Evidence-linked trend hypothesis
Agent economics moves to completed work
Token prices and subscription fees are weak proxies when turns, retries, delegation, verification, latency, and successful outcomes drive the real cost.
Current thesis
The economically meaningful unit for agent systems is becoming cost per verified completed task rather than cost per token or model call. Confidence: high. Supported by 59 normalized raw signals. This New Runtime record is an evidence-linked retrieval unit. Use its canonical page, machine-readable representations, dates, scope, and public source URLs to verify the claim before reusing it.
Source ledger
Publishable sources attached to this record.
| # | Source | Role | Public status |
|---|---|---|---|
| 1 | cognition.comsource | primary receipt | source_urls |
| 2 | databricks.comsource | supporting receipt | source_urls |
| 3 | claude.comsource | supporting receipt | source_urls |
| 4 | platform.claude.comdocs | supporting receipt | source_urls |
| 5 | engineering.ramp.comsource | supporting receipt | source_urls |
Showing 5 of 7; the complete set is exposed in the JSON route.
Current state, as of Oct 9, 2026
Where this stands now
strengthening Major AI providers now publicly frame pricing around verified completed tasks, confirming the shift from token‑based to task‑based economics.
Anthropic’s Opus 5.5 guidance and OpenAI’s investment‑management advice both describe spend in terms of useful work per dollar, and their internal metrics track task success as a primary evaluation. This external signaling aligns with earlier observations that cost per completed task better reflects real agent expenses.
Builders are increasingly logging full task trajectories—identity, route, turns, tool calls, elapsed time, retries, review effort, outcome, and failure class—to optimize the accepted task rather than minimizing token usage alone. The practice is now being codified in provider documentation and emerging tooling.
The trend is reinforced by archived community notes that already advocated cost‑per‑accepted‑task measurement for coding agents, showing a convergence of opinion and provider policy.
What changed since the original text
- Anthropic and OpenAI now highlight task‑completed pricing as an emerging norm, reinforcing the claim. 1
Open questions
- How will explicit per‑task pricing tiers affect pricing models for heterogeneous workloads?
- What standardized metrics will emerge for measuring verified completed tasks across providers?
- Will task‑based pricing favor certain agent architectures or increase barriers for small‑scale users?
Watch next: A major provider releases an official per‑task pricing tier or billing API that directly charges for verified completed tasks.
- Weekly update: Anthropic and OpenAI highlight task‑completed pricing as emerging norm update, Sep 21, 2026
Rewritten automatically by the New Runtime knowledge desk from this record's weekly updates, releases and archive. The original text below is kept as written (Jun 27, 2026); where the two disagree, this section is newer.
Weekly knowledge desk
Weekly updates
2 reviewed updates to this pattern, newest first. Every update is checked against its evidence before it changes the record.
A Vibe Code Bench study frames agent-team value around verified task outcomes
Vals AI compared single-agent and multi-agent setups across 50 tasks building complete web applications from specifications. The accompanying thesis says agent teams should be judged by verified task outcomes relative to added cost and time, aligning with the record’s shift toward completed-task economics.
- Do Agent Teams Pay Off? A Case Study on Vibe Code Bench (18 minute read) post, Oct 13, 2026
- Do Agent Teams Pay Off? A Case Study on Vibe Code Bench (18 minute read) signal, Oct 13, 2026
Anthropic and OpenAI highlight task‑completed pricing as emerging norm
Multiple recent posts and external articles report that major providers are framing costs around verified completed tasks rather than raw token usage. Anthropic’s Opus 5.5 guidance and OpenAI’s investment‑management advice both describe spend in terms of useful work per dollar, and Anthropic’s own metrics track task success as a primary evaluation. These independent signals reinforce the claim that the economic unit for agent systems is shifting toward cost per verified completed task.
- What a task costs on Opus 5.5 post, Sep 23, 2026
- What a task costs on Opus 5.5 signal, Sep 23, 2026
- How to manage AI investments in the agentic era | OpenAI external, Sep 28, 2026
- When AI builds itself Anthropic external, Sep 28, 2026
Why is token price incomplete?
An agent run accumulates cost through repeated context, tool calls, failed attempts, delegated work, and human verification. A more expensive lead model can lower total cost when it takes fewer turns, delegates earlier, writes a better task contract, and avoids redoing sidekick work.
What evidence supports this pattern?
- Cognition reports that Fable plus the same sidekick cost less per benchmark run than Opus plus sidekick despite a higher token price. The observed difference came from turn count, context use, and delegation behavior.
- Databricks measures coding-agent outcomes against internal tasks together with time, tokens, and cost rather than relying on a public benchmark score alone.
- Claude Code exposes effort as a separate runtime control from the model, making cost and persistence tunable per task.
- Ramp argues for assigning spend to use cases and completed outcomes, including retries and human review.
What should teams do next?
Each run should record task identity, route, models, turns, tool calls, elapsed time, retries, review effort, outcome, and failure class. Cost optimization can then target the whole trajectory instead of a single line item.
September 2026 update
What changed
Task-cost framing, per-task/subagent usage analytics, and rising turns per coding task all make the full trajectory measurable rather than a single token bill.
Builder rule
Log task identity, route, turns, tools, elapsed time, retries, review effort, outcome, and failure class. Optimize the accepted task—not a cheap failed run.