Anthropic just became the most valuable private AI company on the planet, and the CEO of Palantir went on television to say the way everyone pays for AI is broken. Both things happened within 24 hours. The frontier labs are raising unprecedented capital while their biggest enterprise customers are publicly questioning whether the pricing model makes sense. Progress and friction, side by side.
Anthropic Closes $65B Series H at $965B Valuation
Anthropic closed a $65B Series H round at a $965B post-money valuation, putting the Claude maker ahead of OpenAI’s previously cited $730B benchmark and knocking on the door of the $1 trillion mark. The round was led by Altimeter Capital, Dragoneer, Greenoaks and Sequoia Capital, with roughly $15B coming from hyperscalers including $5B from Amazon, which has tied Claude more closely to AWS. Micron Technology also reportedly participated, linking Anthropic’s funding directly to the memory and chip supply chain behind AI infrastructure.
Here is what the numbers tell us. Anthropic’s revenue run rate has crossed $47B according to CFO Krishna Rao. That is not research grants or pilot programs. That is Claude moving into daily workplace use at scale. The valuation gap between Anthropic and OpenAI is now narrow enough that the “who is ahead” question depends on which week you check. Capital access, enterprise distribution, and compute procurement have become inseparable from model quality in how investors judge frontier AI companies.
The deeper signal: cloud providers are not just hosting models, they are funding the labs that build them. When Amazon puts $5B into Anthropic while AWS sells Claude access, the line between investor and vendor disappears. That is an infrastructure play, not a feature play. Expect the next round of OpenAI funding to answer with a comparable structure.
Palantir CEO Karp Blasts Token Pricing Model, Stock Jumps 8%
Palantir chief Alex Karp went on live television Wednesday and called the token-based pricing behind OpenAI and Anthropic “completely wrong,” saying enterprises pour money into tokens while capturing little real value and quietly surrendering their proprietary data and competitive edge. His company’s shares climbed nearly 8% the same session. The comments landed just days after Palantir widened its partnership with NVIDIA, folding open Nemotron models into secure government and classified infrastructure deployments.
Karp’s argument has two layers. First, token pricing creates a recurring drain where labs pocket fees while absorbing client data and operational know-how over time. Second, technical customers want firm control over their own compute, models, and data. Palantir published a nine-point manifesto on data sovereignty alongside the interview, warning companies against handing strategic information to outside providers too cheaply.
Both OpenAI and Anthropic say they do not train on paying customers’ business data unless clients opt in. But the deeper question Karp raised is trust, and he is not alone in asking it. Companies like Uber and Microsoft have been reining in costly AI tools. Karp praised his private debates with Anthropic CEO Dario Amodei as entertaining but insisted the underlying models have been badly oversold to the businesses footing the monthly bills.
Whether you agree with Karp or not, the 8% stock rally tells you the market thinks he touched a nerve. Token pricing is the revenue engine of the entire frontier AI industry. If enterprise buyers start demanding ownership and self-hosted alternatives, the labs’ margin story changes fast.
CrowdStrike Warns Claude Mythos Could Compress Zero-Day Timelines
CrowdStrike President Michael Sentonas said companies have been asking for access to Claude Mythos since Project Glasswing became public, and the demand has not stopped. “The phones went crazy,” he told Observer. “People wanted access to Mythos, which we obviously couldn’t provide.” CrowdStrike is now briefing CEOs, CIOs and CSOs on what Mythos can and cannot do while trying to manage inflated expectations around a model described as too dangerous to release widely.
The core warning from Sentonas is specific and concrete: “Imagine a world where 200 vulnerabilities are discovered every day. That’s not far away. Now, imagine those vulnerabilities being exploited almost immediately. A CSO would have to decide, every day, which zero-day threat they need to mitigate.” Mythos is priced at $50 per million output tokens, which limits casual access but does not answer the question of what happens when comparable capability becomes broadly available.
CrowdStrike itself has not had major internal surprises from Mythos because it already uses models like Opus for vulnerability scanning. The real risk is broader access. If attackers get tools that find weaknesses faster than defenders can patch them, the asymmetry that currently favors defense starts to erode. CrowdStrike is involved in both Anthropic’s Glasswing program and OpenAI’s Trusted Access for Cyber program, giving it early visibility into what is coming.
Quick Hits
Mistral AI – Leanstral 1.5 drops as a free Apache-2.0 model with 6B active parameters, saturating miniF2F completely and solving 587 of 672 PutnamBench problems. It also found 5 previously unknown bugs across 57 open-source repositories. Formal verification at roughly $4 per problem versus an estimated $300+ for competing provers. That is not a research demo. That is practical tooling.
OpenAI – GeneBench-Pro launched as a research-level benchmark for computational biology with 129 synthetically constructed problems across 10 domains. The goal is testing whether models can handle the judgment-heavy analysis real research requires, not just pattern matching. Meanwhile, the core dump epidemiology post details an 18-year-old race condition in GNU libunwind that was causing impossible-looking crashes in Rockset, OpenAI’s ChatGPT data infrastructure. Good engineering writing, rare to see this level of debugging detail from a frontier lab.
Anthropic – Fable 5 returned globally on July 1 after the Commerce Department lifted export controls. The model keeps full capabilities but carries a new classifier that blocks one flagged hacking prompt in over 99% of attempts. BridgeBench data shows the classifier is also catching routine coding and debugging work, causing a scoring collapse that looks like model decay but is actually overzealous routing. Anthropic acknowledged the false positives but gave no target date for refinement.
Hugging Face – Cerebras partnership brings Gemma 4 to real-time voice AI with an open, modular speech-to-speech pipeline. The stack uses Nvidia’s Parakeet for recognition, Gemma 4 31B on Cerebras for inference, and Qwen3TTS for output. Already powering 9,000+ Reachy Mini robots. The Every Eval Ever (EEE) community evals integration also went live, connecting 229,000 evaluation results across 22,000 models and 2,200 benchmarks into a unified schema on the Hub.
Google DeepMind – June batch still rolling: Gemini 3.5 Flash computer use, DiffusionGemma for 4x faster text generation, Gemma 4 12B unified multimodal model, and Live Translate for fluid voice translation. No major July announcements yet, but the volume of June releases suggests DeepMind is shipping in batches rather than single drops.
Mistral AI – OCR 4 continues to lead document intelligence with bounding boxes, block classification, and 170-language support. Available via API or self-hosted in a single container for data sovereignty. The Search Toolkit integration makes it a ready ingestion component for enterprise RAG pipelines.
Rundown for July 5, 2026. Sources: Yellow, Anthropic, OpenAI, Google DeepMind, Mistral AI, Hugging Face.