← /editions
#145

Gartner Says Open Models Handle 90% of Agentic Work at 1/20th the Cost

// Gartner says open models can handle about 90% of agentic work at roughly a twentieth of the cost of a frontier LLM, which makes routing between cheap and frontier models the default agent design.

$ today_in_10_seconds

Gartner says open models can handle about 90% of agentic work at roughly a twentieth of the cost of a frontier LLM. If that holds, production agents should run on a cheap model by default and escalate to a frontier model only when needed, instead of using one big model for everything. It's a one-story day, and this is the story that matters.

$ lead_story

Gartner Says Open Models Handle 90% of Agentic Work at 1/20th the Cost

Gartner, as reported by The Register, puts a number on what many agent builders have suspected: open models can cover about 90 percent of agentic work at roughly a twentieth of the cost of a frontier LLM. Agents burn through tokens far faster than chat. Every plan, tool call, retry, and reflection step is another inference, so per-token pricing compounds. A 20x cost gap on most of those steps decides whether an agent stays a demo or makes financial sense at scale. Builders should stop treating model choice as a single decision and start treating it as a routing problem.

// why_it_matters: For agents, cost per step is the product, and Gartner says most steps don't need frontier pricing.

read more →

$ also_worth_attention

  • The real question is which 10%

    A headline number like '90% of agentic work' hides which tasks fall in the other 10%, and that's where your agent breaks. Long-horizon planning, ambiguous instructions, and recovering from failed tool calls are where smaller models tend to stumble. Test open models on those cases before assuming the 90% covers your workload.

    link →
  • Routing becomes core infrastructure

    If most steps run on cheap models and a minority escalate, the router becomes the most important part of your agent stack, more than any single model. Testing whether your agent escalates at the right moments now matters as much as testing its output quality.

    link →
  • Self-hosting changes your agent's risk profile

    Open weights let you run agents inside your own infrastructure, pin model versions, and stop worrying about a provider silently changing behavior mid-workflow. The trade-off is that you now own serving, latency, and upgrades.

    link →

$ bigger_picture

The trend behind the headlines

AI models are commoditizing the way most technology does. New capability appears first in the most expensive frontier models, then spreads to cheaper ones fast enough that the frontier only needs to handle the hardest cases. Now that Gartner has put a number on it, CFOs have cover to ask why every agent call goes to the priciest model. For founders, 'we use the best model' was never a durable advantage. The advantage lies in how you coordinate models, how you test them, and the data your agents accumulate. Teams that built to swap models easily from day one get a cost cut for free, while teams locked into one provider face a migration project.

$ tool_of_the_day

LiteLLM

LiteLLM is an open-source, OpenAI-compatible gateway that sits in front of 100+ model providers, including self-hosted open models, with built-in fallbacks and spend tracking. It's the fastest way to test Gartner's claim on your own workload: point your agent at it, send the easy steps to an open model, and compare cost and success rate.

explore LiteLLM →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime