Export limits, cheaper inference and coding agents that must ship

Alternatives to restricted Anthropic models

Restrictions on Anthropic's models appear to be creating openings for labs outside the US. Sakana AI released Fugu, a coding model that scored 93.2 on LiveCodeBench, days after Claude access was restricted in certain markets. TechCrunch reports that several Asian startups have launched models positioned as alternatives to Anthropic's Mythos line as the export ban continues. For teams outside the US that lost Claude access, Fugu may be worth evaluating for agentic coding, though one benchmark score says little about how it holds up on real workloads.

Budgets keep rising as token costs bite

Enterprise budgets are growing, and so is attention to what each token costs. An RBC survey of CIOs found more than half already run AI in production, with budgets still growing despite concern about per-token costs; OpenAI reportedly leads enterprise deployments. Even AI vendors lean on rivals: Salesforce employees are reportedly puzzled that the company promotes Anthropic's Claude inside Slack. Further up the stack, Meta, Microsoft and other hyperscalers have pushed combined AI data center commitments past $850 billion, a buildout that suggests accelerator supply may stay tight while capacity comes online.

Two releases attack cost from different angles. DeepSeek open-sourced DSpark, an inference framework it says speeds up LLM serving by up to 85% through orchestration-level changes rather than model architecture, which should make it applicable to many self-hosted stacks. Cognition took a routing approach: Devin Fusion sends each task to a frontier or cheaper model depending on the job, cutting costs 35% on Cognition's own FrontierCode benchmark. Both figures come from the developers' own measurements.

Coding agents meet review, maintenance and supply chains

Faster code generation does not automatically mean faster shipping. InfoQ's reporting around Xcode 27's agent device hub notes that AI tools speed up writing code while delivery timelines lag, with the bottleneck moving to review, testing and integration. Evaluation is moving toward that harder work: RoadmapBench tests agents on long-horizon development across version upgrades, closer to everyday maintenance than one-off tasks. Training has a related limit, and an essay on extending RL beyond verifiable tasks examines what happens when no test suite supplies a reward, as in writing, research or long-horizon planning.

As agents gain autonomy, the controls around them matter more. A researcher showed that Claude Community Plugins were exposed to repo-jacking, and commit SHA pinning prevented exploitation, a concrete case for pinning plugins by SHA rather than branch or tag. Okta launched Okta for AI Agents Core, which it calls the first independent identity platform for agent lifecycle governance in regulated environments. For coordination between agents, the open-source Ablo project aims to standardize how agents share context and hand off tasks.

Google's moves on data, science and pricing

Google is pushing on several fronts at once. Google Cloud proposed an Open Knowledge Format for sharing structured, AI-ready data, intended to let LLM pipelines use enterprise knowledge bases without heavy custom ETL; as a proposal, its value depends on adoption. Google Cloud also plans to sell specialist science models built with SandboxAQ, a domain-specific line alongside general foundation models. On the consumer side, Gemini's personalized image generation is now free for US users after previously sitting behind a paid tier.