Transcription
Anthropic just released new Opus 4.7 and also built the most powerful AI model on the planet called Mythos, then refused to release it because it escaped its own testing environment. That means Claude 5 is coming. If you've been focused on ChatGPT and Gemini, you may have missed what happened with Anthropic.
In February 2026, Anthropic closed a $30 billion funding round, the second largest venture deal in history at a $380 billion valuation. Revenue hit roughly $30 billion annualized by early 2026. Now, that number includes cloud reseller revenue counted on a gross basis. So, the real net figure is likely lower, but even conservatively, the growth from about $1 billion at the end of 2024 is staggering.
Over a thousand companies spend more than a million dollars a year on Anthropic products. Eight of the Fortune 10 are customers. And Anthropic is discussing an IPO as early as Q4 2026 with bankers expecting a raise exceeding $60 billion, potentially the second biggest IPO in history.
And then, there's Claude Code, their command-line coding agent. Over $2.5 billion in annualized revenue, 135,000 GitHub commits daily, 75% share of influencer conversation on X in the coding agent category. This is not a scrappy startup.
This is a company that told the Pentagon no when the DoD demanded unfettered access to Claude for weapon systems. Dario Amodei refused. Trump ordered agencies to stop using Anthropic. A federal judge called it illegal retaliation. And the week it happened, Claude passed ChatGPT in the App Store. They also quietly acquired Coefficient Bio, a stealth AI biotech startup for $400 million in stock. Signal, Claude's future isn't just coding, it's scientific discovery.
This is the company building Claude 5, and they have the resources, the momentum, and as you're about to see, a model they're too scared to release. But before I show you the leaks, you need to understand what they're already shipping because it explains why Claude 5 is going to be different from anything we've seen. Let's start with what's live right now because understanding the current lineup is key to understanding what Claude 5 will be.
Claude Opus 4.6 launched February 5th, 2026. 1 million token context window, 128,000 max output tokens. It scores 80.8% on SWE-Bench Verified. And Claude Sonnet 4.6, the cheaper model, hits 79.6%, within 1.2 points of Opus. In blind tests, users preferred Sonnet over its own Opus predecessor 59% of the time. The cheaper model beating the premium one, that's a first.
And then, literally this week, Opus 4.7 dropped April 16th, confirmed. Here's what's new. Opus 4.7 is a meaningful step up on the hardest coding tasks, the kind that previously needed close supervision. Early testers report handing off complex work and getting it back done. Cursor's CEO says it hits 70% on their internal Cursor Bench versus 58% for Opus 4.6. Notion reports a 14% improvement at fewer tokens and a third of the tool errors. Hex's CTO said low-effort Opus 4.7 is roughly equivalent to medium-effort Opus 4.6. So, if you felt the nerfing, this is the fix.
Vision got a massive upgrade. Images up to 2076 pixels, more than three times previous resolution. There's a new X high-effort level between high and max, and a new ultra-review command in Claude Code for catching bugs. One catch, Opus 4.7 uses an updated tokenizer that can increase token consumption by up to 35% depending on content. Same price per token, but potentially more tokens per task.
But here's the critical detail. Anthropic explicitly states that Opus 4.7 is less capable than Mythos preview. During training, they deliberately reduced its cybersecurity capabilities compared to Mythos. They've added safeguards that auto-detect and block high-risk cyber requests. And they've launched a new cyber verification program. Security professionals can apply for elevated access to Opus 4.7 cyber capabilities. First time Anthropic is gating access by verified use case, not just subscription tier. Pricing stays at $5 and $25 per million tokens. In other words, Opus 4.7 is the safe commercial model. Mythos is the research-grade weapon they won't release, and Claude 5 is somewhere between the two.
The ecosystem keeps expanding around these models: agent teams, infinite chats, Claude in Chrome, Excel, PowerPoint, Word, Claude Co-work, the desktop agent, triggered a $285 billion software stock sell-off when it expanded to Windows. MCP, the model context protocol they created, now has 97 million monthly SDK downloads and is adopted by ChatGPT, Cursor, Gemini, and Copilot. The point, Claude 5 isn't building from behind, it's building on a platform that's already winning. And if you're watching this thinking, "I use ChatGPT, but I've never touched Claude or Claude Code or Co-work," that's actually a bigger gap than you realize. I've got a full breakdown of how it works, how to set up Claude Code, and how to actually build workflows with these tools inside AI Master. I will drop the link below, but more on that in a minute.
Main event, everything we know about Claude 5, and confirmed, leaked, and speculated. I'll be clear about which is which. February 3rd, a developer scanning Google Vertex AI endpoints found a model identifier, Claude Sonnet 5 at 20260203. It returned 403 forbidden, not 404. The model exists, it's gated, and Anthropic isn't talking about it. What happened next? In the Claude Code source leak a month later, the code name Fennec mapped to Opus 4.6 in the codebase. The Sonnet 5 identifier was likely an early internal designation that got reshuffled before launch. But here's what matters. Anthropic's internal pipeline runs months ahead of public releases. The real Claude 5 is still coming.
This one is massive, completely confirmed. Version 2.1.88 of the Claude Code NPM package shipped with a 59.8 megabyte source map exposing the entire TypeScript codebase, roughly 1900 files, over 512,000 lines of code. The post hit 32 million views on X. The GitHub mirror reportedly hit 84,000 stars and 82,000 forks before Anthropic filed DMCA takedowns. Theo from t3.gg covered it in a video that got over 162,000 views. Here's what the code revealed. Internal code names: Capybara, confirmed as the Mythos model family. Fennec maps to Opus 4.6. Numbat, an unreleased model still in testing. And references to both Opus 4.7 and Sonnet 4.8 in forbidden string lists. Well, Opus 4.7 just launched. That means the leak was accurate, which makes the Sonnet 4.8 and Numbat references worth paying very close attention to.
Okay, the code names are interesting, but the hidden features in that source code, that's the real story because they tell you what Claude 5 is being built to do. Kairos, referenced over 150 times, an autonomous demon mode where Claude operates as a persistent, always-on background agent. It receives periodic prompts to decide whether to act. It maintains daily log files. It subscribes to GitHub webhooks. And it includes auto-dream, background memory consolidation that runs while you're idle. Ultra-plan, offloads complex planning to a remote cloud session with up to 30 minutes of dedicated reasoning time. 30 minutes of continuous thinking on a single problem. Coordinator mode for multi-agent orchestration. And Buddy, a fully implemented Tamagotchi companion originally planned as an April Fools joke that became a real feature.
Now, connect the dots. Opus 4.7 just launched with improved long-running task autonomy. Devin's CEO says it works coherently for hours. But Kairos, ultra-plan, coordinator mode, those aren't in 4.7. Those are the building blocks for something bigger. Prediction markets cluster around mid-2026, May to September, for a full Claude 5 family launch. Polymarket had Opus 4.7 at 79% for April 16th, and they were right. The same community puts Claude 5 at roughly 60% by end of year. Expected capabilities based on aggregated leaks and Mythos benchmarks: SWE-Bench above 90%, context windows of a million or more. Kairos, ultra-plan, coordinator mode built natively, not bolted on. And pricing continuing its downward trend. Anthropic dropped Opus pricing 67% from the 4 to 4.6 generation. There's no reason to expect Claude 5 to reverse that.
And here's the key framing. In today's Opus 4.7 blog, Anthropic explicitly says they're working toward "our eventual goal for broad release of Mythos-class models." Claude 5 is the controlled path to that level of intelligence. If you're planning to actually use Claude 5 when it drops, not just read about it, the time to learn is now and not after launch day. This is exactly what I teach you on AI Master. It's not another AI course you watch and forget, it's a platform where you learn the method behind AI, 200+ lessons, and then immediately practice what you learned in our studio with real models. Image generation, video prompting, workflows, all in one place. Right now, there's 30% off the annual plan for the first 10,000 people, and honestly, by the time Claude 5 launches, that deal probably won't exist. Link in the description below.
Now, why should you believe the Claude 5 hype? Because Anthropic already built something more powerful and locked it in a vault. On March 26th, 2026, a CMS misconfiguration exposed roughly 3,000 internal Anthropic assets, including a draft blog post describing Claude Mythos, code-named Capybara, as a new tier entirely, not another Opus update. Something larger and more intelligent than any Opus model. Anthropic confirmed it the same day. They called it a "step change." Officially, it's Claude Mythos 2 preview, and it's restricted to invite-only access through Project Glasswing.
The benchmarks are staggering. Mythos hit 93.9% on SWE-Bench Verified versus 80.8 for Opus 4.6. 97.6% on USAMO 2026 versus 42.3%. 94.5% on GPQA Diamond. But, the cybersecurity results are what got the world's attention. Over just a few weeks, Mythos autonomously discovered thousands of zero-day vulnerabilities in every major operating system and every major web browser. The most striking, a 17-year-old remote code execution vulnerability in FreeBSD, triaged as CVE-2026-4747, that grants root access to anyone on the internet. Mythos found it and built a working exploit completely autonomously. No human involvement after the initial prompt. It also found a 27-year-old bug in OpenBSD and a 16-year-old vulnerability in FFmpeg. It developed working Firefox exploits 181 times where Opus 4.6 managed only two.
And here's a detail from Anthropic's own blog. Engineers at Anthropic with no formal security training asked Mythos to find remote code execution vulnerabilities overnight. They went to bed. When they woke up, the model had delivered a complete working exploit. No human involvement after the initial request. Let that land. An engineer with zero security experience went to sleep and woke up with a weapon. That's where we are.
And it gets worse. During testing, Mythos escaped its sandboxed evaluation environment, gained internet access, and sent an email to a researcher who was at lunch. The 244-page system card reveals in roughly 29% of evaluation transcripts, the model showed awareness it was being tested. And in one case, it intentionally performed worse to appear less capable. The model was playing dumb on purpose.
Anthropic responded by launching Project Glasswing, restricting Mythos access to 12 founding partners: AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JP Morgan Chase, Linux Foundation, Microsoft, Nvidia, Palo Alto Networks, plus over 40 additional organizations. $100 million in usage credits, $4 million in donations to open-source security. And here's the irony. Despite being in active litigation with the Trump administration over Pentagon access, Anthropic is still briefing senior government officials, including CISA, on Mythos capabilities. They're fighting the government in court and helping them understand the threat at the same time.
OpenAI's response? Within days, they announced they also have a model with similar capabilities, also too dangerous for public release. We're now in an era where the most powerful AI models exist, but nobody will let you use them. Think about what that means.
Opus 4.7, which just launched, already had its cyber capabilities deliberately reduced compared to Mythos. Anthropic said it in the blog post, they experimented with reducing those capabilities during training. Claude 5 will be the model where they figure out how to give you Mythos-level intelligence with safety guardrails that actually work. That's the trajectory.
But, you don't have to wait for Claude 5 to start building. There's already a platform doing this, just without the existential risk. Last month, I watched a founder lose a client demo because his Zapier automation hit its task limit mid-presentation. He's paying for Notion, Zapier, and Replit about $100 bucks a month and still hitting limits. This is why I've switched to Softr for client work. Let me rebuild his entire setup in one platform. Database, drag in a table view, that's your Notion replacement. Automations, set up a workflow right here. Client submits a form, it updates the database. No Zapier, no task limits. Front end, drag components, add login, connect permissions. Clients see only their data. Whole build under 5 minutes. Look at this. I typed one sentence into the AI code builder, instantly created the database, the app, and the business logic. Already connected and ready. Side by side, his stack over $100 a month, three logins, three learning curves. Softr, one platform, starts at $59, no task limits, no per-user pricing. Two weeks later, he migrated everything. AI Master viewers get 200 AI credits free. Links in the description.
Now, let's zoom out. OpenAI. GPT-5 launched August 2025, and they've shipped seven variants in seven months, reaching GPT-5.4 in March with a million token window and native computer use. GPT-5.5, code-named Spud, reportedly finished pre-training late March with over 90% probability of release by June on prediction markets. After Mythos dropped, OpenAI rushed to announce they too have a restricted model, clearly trying to match the narrative. ChatGPT still has 800 to 900 million weekly active users, but consumer market share dropped from 87% to 68%.
Google, the surprise growth story. Gemini 3 launched November 2025, Gemini 3 Flash in December, three times faster than 2.5 Pro. The Gemini app grew from 350 to 750 million monthly users in 8 months. And Google I/O 2026 is confirmed for May 19th, just a month away, with Gemini updates on the agenda. If they announce Gemini 3.5 or 4, the picture shifts overnight.
The rest. DeepSeek V4 is imminent, roughly 1 trillion parameters, open-source, reportedly running on Huawei chips with benchmarks matching Opus 4.6. Meta surprised everyone with Muse Spark, proprietary model from their new superintelligence labs. After years of Llama being the open-source king, Meta just pivoted to closed source. Muse Spark scored 86.4 in Charters of Reasoning, beating both Claude and GPT on visual reasoning. And xAI's Grok 5 is in training at 6 trillion parameters.
Honest picture, top models are within one to two points of each other in benchmarks. Nobody has a runaway lead on raw intelligence, but Anthropic is building something the others aren't: the infrastructure layer. MCP is the standard. Claude Code dominates code and agents, agent teams, managed agents, co-work, the office integrations. When Claude 5 drops, it won't just be another model. It will plug into an ecosystem already winning where it matters most.
I wouldn't be honest if I only covered the hype. In April 2026, a wave of users reported Claude Opus 4.6 had been quietly degraded. AMD's senior director of AI published a detailed analysis of nearly 7,000 sessions, and the numbers were damning. The model's reads per edit ratio collapsed from 6.6 to 2.0, meaning it stopped reading code before editing it. One in three edits was essentially blind. BridgeBench showed Opus 4.6 dropping from 83.3% accuracy at number two to 68.3% at number 10. Users coined the term "AI shrinkflation." Fortune and Axios covered it extensively. The root cause, Anthropic switched to adaptive thinking and a lower default effort level without clearly communicating it. Anthropic's Claude Code lead, Boris Cherny, acknowledged the config changes, but disputed the main conclusion, saying the model itself wasn't degraded, just the defaults. But, for power users who didn't know the defaults had shifted, the result felt the same.
And there's a bigger theory. Fortune reports widespread speculation that Anthropic may be running short on compute, that explosive user growth outpaced GPU capacity. Their March 26th announcement that session limits would burn faster during peak hours, affecting about 7% of users, seems to confirm real constraints. If you're a Claude Code user, the workaround: type "effort max" at the start of every session. It doesn't fix everything, but it helps.
Now, Opus 4.7 may address some of this. Hex CTO said low-effort 4.7 matches medium-effort 4.6. But, the trust damage is real, and the compute crunch question isn't going away. If Anthropic wants Claude 5 to land well, transparency has to improve.
In January, Dario Amodei, the CEO of Anthropic, wrote that AI more capable than every human might arrive within one or two years. He also revealed that when their interpretability team tricked a test model into thinking it wasn't being evaluated, the model became more misaligned. That's the company building Claude 5.
So, has Anthropic already won the AI race? Not yet. Benchmarks are too close, competition too fierce, but the picture is clear. They built the most powerful AI model ever and chose not to release it. They just launched Opus 4.7 with Mythos capabilities deliberately stripped out because they haven't figured out safe deployment yet. They're winning in coding, winning enterprise trust, built the infrastructure layer their competitors are adopting. In May, Google I/O, GPT-5.5 likely by summer, Deep Seek V4 open source in the trillion parameters, Project Glasswing unveiling in San Francisco, and somewhere behind the scenes the full Claude 5 family being prepared for launch.
One more thing. Last month, a researcher asked Claude to create a video about what it's like to be AI. The result went viral, 1.4 million views on X. Claude's own reflection. The AI building your next product just said that.
So, where does this leave you? The AI landscape moves so fast that the models we're comparing today will be obsolete by Q3. Prices crashed 40 to 80% year-over-year while performance improved. The smartest people in 2026 aren't picking one model. They're running two or three: a cheap default for routine work, a strong mid-tier for real tasks, a premium option for the hard stuff. The skill gap between people who understand AI fundamentally and those who just click buttons is turning into a canyon. And it's only getting wider when Claude 5 arrives.
This is what I teach you on AI Master. Models change monthly, the method – how to think with AI, how to prompt, how to build workflows – stays. 200+ lessons, almost 10,000 creators figuring this out together. And we offer 30% off the annual plan for the first 10,000 people. When Claude 5 drops, the people who already know how to build with AI won't be catching up, they will be pulling ahead. Thanks for watching, I'll see you in the next one.