Claude Opus 5 Launch | AI Field Notes #74

AnalysisThe pitch is that how an agent talks is now worth paying for. Cognition, the company behind the Devin coding agent, acquired The Interaction Company of California, maker of the messaging assistant Poke, in a deal it put in the low nine figures. Poke lives inside iMessage, WhatsApp and Telegram and handled more than 100 million messages in three months by chatting through scheduling and travel tasks. Cognition wants that conversational feel wired into Devin so the tool reads like a co-worker who can take a joke rather than a terminal that spits diffs. The models are commoditizing, so the interface becomes the moat.

AnalysisThe headline number is the price that did not move. Anthropic shipped Opus 5 on July 24 at five dollars per million input tokens and twenty-five per million output, the same sticker as Opus 4.8, while claiming it more than doubles the older model's score on its internal Frontier-Bench and triples the next-best result on ARC-AGI 3, a reasoning test built to resist memorization. On coding it lands within half a percent of Fable 5, a rival frontier model, at half the cost per task. Capability is climbing a step function while the bill per token flattens, which quietly changes the math on every agent you were told to ration.

AnalysisThe customer list is the story. AMD used its Advancing AI event on July 23 to unveil Helios, a rack-scale system Lisa Su called the industry's highest-performance AI rack, and named Microsoft, OpenAI, Meta, Oracle and Anthropic as buyers. Anthropic's slice alone runs up to 2 gigawatts of Instinct MI450 GPUs, enough electricity to light a small city. Helios ships later this year and claims to beat Nvidia's Vera Rubin systems on several metrics AMD did not fully detail. Nvidia has owned the training rack for years, so the buyers hedging their supply is what actually shifts here.

AnalysisA model that watches how a billing clerk clicks through her afternoon is the bet here. Prentis, co-founded by Ritankar Das with Reid Hoffman and Mark Pincus, is raising $100 million at a $1 billion valuation to build agents that learn routine office workflows by imitation. Its Hive-32B model already claims to beat OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.6 on computer-use tests, where a model drives real software, at roughly a tenth of the API cost, and the eight-month-old company says it has signed contracts worth up to $50 million across healthcare, manufacturing and apparel. The target sits below the chat window, in the desk work itself.

AnalysisThe attackers automated first, so the defense is racing to catch up. AegisAI, founded by two former Google security engineers who worked on Safe Browsing and reCAPTCHA, raised a $36 million round led by Battery Ventures to build agents that scan email for the tells of AI-generated spear phishing, the targeted scam messages crafted for one specific recipient. Co-founder Cy Khormaee says AI-powered attacks now slip past existing filters more than half the time, close to twice as effective as the older kind. Customers already include LangChain and the fintech Mesh. The same tools writing the lures are being turned to catch them.

AnalysisChoosing a generative model has become a full-time job, so Runway is automating the choice. On July 23 the company launched Media Router through its developer platform, a tool that sends each image, video or audio request to whichever model, its own or a third party's, best fits a developer's stated priority of quality, speed or cost. Runway leans on its in-house creative team's taste to score things like how a video model handles motion or lip-sync. Six launch customers including Adobe, Cloudflare, ElevenLabs and Expedia are already wiring it in to generate media inside their own products. When the models multiply, the router becomes the product.

AnalysisDoubling your valuation in seven months on a chip that has not shipped in volume takes some nerve. Etched closed a $300 million round on July 23 at a $10.3 billion mark, up from $5 billion in December, with Sequoia leading and Peter Thiel, Andrej Karpathy and SK Hynix in the mix. The company builds silicon tuned only for inference, the work of running a trained model for users, splitting the job into a low-voltage prefill chip and a shared memory fabric that lets many chips pool memory, and it says it already holds $1 billion in booked orders with TSMC making the parts. The wager is that general-purpose GPUs waste money once a model is trained.

AnalysisUber's founder wants to automate the physical world, and a16z just handed him $1.7 billion to try. Travis Kalanick's holding company Atoms, built on the bones of his CloudKitchens venture, raised the round on July 22 with Andreessen Horowitz leading, Ben Horowitz taking a board seat, and Uber itself among the backers. Atoms bought Pronto, a heavy-industry automation firm, and plans to move into mining and robotics, with Kalanick describing manufacturing as compute and real estate as storage. Most of the cash is earmarked for hiring. The bet is that AI's next decade runs on trucks and drills more than chatbots.

AnalysisA profitable software company shedding a fifth of its people is the shape 2026 layoffs keep taking. Monday.com cut roughly 630 workers on July 22, about 20 percent of headcount, and booked $45 million to $55 million in restructuring charges, saying it wants a leaner operation built around its AI work platform of agents and no-code automation. The firm is not in financial trouble. It rebuilt the product around AI agents working next to employees, then decided it needed fewer employees to sell that story. Across tech, 78 percent of this year's cuts cited an AI refocus. The rebrand and the redundancy arrive together.

AnalysisThe grace period ends in nine days. On August 2 the European Union's AI Act, its risk-based law for artificial intelligence, switches on the obligations for high-risk systems and general-purpose models, covering conformity checks, transparency notices and record-keeping. The Commission paired the date with a July action plan on AI and cybersecurity and a push to build EU capacity to test frontier models before they reach the market. Full enforcement against some high-risk systems embedded in products stretches to 2027, so the teeth arrive in stages. Any team shipping AI into Europe now works against a compliance calendar with a real deadline.