Cloudflare Wants AI to Do More Without Running Up the Bill

Morning, {{first name | folks}}! AI is advancing on several fronts this week, from Cloudflare’s push to cut the cost of routine AI tasks to OpenAI’s progress on legacy code. Meanwhile, Microsoft is tackling security risks from autonomous agents, Amazon is making its AI chips easier to adopt, and Exa is stepping up the race to build search engines for AI.

Cloudflare has launched Clairic, a model designed to handle routine tasks like pulling figures from invoices, calculating totals, and estimating costs. Instead of using a large language model for every step, Clairic returns the values directly, without generating a full text response.

In Cloudflare’s tests, a workflow using Clairic alongside its Clef decision model cost 49% less and ran about 3.2 times faster than using a language model throughout. Clairic is available in preview on Workers AI, with open model weights planned for release.

OpenAI is improving its models’ ability to write and migrate code in older languages like COBOL and Fortran, which still run critical business systems. Modernizing this software can be costly and complex, especially when documentation is incomplete and decades of business logic are buried in the code.

In OpenAI’s COBOL migration test, GPT-6.1 Sol reached 97.9% accuracy at higher reasoning effort, compared with 5.4% for GPT-5.6 Sol. Better AI support could help businesses upgrade aging systems without having to start from scratch, though real-world migrations still need expert review and testing.

Microsoft is warning that AI agents could create a new security headache for businesses as they gain access to company data, internal tools and critical systems. Unlike traditional cyberattacks, the risk may come from an agent carrying out an action it was allowed to take but that turns out to be harmful.

Microsoft is developing Guardian Agents to watch how AI agents behave, spot risky actions and step in when needed. As companies give AI more responsibility, the challenge is making sure these systems can act independently without putting sensitive information or business operations at risk.

Amazon is making its Trainium chips easier to try for developers already using PyTorch. Its new TorchNeuron tool lets them run existing models on Amazon’s hardware with a configuration change, rather than rewriting code to fit a different chip.

Amazon plans to open-source the tool, giving developers a way to test Trainium without rebuilding their existing workflows. The move tackles one of the biggest barriers to trying alternative AI chips: the time and effort involved in moving software to new hardware. For AWS, it’s another push to attract teams that have built their AI infrastructure around Nvidia.

Exa has rolled out Exa 3.0, an upgrade to its search engine designed for AI agents. The company says it beat Perplexity, Brave, and Parallel across eight benchmarks, including tasks involving hard-to-find information, coding and financial research.

Alongside the launch, Exa introduced Exa Lite at $4 per 1,000 requests and opened up its Dynamic Highlights feature to more users. The feature cuts the text sent to AI models by up to 95% compared with full-page content, helping developers save on costs when building agents that search the web at scale.