Code users. So you can use the latest intelligence wherever you use Quad Code. We're so excited for you to try out these new features and for you to let us know what you think. We hope they continue to help you close the gap between idea and a shift product. And this is really what every talk today was pointing at. Dian's capability curve, Angelo's agents that run on infrastructure that you control, and what I just showed you. These are three layers to one story. The remaining gap is just how fast we can put these great capabilities to work for us. I encourage you to spend the rest of today exploring these layers. Join research talks if you want to learn more about the latest model capabilities. Join cloud platform sessions if you're building your own agents for your end users or join quad code workshops if you want to learn more ways to bring quad code into your day-to-day development work. All of this runs on Quad Fable 5, the best model we've ever shipped for aentic work and it's live today. Thank you all and enjoy Code with Claude.