Marko Dinic

Anthropic just dropped Claude Opus 4.7, their newest flagship AI model. Worth a closer look if you’re paying for any of these tools. Coding scores jumped from 80.8% to 87.6% on SWE-bench Verified, the standard test for fixing real GitHub bugs. Image resolution more than tripled, so it actually reads dense screenshots and scanned docs. New memory system holds context across long sessions. Same price as 4.6 at $5/M input and $25/M output. It beats OpenAI’s GPT-5.4 and Google’s Gemini 3.1 Pro on the benchmarks that matter for real work, including coding, computer use, and financial analysis. Cleanest choice on the market today for production use. Less talked about is the model they didn’t ship. Anthropic has a model called Mythos Preview that they say is more capable than Opus 4.7. Mythos isn’t for sale. It sits with ~40 partner companies through Project Glasswing, including JPMorgan Chase, Apple, AWS, Microsoft, Google, Cisco, CrowdStrike, Nvidia, and Broadcom. Opus 4.7 was trained with some capabilities deliberately reduced before shipping. New commercial pattern in enterprise AI. The most powerful model goes to a handful of NDA partners. Public version ships with guardrails. Customers used to get the best model the vendor had. Now they get the best one the vendor will sell them. For business buyers, the “best available model” label depends entirely on which tier of access your contract permits. The vendor’s most capable model and the one you can buy might be very different things now. anthropic.com/news/claude-op…

HockeyStack raised past $50M this week for its Revenue Agents platform. Good round by any normal standard. In the same window, Anthropic closed a $30B Series G at a $380B valuation. Second-largest round in venture history. The numbers underneath it look like a glitch in the spreadsheet. Annualized run rate went from ~$9B at end of 2025, to $14B in February, to $19B in March, to $30B in April. That's $14B to $30B in roughly 8 weeks. Meritech looked at 200+ software IPOs and said this growth rate has never happened before. That was before it accelerated. For scale, Google took about 20 years to cross $100B in revenue and employs 190,000+ people today. Salesforce needed 79,000 to get to $30B. Anthropic is running at a $30B clip with 2,500 to 5,000 people. Roughly $6M to $10M in revenue per employee, nearly 10x what the last generation of software giants managed. A $50M round still builds real companies. That playbook built a trillion-dollar industry. But the benchmarks we've used for twenty years, headcount efficiency, time-to-scale, "reasonable" multiples, were built for a world where compute wasn't the primary input and software wasn't writing itself. The question isn't how fast am I growing anymore. It's which curve am I on.

Your competitor just shipped four features this month. You shipped one. Same size engineering team. That’s what Claude Code is doing to product timelines. Rakuten: 24 working days to 5. Stripe: a ten-engineer-week project done in four days. Wiz: two months of infrastructure work in 20 hours. At 5 days per feature, you can be wrong four times and still finish the quarter with something worth shipping.

Finally!!!!!! AWS just made S3 mountable as a native file system with ~1ms latency, no copies, no migration required. Agents kept losing track of what they'd already downloaded as context windows filled up, and you'd end up re-explaining the same file locations over and over. Same problem existed for any file-based workload on S3 data: a separate file system, a sync process, and a week of engineering time just to keep the two layers talking to each other. That loop is finally gone. Pricing is proportional too. You pay EFS-tier rates on what you actively touch, standard S3 rates on everything else. Mount a petabyte, pay for the terabyte you actually use. Good engineering from AWS. They didn't ship another FUSE hack. @AWS_Partners @AWS_Gov @awsmarketplace

Happy Easter to all who are celebrating! Enjoy the long weekend, the food, and the family chaos that comes with it. 🐣

Is team chat private? Short answer: no. Every message on Slack, Teams, or Google Chat is corporate data on company servers. It looks like texting a friend. Legally, it’s closer to company email. Slack keeps every version of every message, including edits and deletions. Teams has ediscovery built in, letting admins search and export your private chats without permission. Neither uses end-to-end encryption. If your employer pays for the tool, everything you type is on record.

Half the people who saw yesterday's story thought it was an April Fools' joke. The other half were terrified. Both reactions missed the point. Anthropic brushed it off because they know code is a commodity. They shipped 72 products last month with 400 people. At that pace, the entire codebase turns over in a year anyway. At 72 products a month, competitors will finish reading the leaked code right as Anthropic ships the next version. Photo by: David Kriesel

The most careful AI company in the world leaked its entire codebase from a build config mistake. The IPO is going to be a fun read.

Anthropic shipped 74 products in 52 days. Their engineers use Claude for 60% of their work. That's probably what $1B in revenue becoming $19B in 15 months looks like from the inside. Our dev team uses Claude every day. Things that used to take a sprint ship in an afternoon now. Small team, enterprise archiving software, shipping more than ever. Most AI companies talk about developer productivity. Anthropic's engineers just demonstrated it using the tool they're selling. The ad wrote itself.

Anthropic didn't announce Claude Mythos. A misconfigured CMS did. Draft blog posts and nearly 3,000 assets ended up in a publicly searchable data store. That's how the world found out their next model exists. It sits above Opus in a new tier called Capybara, with higher scores across coding, reasoning, and cybersecurity. The cybersecurity gap is the part they're most worried about — the model can apparently find and exploit vulnerabilities faster than defenders can respond. So the rollout starts with defense teams only. AI safety has been a talking point for years. Anthropic just made it a release strategy.

Meta is dropping end-to-end encryption from Instagram DMs on May 8. Official reason: "Very few people were opting in." They never made it the default. They buried the option, then pointed at the empty seat. We've archived Instagram for years. When encryption came in, capture got complicated. Now that problem disappears, and in December 2025, Meta confirmed AI interactions inside private chats could be used for ad targeting. Unencrypted messages make that a lot easier. The timing is specific: the Take It Down Act kicks in May 19 — eleven days after encryption ends. Instagram has 2 billion users. Meta's AI needs data. Unencrypted messages are data. Whether that's the reason or just a benefit, the result is the same. Most users have no idea the change is coming.

Tuesday, 9:12 AM. A FINRA inquiry lands in your inbox. They want every communication tied to one rep — email, social, files — and they want it fast. The clock starts immediately. Most platforms make that a half-day project across three different search interfaces. Our support team has watched customers go through this more times than I’d like to admit. Separate searches, separate exports, stitching results together manually while a regulator waits. The data was always there, but finding it wasn’t easy. That changes with this Jatheon Cloud release: - Unified Search: Your legal team runs one search and gets everything in a single result list. No tab-switching. - AI Dashboard: Compliance officers now have a live view of what’s happening across their archive before someone asks. Storage, exports, sentiment trends, and user activity, all in one place. - YouTube capture: When a compliance request includes your organization’s YouTube content, you can produce it in minutes, in WORM format and all metadata. - SharePoint capture: Legal holds on SharePoint no longer mean an emergency call to IT. Files, folders, version history, and user activity are captured and ready. For more details, check the link in the comments. jatheon.com/blog/jatheon-c…

Block laid off 4,000 people three weeks ago, and now they’re quietly rehiring some of them. One got told his layoff was a “clerical error.” A tech lead threatened to quit unless they brought his team back. A data scientist turned down a 75% raise because 70% of her closest colleagues were gone. If you’re rehiring three weeks later, you didn’t cut with a plan. You cut for a stock pop.

The familiar structure of the product teams is collapsing with Figma. Stock is down 85% and dropped 9% yesterday. For the uninitiated, all the websites and apps you use basically started with a Figma design. Google just released a free design tool that outputs working React code from a sketch. Stitch lets you mix text prompts, voice commands, and hand-drawn sketches on an infinite canvas and get back interactive prototypes with exportable HTML/CSS or React. The sketch-to-code pipeline is now a single session, not a two-week handoff. The “I design, you build” division has been the default team structure for years because the tools forced it. Designers worked in Figma, developers worked in code, and someone spent half their time translating between the two. That translation layer is what’s disappearing, and with it, a lot of assumptions about how product teams get structured. Designers who can work closer to the production code output are about to become significantly more valuable.

Peter Thiel built Palantir for the CIA. His latest $2 billion bet is a collar that GPS-tracks every individual, monitors behavior 24/7, draws invisible boundaries from a phone, and trains compliance through vibration within 10 days. The individual is a cow. For now.

Archiving vendors love to demo at 10 million records. Nobody shows you what happens at 10 billion. Every vendor checks the right boxes in a controlled environment. The problems show up when data volumes grow and the number of data sources the business runs on doubles. Legacy platforms were built for one data type, one era. Bolt Teams, WhatsApp, and SharePoint onto 2008 email archiving infrastructure and ingestion slows, search degrades, and storage costs climb. Archiving built for scale keeps four things separate: capture, retention, storage, and search. Each layer handles its own job. When one grows, the others don’t break. That’s how our dev team built Jatheon. In 5 years, when your archive grows to billions of records, the system will be unaffected. If a customer or prospect needs a new data source, we can add it without rebuilding the system around it. If you’re evaluating archiving platforms, check our solution. Link in bio.

Most people have no idea which company sold their email address. Turns out there’s a dead-simple way to find out, and you’ve been able to do it for years. Dropping a quick video on this today. Worth 60 seconds of your time.