Activity
113K followers
-
Mike Krieger shared thisWe just shipped Claude Sonnet 5.5! Sonnet 5.5 is the model to help you get more work done, faster. It runs 30%+ faster than Sonnet 5 and typically needs far fewer tokens, so most work costs up to 30% less A good complement to the more complex work done by Opus 5.5, Sonnet 5.5 is strong at scoped everyday work tasks, fixing bugs, and has a great eye for design. Haiku 5.5 will round out the 5.5 family in the coming weeks! More in the blog: https://lnkd.in/gS7AU4RB
-
Mike Krieger shared thisOpus 5.5 is out today! It performs at the level of Claude Fable 5.1 on most work, leads in coding and knowledge work, and it’s a really great writer. It performs at the level of Claude Fable 5.1 for most tasks, and costs 40% less to run than Opus 5. Opus 5.5 is priced at $4 per million input tokens and $20 per million output tokens, and we’ve reduced cache reads by 60% for usage billed by tokens. It also uses fewer tokens per task, so work runs about 40% cheaper than on Opus 5, and generates output more than 30% faster. We’re also increasing 5-hour session limits by 20% on Pro, Max, and Team plans today. Opus 5.5 is built to be a daily driver for complex coding and agentic work. The team at GitHub shared: "Developers want agents that can take on real software work and finish it. In our testing across GitHub Copilot CLI and VS Code, Claude Opus 5.5 used among the fewest tokens and steps we measured. In VS Code, it solved more terminal tasks than Opus 5 in less than half the steps. More than making individual tasks more efficient, it's making developers' bigger projects more achievable." Safety continues to be a huge priority for us. This is the first model to launch since we called for pacing the frontier. Pacing means keeping our safety practices ahead of model capabilities. In internal alignment testing, Opus 5.5 achieved the best scores of any model to date, and is the first Opus model to launch with a similar class of safeguards to Fable 5.1. It was also tested before release by external evaluators including Frontier Design and METR. Opus 5.5 is available today in the Claude apps, Claude Code, and on the Claude Platform, including through AWS, Google Cloud, and Microsoft Azure. And more good things on the way! Claude Sonnet 5.5 and Haiku 5.5 are launching in the coming weeks. These models will come with many of the same improvements to performance, efficiency, and safety. Read more about Opus 5.5 here: https://lnkd.in/gewgKTvFMike Krieger shared thisWe’re introducing Claude Opus 5.5, the first model in our new Claude 5.5 family. It performs at the level of Claude Fable 5.1 for most tasks, and costs 40% less to run than Opus 5. Opus 5.5 is our first model since we called for pacing the frontier. As with previous models, it was tested by external evaluators before release, including METR and Frontier Design. On our most comprehensive alignment test, it achieves the strongest score to date. It outperforms both Opus 5 and Fable 5.1 on nearly every benchmark we report, leading on agentic coding and real-world knowledge work. It’s also more efficient. Opus 5.5 uses fewer tokens per task at a lower price per token, and generates output more than 30% faster than Opus 5. Opus 5.5 communicates more naturally, addressing some of the most common feedback we heard on Opus 5. It puts the most important information up front and follows the writing rules you give it. Claude Opus 5.5 is available today. We're also raising five-hour limits on Pro, Max, and Team plans, so there’s more room to explore what it can do. Read more: https://lnkd.in/ed8v38cg
-
Mike Krieger shared thisThe best intelligence should also be easy to use. Today, Claude Cowork and Chat start becoming one Claude. Making models more capable isn't the only hard work. We also need to match that capability with products people can actually reach. Claude users told us, more and more, that they weren't sure which product to use for a given task. That question was getting in the way before the work even began. Now you bring the task. Claude figures out what it needs. I've been on the unified version for a few weeks. The thing that changed: Claude handles the decisions about how to get the work done. I keep my thinking on what I actually want to make. That shift extends to the output too. Claude Docs, Claude Slides, and Claude Design are now in every conversation. Ask for a document, a deck, or a design. Get back a file you can edit, share, and download. We think this brings many more people to the moment where Claude's capability becomes real for them. Try it and let us know what you think. We'll be getting it dialed in over the coming weeks.
-
Mike Krieger shared thisWe just launched Fable 5.1! I was testing it out a couple weeks ago and I asked Claude to build an interactive model of the human brain. No external assets, no reference images. It sculpted every fold of the cortex in code, then traced one sentence, "Could you pass the salt?", all the way through the thought process. Fable 5.1 is the upgrade to Fable 5 we built from customer feedback, and it builds on Fable 5’s strengths: long-running, agentic work across coding and knowledge work. The model really shines at finishing long-running, complex tasks more efficiently. Give it a concrete target, like "make this test run 10% faster" or "find the root cause of this bug", and it works at the problem until it gets there, communicating blockers clearly as it moves through the work. In early testing, Shopify said “Fable 5.1 is more comfortable with long, unattended work than Fable 5. I've had workflows run for a long stretch without losing the plot: it keeps its own records, reprioritizes as things change, and picks up where it left off." We’ve also made the safety classifiers more precise. Bio classifiers intervene on benign requests 85% less often than the ones Fable 5 launched with, and Claude Code users can expect an average of around 60% fewer interventions per session from our cyber safeguards compared to Fable 5. We’re also enabling vulnerability finding in source code.. Also, we dropped cache read pricing by 75%. This change lowers costs for customers by up to 45% depending on the workload due to the cache heavy token mix in agentic workloads. Processed input/output pricing remains unchanged at $10/$50M per million tokens. Fable 5.1 is on the API and in paid Claude plans today. There's more in the blog: https://lnkd.in/gV3VDW55. I'd love to hear what you build with it!Introducing Claude Fable 5.1 and Claude Mythos 5.1Introducing Claude Fable 5.1 and Claude Mythos 5.1
-
Mike Krieger shared thisClaude Opus 5 is out today! It comes close to Fable 5 intelligence at half the price, and I've really loved playing around with it over the last couple weeks. My favorite game as a kid was Transport Tycoon, so I had Opus 5 build a browser-based 3D viewer for its save files, letting me walk my old maps at eye level while the trains run their schedules. 8 hours and 70k lines of code later, it loads a 4.2 million tile world in under six seconds and holds 60fps. It added a day/night cycle without being asked. The prompt was four sentences. A year ago, a project like this meant builder and critic agents checking each other's work, role prompts, and careful task decomposition. Opus 5 verifies its own work and keeps finding useful improvements for hours. If you're a founder, it's worth asking how much of your product's value lives in the scaffolding around the model, because the models keep absorbing the scaffolding. The benchmarks are great too. On Frontier-Bench, Opus 5 more than doubles Opus 4.8's score at a lower cost per task. On CursorBench it comes within 0.5% of Fable 5's peak score at half the cost per task. Opus 5 is on the API and the default on Claude Max starting today, priced the same as Opus 4.8. Cyber classifiers step in about 85% less often than on Fable 5, with flagged requests falling back to Opus 4.8. Check out the blog for more details! https://lnkd.in/gewNUPnD
-
Mike Krieger shared thisLast weekend I finally built an app I'd been meaning to make for years: a tracker for all the books, music, and movies I want to get to. It went from idea to working in a weekend, mostly while the model ran on its own. That model is Claude Fable 5, and it's out today. Fable is a Mythos-class model. The most capable class of systems we've built, and the first one we've made generally available. It's state of the art on nearly every benchmark (SWE-bench Pro went from 69.2 with Opus 4.8 to 80.3), with a lead that grows as tasks get longer and harder. With earlier models, you broke a project into model-sized tasks and stitched the results together. Fable holds the whole project. It plans, runs for hours or days, checks its own work, and comes back when it's done. Lovable's CTO put it well: Fable "understands what builders mean, not just what they type. Apps that took a hundred prompts a year ago, it now one-shots." Requests touching cybersecurity or biology fall back, transparently, to Opus 4.8 while we scale trusted access programs. More than 95% of sessions never encounter one. For builders, it's worth rethinking what you scope. Projects we used to plan a quarter around are now weekend experiments. $10/$50 on the API, in paid Claude plans starting today. More in the blog: https://lnkd.in/gXAbGSxb
-
Mike Krieger shared thisWe just shipped Claude Opus 4.8. It's the most capable model we've put out and the best you can build on right now, outside the Mythos-class systems we're still testing under Project Glasswing. SWE-bench Pro went from 64.3 to 69.2. But the improvement I keep coming back to is honesty. Opus 4.8 is about 4x less likely than 4.7 to let a flaw in its own code slide past unremarked. It tells you what it's unsure of instead of dressing up thin progress as finished work. For anyone whose agents run with real oversight cost, that's worth more than another point on a leaderboard. Where that compounds is on big, long-running jobs, which is why I'm most excited about dynamic workflows, launching in research preview today. Claude plans the work, fans out across hundreds of parallel subagents in a single session, and verifies its own output before handing it back, with 4.8 letting those agents run longer before they report. This is aimed at the work that used to take a quarter and a working group: codebase-scale migrations, sprawling refactors, and bug fixes across hundreds of thousands of lines, graded against the test suite you already trust. Available in Claude Code for Enterprise, Team, and Max. One practical note for fellow CTOs: Opus 4.8 defaults to "high" effort, and that's the right setting for most work. There's an "xhigh" level for the even harder, long-running tasks. It's strong, but it's token hungry, so reach for it deliberately. We've made the headroom for it on both sides: higher rate limits for subscribers, and fast mode now 2.5x the speed at 3x lower cost ($10/$50) for API customers. Standard pricing is unchanged at $5/$25. Available today. More in the blog: https://lnkd.in/gxxEeKxn
-
Mike Krieger shared thisClaude Opus 4.7 is live today! I've been using it for a couple weeks and the thing I keep coming back to is how much further I can push a single prompt. Ambiguous, multi-step work that I used to break into pieces now goes in as one task, and it checks its own work before telling me it's done. That self-verification loop is the part that changes how it feels to build with. The numbers from early partners track with that. Cursor's internal bench cleared 70%, up from 58% on 4.6. Notion saw a 14% lift on their evals with a third of the tool errors. Also shipping today on the platform: higher-resolution image support (3x the pixels of prior Claude models), task budgets in public beta for guiding token spend on long runs, and a new xhigh effort level for tuning the reasoning/latency tradeoff. Let us know what you think! https://lnkd.in/gRUHa38h
-
Mike Krieger shared thisSonnet 4.6 just launched! For most people, this should be your new default model. Computer use is the headline. We've been building toward this for 16 months, and Sonnet 4.6 is by far our strongest model for it. Early users are seeing near-human performance on tasks like navigating complex spreadsheets and multi-step web forms. And Sonnet 4.6 approaches Opus-level intelligence at the same price as Sonnet 4.5. Also shipping today: MCP connectors in Claude in Excel (S&P Global, LSEG, FactSet, PitchBook, Moody's, Daloopa), a 1M token context window in beta on the API, and web search and fetch tools that are now more accurate and token-efficient. We can’t wait to see what you build! https://lnkd.in/gA7NhK8S
-
Mike Krieger reacted on thisMike Krieger reacted on thisTen years ago I sat outside this sign. That was as close as I got. This time I’m in. I’ve left Anthropic and stepped into YC 🟧 The photo is 2015. I was a founding engineer at Sonar, a startup in San Francisco. Those were special years. We were never in YC. I always wanted to be. Anthropic was a ride 🎢 I started when AI was a tool people used to help with PRs; by the time I left it was writing all of my code. What it meant to be an engineer changed: we had to reinvent how we work while we worked, and find answers that didn’t exist before. Best job I’ve had. Now, as a founder, I’m leaping into more of it: even less structure, and a tomorrow that’s even less defined. Thank you to the people who believed in me before there was anything to believe in: Shoya Philipp Dudchuk Simon Trevor Igor Finbarr Ilya Anna Vincent Hinchey Jiangda Mansoor Sina Meraji Yaroslav Bulatov and many others. And to everyone I worked with at Anthropic. What I’m building gets its own post, next. I want to reimAgine the surface where AI and humans collaborate 🤖 Follow along if you want to watch it get made 🧑🍳🍲
-
Mike Krieger reacted on thisMike Krieger reacted on thisOne year at Anthropic. Here is what I built. → An internal app deployment framework used by nearly every Anthropic employee → This framework powered shipping Claude Design, and more → Tuned that framework to 5-second deploys → The first internal version of Claude Code Artifacts. It went viral → An internal BI tool the whole company uses → A bunch of other things ❤️ to all the Anthropic people I got to work with! I’ll miss y’alls! Cheers!
-
Mike Krieger liked thisMike Krieger liked thisI'm both excited and nervous: Wanderly school pilots kicked off this week! 😅 I’ve been so heads-down this summer building that I haven’t had a chance to update y’all after my last newsletter calling for help to get Wanderly into 5 schools. I asked for 5 schools because I was really counting classrooms, estimating 1–2 per school for a goal of about 10. I ended up with 11, across just 2 schools that were eager to sign up more classrooms than expected. Wish me luck! Full details in the newsletter post in the comments :D
Experience & Education
Languages
-
Portuguese
-
Recommendations received
1 person has recommended Mike
Join now to viewOther similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top contentOthers named Mike Krieger in United States
130 others named Mike Krieger in United States are on LinkedIn
See others named Mike Krieger