Google Releases Gemma 4 Under Apache 2.0; Anthropic Limits Claude Access

Google drops its custom AI license and opens Gemma 4 to developers under Apache 2.0 — as Anthropic signals capacity strain by pulling Claude subscription coverage from third-party integrations.

🤖 Google Releases Gemma 4: Four Sizes, Apache 2.0 License, Edge-Optimized

Decoded: Google released Gemma 4 on April 3, the first major update to its open-weight model family in over a year. The release includes four models: Effective 2B (E2B) and Effective 4B (E4B), designed for mobile devices and edge hardware including smartphones, Raspberry Pi, and Jetson Nano; and 26B Mixture of Experts and 31B Dense, which run on a single 80GB Nvidia H100 unquantized or on consumer GPUs when quantized. The 26B MoE model activates only 3.8 billion of its 26 billion parameters during inference, delivering high tokens-per-second throughput at significantly lower compute cost than equivalently sized dense models. Google collaborated with Qualcomm and MediaTek on mobile inference optimization and claims near-zero latency on E2B and E4B on current flagship device silicon. All Gemma 4 models are released under the Apache 2.0 open-source license — replacing the custom Gemma license that had restricted commercial use since the original Gemma launch. The models are available now on Google AI Studio, Hugging Face, and Kaggle. (Ars Technica, Google official blog, April 3, 2026)

Why it matters: The Apache 2.0 license switch is the most strategically significant element of the Gemma 4 release. Google's prior custom license created procurement friction for enterprise legal and compliance teams that Meta's Llama 3 — released under Apache 2.0 — did not face, and Llama's rapid enterprise adoption followed directly from that licensing choice. By matching Llama's license terms, Google removes the primary objection that pushed enterprise workloads toward Meta's open models and away from Gemma. The 26B MoE architecture is also investor-relevant: activating 3.8 billion parameters at inference while maintaining 26B parameter quality delivers meaningfully lower inference costs per token at production scale — critical for enterprises evaluating on-premises or hybrid cloud deployment economics. For Google (GOOGL), Gemma 4 captures developer and enterprise workloads that would otherwise flow to Llama or Mistral on AWS or Azure infrastructure, redirecting those deployments toward Google Cloud and Vertex AI as the preferred fine-tuning and serving platform. Qualcomm's (QCOM) role in optimizing the edge models positions its Snapdragon platform as the preferred silicon for on-device Gemma 4 inference across Android devices.

🔐 Anthropic Pulls Claude Subscription Coverage from Third-Party Tools Citing Capacity

Decoded: Anthropic announced on April 4 that Claude subscription plans — including Claude.ai Pro and Max tiers — will no longer cover usage on third-party applications and AI tools, effective April 4 at 12pm PT. Users accessing Claude models through third-party integrations built on Anthropic's API will need to pay standard API usage fees rather than having third-party consumption count against their subscription allowance. Anthropic cited capacity management as the driver of the change. The policy separates the subscription product, which covers direct Claude.ai interactions, from API-mediated access through developer-built tools and enterprise integrations, which now route exclusively through Anthropic's metered API billing tier. (The Verge, April 4, 2026)

Why it matters: The policy change is a signal of simultaneous demand pressure across both direct consumer and developer-mediated channels — a consequence of Claude's rapid deployment inside enterprise products including Microsoft Copilot Cowork, launched March 30, and Apple's forthcoming iOS 27 Siri Extensions. Pulling subscription coverage from third-party consumption pushes developer-platform revenue onto metered API billing rather than allowing third-party app usage to drain flat-fee subscription capacity. For enterprises building Claude integrations, the shift raises per-user inference costs on productivity workflows, which will flow into pricing decisions for tools built on Anthropic's API. Anthropic is expected to seek additional funding ahead of a potential IPO; a capacity management action at this stage signals demand exceeding infrastructure scale and a deliberate move to monetize the developer channel separately from direct end-users.

Stay decoded. See you tomorrow.

— The Get AI Decoded Team