AnalysisNov 19, 202611 min read5 sources

Choosing which tasks go to agents may save more than OpenAI's new discounts, but only if failed runs cost as much as successful ones

OpenAI's report puts agent success at 79% on one research task and 25% on another. If failed runs cost about as much as successful ones, that spread moves the cost per finished task more than a 40% token discount does. OpenAI has not published spend by task type, so teams have to measure it themselves.

In one minute

OpenAI's report puts agent success at 79% on one research task and 25% on another. If failed runs cost about as much as successful ones, that spread moves the cost per finished task more than a 40% token discount does.

OpenAI has not published spend by task type, so teams have to measure it themselves.

Summary

OpenAI's November 18 update on research acceleration says coding agents now finish 46% of two-to-four-hour research tasks with no human stepping in, up from 23% in March. Results vary widely by task: 79% for monitoring training runs, 25% for writing inference and serving code. The report came out in the same weeks as cheaper GPT-6.2 Astra tokens and a discounted background mode in the Responses API. Together these let a team sketch a cost per finished task, though one key input is missing: how much a failed run costs.

What happened

The update covers three parts of AI research inside OpenAI: inputs, process and outputs. On inputs, agent spend per researcher at API list prices rose 62% at the median between September and October, to $2,400 a day, and 105% at the 90th percentile, to $23,000. Agent workdays per researcher, counting eight hours of agent runtime as one workday, grew elevenfold to 5.7. On process, zero-intervention success rose from 23% to 46% on two-to-four-hour tasks and from 13% to 35% on four-to-eight-hour tasks. By task type, monitoring runs for issues reached 79%, launching training runs 35%, building evaluation, RL and agent environments 29%, and writing inference and serving code 25%. OpenAI warns that these rates mix gains in model capability with changes in what researchers chose to delegate. Researchers OpenAI interviewed said agents still need close supervision. They described agents rabbit-holing on unneeded steps, writing low-quality pull requests and making invalid assumptions in analysis. In the same weeks, OpenAI launched GPT-6.2 Astra with input and output token prices 20% below GPT-6 Astra. It also opened a public beta in which background-mode Responses API work can accept a two-hour or eight-hour target for the first response in exchange for up to 40% off Standard per-token pricing. The report does not break token spend down by task type, and it does not say how much of the spend goes to runs that failed.

What the archive adds

New Runtime's earlier coverage of the GPT-6.2 launch listed API prices per million tokens: Astra at $8 input, $0.40 cached input and $40 output, and Sol at $2, $0.10 and $10. That coverage advised readers to compare cost per completed task on their own prompts rather than headline benchmark quality. An earlier New Runtime note on the background-mode beta pointed out that "up to 40%" is a ceiling, not a guaranteed rate. OpenAI's API pricing page now gives the split: 30% off for a two-hour window and 40% off for an eight-hour window on eligible requests. Google's Gemini pricing page lists Batch and Flex inference at 50% off standard. Artificial Analysis measured Intelligence Index cost per task at $4.00 for GPT-6.2 Astra (max) and $1.01 for GPT-6.2 Sol (max), but those are benchmark-suite costs, not costs of production research tasks. On self-reports, METR's 2025 study found that experienced open-source developers believed AI had sped them up by about 20% when it had not. METR has since published an independent assessment of AI R&D acceleration at Anthropic dated November 2026, so outside, task-level checks of lab claims are starting to appear.

Trend signal

Labs are moving from benchmark scores to measures of how much work agents actually do: agent workdays, success rates by task length, and spend per researcher. Pricing is moving in the same direction. Cheaper output tokens, cached-input discounts and deferred completion windows all make long, unattended runs cheaper. What OpenAI's report adds is a public admission that progress is uneven. Task types succeeding 79% of the time and 25% of the time sit in the same research loop, and OpenAI itself says the weaker ones could become bottlenecks. Price lists, meanwhile, are still quoted per token and do not depend on task type. Vendors report success rates by task type, but cost by task type is still not reported. For buyers, cost per task type is the number that matters.

Why now

Three things are driving this now. First, success on long tasks crossed a usable threshold this year: OpenAI reports zero-intervention success on two-to-four-hour tasks rising from 23% to 46% between March and October, so pricing multi-hour delegation is now worth doing. Second, OpenAI lowered the per-unit cost of long runs. GPT-6.2 Astra tokens are 20% cheaper than GPT-6 Astra, and OpenAI's pricing page lists 30% and 40% discounts for two-hour and eight-hour background windows, against Google's listed 50% for Batch and Flex. Third, OpenAI says labs should publicly report indicators of recursive self-improvement, so publishing spend and success data serves a governance aim alongside a commercial one. That also helps explain why the report gives success by task type but not cost by task type.

Connections to other events

The worked example below is a scenario with stated assumptions, not a measurement. Assume three things: every agent attempt at a task costs roughly the same in tokens; a failed attempt is simply retried; and no human time is counted. Under those assumptions, the expected number of attempts per finished task is 1 divided by the success rate:

  • monitoring runs (79%): about 1.3 attempts;
  • building evaluation, RL and agent environments (29%): about 3.4 attempts;
  • inference and serving code (25%): about 4 attempts.

That is a gap of roughly 2.7x to 3.2x in agent spend per finished task between the strongest and weakest of these task types.

The background-mode discount cuts the per-token price to 0.7x for the two-hour window or 0.6x for the eight-hour window, according to OpenAI's pricing page. That is at most a 1.67x saving, and it applies only to work that can wait hours. So in this scenario, choosing which tasks to delegate moves cost more than choosing a price tier.

Two unknowns can overturn this either way. First, failed runs may cost less than successful ones because they stop early, or more because agents rabbit-hole, which OpenAI's interviewees report. Second, human review time is left out entirely. In practice a failure is often handed to a person rather than retried, and an hour of that person's attention may cost more than the tokens. OpenAI publishes neither number. So our conclusion is conditional: task routing probably matters more than discounts, but nobody outside OpenAI can confirm that from what has been published.

Why it matters

For builders, the most useful unit is cost per finished task, broken down by task type, not spend per seat or per day. Tasks that score well and can wait, such as run monitoring, scanning data and answering technical questions, are where delegation is cheapest and where background windows add savings on top. Tasks that score poorly, such as serving code and environment building, are better run as one agent draft followed by human review, not as open-ended retry loops. Buyers should ask vendors for success rates and token spend by task type on their own workloads.

The strongest counter-argument is selection. OpenAI says its success rates partly reflect what researchers chose to delegate. The 25% on serving code may mean researchers send their hardest serving problems to agents, not that agents are weak at serving code in general. If so, a team that delegates easier serving work could see much higher success, and the task map would not carry over to other teams. The way to separate selection from capability is an outside, task-level test: a fixed set of tasks from each category, run monthly without anyone choosing what goes in, recording success and token spend. If that test shows serving-code and monitoring tasks scoring about the same, the gap is about selection and workflow, and routing by task type matters less than our scenario suggests.

What happens next

  • OpenAI's next research-acceleration update will still not publish token spend or cost broken down by task type. (3 months, 75%). Watch for: Any OpenAI post or methods note that reports dollars or tokens per O*NET task category. Resolves on 2027-02-19, yes if: No post on openai.com dated between 2026-11-19 and 2027-02-18 reports agent token spend or dollar cost per research task type; checked on 2027-02-19.
  • If OpenAI publishes another research-acceleration report within three months, serving-code success will still be below 45%. (3 months, 35%). Watch for: A new OpenAI acceleration post with an updated O*NET task chart. Resolves on 2027-02-19, yes if: openai.com lists a research acceleration report dated between 2026-11-19 and 2027-02-18 that gives zero-intervention success below 45% for writing code for inference and serving stacks; checked on 2027-02-19.
  • OpenAI will not raise its background-window discount above 40% within three months, leaving it below Google's 50% Batch rate. (3 months, 75%). Watch for: Any change to the background-mode beta terms, or the windows leaving beta. Resolves on 2027-02-19, yes if: The OpenAI API pricing page lists no background or deferred-window discount above 40% for GPT-6.2 Astra on 2027-02-19.

Limits and uncertainty

The cost multiples above depend on three unverified assumptions: attempts cost the same across task types, failures get retried, and human review is excluded. OpenAI does not publish token spend by task type, the cost of failed runs, or review time. Success rates mix capability with delegation choices, as OpenAI notes, so they may not apply to teams with a different task mix. All figures come from inside OpenAI, using its own harnesses and models, and no outside party has checked them. The background discount split (30% and 40%) comes from OpenAI's pricing page, and the effective discount on mixed workloads is not public. Success rates are also rising fast: four-to-eight-hour success went from 13% to 35% between March and October. Any task-type map may be out of date within a quarter, which is a reason to re-measure, not to assume the gap is permanent.

Competing readings

  • For outside teams, which tasks go to agents probably moves cost per finished task more than the price tier does, provided failed attempts cost about as much as successful ones (likely, about 55%). What would tell it apart: Success ranges from 79% to 25-29% across task types in the same research loop (openai.com); The background discount saves at most 1.67x per token and applies only to work that can wait (x.com)(openai.com).
  • The task-type gap mostly reflects what researchers chose to delegate and how they work, not what agents can do, so it says little about cost for outside teams (unlikely, about 30%). What would tell it apart: OpenAI says success rates mix capability with delegation choices (openai.com); Researchers describe changing how they delegate after being burned, which shifts the task mix (openai.com).
  • Per-token price and discounts are the main cost lever, because failed runs stop early and cost little (very unlikely, about 15%). What would tell it apart: Spend is growing fast, so per-token discounts compound across large volumes (openai.com)(x.com); Average success is rising quickly, which shrinks the number of retries (openai.com).

Base rate

Of 10 earlier research stories about the same companies or product lines in the New Runtime archive over the last 12 months that are at least 3 months old, 10 (100%) were followed within 3 months by a release story on the same company or product line. In New Runtime's archive, 10 of 10 earlier research stories on the same companies were followed by a release story within 3 months. Ten cases is a thin sample, and it measures release cadence, not whether delegation pays off, so it does not bear on this conclusion.

What would change this view

  • OpenAI or METR publishes token spend per task type showing failed runs on serving code or environment building cost far less than successful monitoring runs (signal): The same-cost-attempt assumption fails, and the gap in cost per finished task is much smaller than about 3x.
  • An independent test on a fixed task set, with no one choosing which tasks go in, finds serving-code and monitoring success within 15 points of each other (model-listing signal): The jagged map mostly reflects delegation choices, and routing by task type matters less than this take argues.
  • OpenAI's pricing page raises the background-window discount to 50% or more, or cuts Astra list prices again (pricing signal): Choice of price tier could rival choice of tasks as the main cost lever.

Skill to build now

Measuring cost per finished task for agent work, broken down by task type. First exercise: run a one-week pilot. For every agent run, record the task type (monitor, analyze, build, launch, communicate), total token spend, number of retries, minutes of human review, and whether it finished with no intervention. At the end of the week, divide spend plus costed review time by the number of finished tasks for each type. Only then pick which types to move to background windows and which to keep with a human.

Reader takeaway

Before moving more work to agents or onto background windows, run a one-week pilot. Tag each agent run by task type, and log total token spend, retries, human review minutes and whether the run finished without intervention. Move work only after you can rank task types by full cost per finished task. Until your own data says otherwise, send failures on low-success types to a person after one attempt.

Open question

Is the 25% success on serving code a real capability gap that will close as models improve, or is it mostly a product of researchers sending agents their hardest serving problems, so that outside teams would see a much smaller gap?

Evidence / sources

  1. 1openai.com/index/research-acceleration-jagged-frontier
  2. 2openai.com/index/introducing-gpt-6-2-astra-sol
  3. 3deploymentsafety.openai.com/gpt-6-cyber
  4. 4x.com/OpenAIDevs/status/2123188195202617408
  5. 5t.me/qwgai/4338
All Analysis

Next reads

Read next

Open the graph
  1. 01topicAgentsExplore the agents topic hub.
  2. 02archiveAnalysisOpen the latest authorial analysis.
  3. 03sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract