Hybrid Routing Cuts AI Token Costs 30-50%

HPE's internal figures show cost-aware model routing could ease budget pressure from agentic coding tools, which are not yet a general product.

Updated on Nov. 2, 2026 in Artificial Intelligence

Isometric editorial illustration of a row of dark server cabinets in a clean data hall, with cable bundles and amber status lights, in a muted oxblood and slate palette. — Hewlett Packard Enterprise says its internal hybrid AI platform cut token costs 30-50% by routing routine work to open-weight models and escalating only complex jobs to frontier models. AI Illustration. Upload story photo >

Live Poll

Do you think AI tools are getting cheaper or more expensive for you?

HPE says its internal hybrid AI platform, which routes work between frontier models and open-weight models on its own private cloud, cut token costs by 30-50%. The company reports these figures from internal use, so the platform should be read as an in-house deployment rather than a product on general sale.

Why it matters

Agentic coding tools consume tokens far faster than chat, so per-token pricing is becoming a budget problem for enterprises. HPE's routing approach, which sends routine work to cheaper capacity and escalates only complex jobs, points toward cost-aware AI infrastructure as a central design question.

The figures come from HPE's internal reporting, with the 30-50% token cost reduction measured against running the same work without hybrid routing. A single agentic coding task can consume 10 to 50 times more tokens than a basic chat prompt, and the savings claims have not been independently verified.

The players

HPE

Hewlett Packard Enterprise, a major enterprise infrastructure company whose HPE Private Cloud AI platform hosts open-weight models for internal and customer workloads.

Marie Myers

HPE's Chief Financial Officer, who stated the private cloud cost reduction figures.

Gartner

A research and advisory firm that tracks enterprise technology markets and projected that AI coding token costs will surpass developer salaries by 2028.

Uber

The ride-hailing and delivery company that reportedly exhausted its 2026 AI budget by April after expanding its use of Anthropic's Claude Code.

GitHub Copilot

Microsoft-owned GitHub's AI coding assistant, which moved from flat-rate pricing to metered credits.

The details

The platform acts as a router that sends each request to the model best suited to it. Complex work goes to frontier models, the largest proprietary systems, while common operations run on open-weight models, whose trained parameters are publicly available, hosted on HPE Private Cloud AI. Routine tasks default to cheaper capacity, and jobs escalate to larger models only when needed. Coding agents are a major cost driver because they plan, write code, run tests, find errors and loop back to fix them, consuming tokens, the units of text a model processes and bills for, at every step.

September 2026: HPE CFO Marie Myers stated the private cloud cost reduction figures.
October 2026: HPE conducted internal analysis of hybrid platform token costs.
April 2026: Uber reportedly exhausted its 2026 AI budget after expanding its use of Claude Code.
2028: Gartner predicts AI coding token costs will surpass average developer salaries.

The Tech Race

GitHub Copilot's move from flat-rate pricing to metered credits marks the industry's broader shift toward usage-based billing for AI coding tools. HPE's hybrid routing approach is a competing response: it keeps routine work on cheaper infrastructure rather than passing metered costs directly to customers.

Teams using AI coding agents face rising per-task token bills, and the budget effects are already visible, as Uber reportedly exhausted its 2026 AI budget by April. Availability of HPE's routing approach to outside customers has not been announced, so the timeline for other organizations to adopt similar setups is unconfirmed.

The takeaway

Cost-aware routing between frontier and open-weight models is becoming a central design question for enterprise AI. Watch whether HPE publishes its cost methodology or extends the platform to customers, and whether Gartner's 2028 projection on coding token costs holds as metered pricing spreads.