🚀 OpenAI and Anthropic Drop New Models Hours Apart

Anthropic has a new Opus, and the interesting part is how close it now gets to Fable at a much lower price. For teams using Claude at scale, that makes Opus 5.5 a much more serious option. The bigger story is how quickly frontier-level capability is getting cheaper. Let’s dive in!

In today’s insights:

LATEST DEVELOPMENTS

Evolving AI: Anthropic launched Claude Opus 5.5, its strongest Opus model yet with Fable 5.1-level performance at a much lower cost.

Key Points:

Opus 5.5 is built to use less compute while handling longer coding and agent tasks. One early tester audited and fixed a 200,000-line codebase in under three hours. Opus 5 took more than 20 hours on the same job and used 2.5× as many tokens. The new model also generates output more than 30% faster than Opus 5. It follows writing instructions more closely and surfaces important information earlier, making long sessions easier to follow.

Anthropic is shrinking the gap between its premium and everyday models. Opus 5.5 now reaches Fable 5.1-level performance on most work while standard token prices are 60% lower, making stronger Claude models practical for far more coding, research and agent tasks. As these jobs run for hours and consume more compute, companies will care less about which model tops a benchmark and more about which one finishes reliable work at the lowest total cost. That puts pressure on Fable’s premium positioning and raises competition across the frontier model market.

With Skydive, you can build a team of AI agents to take work off your plate.

Create agents for customer support, sales, marketing, engineering, ops, and more. Give each one a role, connect the tools they need, and hand off the work.

Your agents can work on their own or together, sharing context and handing off tasks to get bigger jobs done.

Start with one agent. Build a whole team around you.

Make your first agent

Evolving AI: OpenAI has launched Sol and Luna, filling out GPT-6 as a three-model family built for different levels of capability, speed and cost.

OpenAI’s new Sol and Luna bring GPT-6 upgrades to its faster, cheaper models, with gains across coding, agents, factual accuracy and alignment. Sol is aimed at more demanding work, while Luna is built for high-volume tasks where speed and cost matter more. Together with Astra, they complete the current three-model GPT-6 family. Terra, which sat between Sol and Luna in the GPT-5.6 lineup, has been dropped from the new family. OpenAI also improved prompt caching, giving reused input a 90% discount and lowering the cost of long-running agent workflows.

OpenAI is showing that the real value of a frontier model may extend well beyond selling that model itself. Sol and Luna carry training advances developed for Astra into the rest of the lineup, turning one frontier push into improvements across multiple products. If OpenAI can keep repeating that cycle, Astra becomes an R&D engine for future generations rather than just its most expensive model. Terra disappearing also points toward a simpler model strategy, where stronger core tiers leave less need for models in between. Over time, a major advantage could be how quickly a lab can spread each frontier breakthrough across its entire product line.

Evolving AI: a16z is backing a new San Francisco school built around AI, projects, and real startup work for students leaving high school.

Andreessen Horowitz is taking a very different approach to education with the Academy. Students will spend most of their time building their own projects, taking short classes from experienced founders and operators, and working directly with companies through co-ops. HAA’s core bet is that with AI, strong mentorship and direct access to companies, ambitious students can start doing serious work at 18 instead of spending years preparing for it. The first one-year program will not offer a college degree or academic credit, so for now it is closer to an alternative path than a replacement for university.

OpenAI, Anthropic and other major AI companies are now helping shape how young people are trained before they even enter the workforce. The bigger shift is that AI lets students attempt harder work much earlier, which weakens the old assumption that years of classroom training must come first. If HAA can show that projects, company experience and access to powerful tools produce strong outcomes, traditional universities may face pressure to rethink how much of education is spent preparing students for work versus letting them do it.

QUICK HITS

🍎 Apple showed a trillion-parameter AI model running across four Mac Studios from a single wall outlet.

☎️ Meta is testing human concierges inside Muse, with contractors handling some calls the AI agent is asked to make.

🔐 Microsoft disrupted the EvilTokens AI platform, linked to 12,000+ compromised inboxes across 10,000+ organizations.

🧬 AWS is backing a neuron-derived video model that its maker says cuts inference costs 80% using measurements from living neurons.

🛡️ Palo Alto Networks launched continuous AI security testing using Claude Mythos 5, GPT-5.6-Cyber and open-weight models.

 *partner link