← /editions
#66

Gartner puts a 5x number on agentic inference costs

// Gartner forecasts inference cost per agentic workflow will rise more than fivefold through 2028 — the Inference Paradox, where falling token prices mask escalating total AI spend. Plus: GLM-5.3 tops a cyber benchmark, Atlassian Rovo can be tricked into exfiltrating Jira and Confluence data, and AWS closes Bedrock Agents to new customers.

$ today_in_10_seconds

Gartner published a number this morning that every team shipping agents should paste into their finance model: inference cost per agentic workflow rises more than fivefold through 2028. Token prices keep falling, total bills keep climbing, and the gap between those two facts is where most agent margins go to die.

$ lead_story

Gartner puts a 5x number on agentic inference costs

Gartner forecasts that AI inference costs per agentic workflow will increase more than fivefold through 2028, as products shift from assistive features to multistep execution — with falling model prices quietly subsidising more complex workflows while total AI spend escalates. The firm has a name for this: the Inference Paradox, where improving unit economics drive up overall AI cost without a clear path to predictable value. The mechanism is the agent loop itself. A chatbot reads a query and answers; an agent reasons, negotiates, and second-guesses itself — and routing a task to an agentic reasoning model costs providers at least five times a basic interaction, often far more as complexity grows. Gartner's prescription is not a cheaper model: analyst Will Sommer argues there is no reliable one-size-fits-all model coming, and competitive AI products will require maintaining complex multimodel ecosystems, with heavily optimised inference tiering, routing and orchestration.

// why_it_matters: Your unit of cost is no longer the prompt — it's the completed task, and nobody's pricing page is denominated that way.

read more →

$ also_worth_attention

  • GLM-5.3 tops a cyber benchmark — and reportedly found a flaw in Cursor

    Z.ai shipped GLM-5.3 on August 14 with an 84.5% CyberGym score, edging past Mythos 5 at 83.8% and GPT-5.6 Sol at 83.6%, while ExploitBench more than doubled to 54.4% — gains Z.ai attributes entirely to security-focused post-training RL on an unchanged base model. The Cursor finding, described as an architectural risk in its Electron-and-Rust codebase permitting arbitrary file writes, was disclosed privately but has no published CVE and no confirmation beyond the vendor's own account. Treat the benchmark as marketing and the capability jump as real: post-training now moves faster than your model evaluation cycle.

    link →
  • Hidden text in a PDF turns Atlassian Rovo into an exfiltration tool

    PromptArmor showed that concealed instructions in an uploaded document can make Rovo gather Jira tickets and Confluence pages, append them to an attacker's URL, and open it — no human approval step, and it works even with org-level web search disabled, because the URL-retrieval tool has no protection against agent-constructed URLs. A separate link-based path found by Varonis was closed server-side; the content-based route was still working when PromptArmor published on August 5. Neither disclosure carries a CVE. The lesson generalises past Atlassian: any tool that can construct a URL is an egress channel.

    link →
  • AWS closes the door on Bedrock Agents

    Bedrock Agents is now Bedrock Agents Classic, shut to new customers as of July 30 and frozen in maintenance mode — the model catalog stops updating, and AWS points net-new work at AgentCore. It was part of a broader pruning of roughly 20 AI and ML services including Kendra and Q Business, and Bedrock Agents lasted two years and eight months. AgentCore is a re-architecture, not an upgrade path, which makes the real question uncomfortable: what's the assumed shelf life of the agent platform you picked this quarter?

    link →

$ bigger_picture

The trend behind the headlines

Three bills arrive with agents, and today's stories are one of each. Inference: the loop multiplies your token spend faster than per-token prices fall. Security: every tool you grant an agent is a potential egress path, and the Rovo chain shows that turning a feature off isn't the same as closing the capability. Platform: the orchestration layer underneath you has a shelf life measured in single-digit quarters. What connects them is that all three are consequences of the same design choice — giving a model the ability to act repeatedly, with tools, unsupervised. The teams that hold margin will be the ones treating the agent loop as a metered, bounded, audited piece of infrastructure rather than a feature flag.

$ tool_of_the_day

LiteLLM

The open-source gateway that makes Gartner's advice implementable. It fronts 140+ providers behind one key, tracks spend per key, user, team, org, tool, agent and MCP server, and enforces hard budgets that stop requests at the cap — plus rate limits and leaked-key protection so a runaway agent loop can't quietly bill you all weekend. Route cheap tasks to small models, gate frontier reasoning behind a policy, and get per-agent cost attribution you can actually defend in a business review.

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