METR's frontier time horizons are doubling every few months, providing substantial evidence that AI will soon be able to automate many tasks or even jobs. But per-task inference costs have also risen sharply, and automation requires AI labor to be affordable, not just possible.[1] Many people look at the rising compute bills behind frontier models and conclude that automation will soon become unaffordable.
I think this misreads the data. The rise in inference cost reflects models completing longer tasks, not models becoming more expensive relative to the human labor they replace. Current frontier models complete tasks at their 50% reliability horizon for roughly 3% of human cost, and this hasn’t increased as capabilities have improved. So, cost isn't an additional bottleneck beyond capability, and we should expect to see automation at roughly the time predicted by METR's capability trendlines.
I define cost ratio as the inference cost of the average AI trajectory that solves a task divided by human cost to complete the same task. Using METR’s data, I examine the trend in cost ratios over time. I show three things:
Across successive frontier models, the cost ratio at each model's 50% reliability time horizon hasn't increased.
Among tasks models successfully complete, longer tasks don't have higher cost ratios than shorter ones.
Time horizon trends barely slow when capping AI spending per task at a fraction of human cost (doubling every ~3 months, even at 1/32x).
Raising the cost ratio by providing additional compute might translate into longer time horizons, but that would be in addition to, rather than causing, current trends. So, compared to extrapolating METR’s data, increased inference-time spending will strictly shorten AI timelines.
In his article on this subject, Toby Ord came to the opposite conclusion; he argued that there is moderate evidence that “the hourly costs are rising exponentially” at the frontier of AI capabilities. In appendix A, I argue that the methodology Ord used to produce this conclusion is unreliable and leads to significant overestimates of hourly model cost.
Thanks to Toby Ord, Buck Shlegeris, Agustín Covarrubias, Alex Mallen, Alexa Pan, Tim Hua, Aniket Chakravorty, Arun Jose, and Francis Rhys Ward for helpful comments on earlier drafts. Thanks especially to Ryan Greenblatt for the ideas that inspired this article.
All figures in this article were generated with code that can be found here.
Evidence from METR
In this post, I analyze publicly available time horizon data from METR. I calculate cost ratio by estimating AI inference cost from token count (using OpenRouter pricing) and dividing by the human cost for each task provided by in the dataset. All calculations are performed using METR’s task weighting. I made some modeling assumptions to fill gaps in the METR dataset (see footnote).[2] These assumptions shouldn’t affect my overall conclusions, because they shouldn’t significantly affect the trend across models.
My analysis has important limitations, which I discuss after presenting evidence.
Cost ratio at models’ 50% time horizon isn’t increasing
To investigate whether cost ratios rise as models improve, I examine the cost ratio at each model's 50% reliability time horizon. For each model, I select tasks that the model successfully completes with lengths between 0.79 times and 1.29 times the calculated time horizon (which is plus or minus 0.1 orders of magnitude).[3] I graph the median cost ratio in the selected set of tasks.
Figure 1: Median cost ratio of successful completions near each model's 50% time horizon (between 0.79x and 1.29x the time horizon), with interquartile range. As time horizons increase from minutes to hours across successive models, cost ratios remain well below 1 (dashed line), with no upward trend.
Cost ratios show no upward trend across models, and remain well below 1. As models' 50% time horizon increases exponentially, companies can cheaply use them to complete longer tasks.
I filter for passing tasks to show that cost isn't an additional limitation on top of capabilities. If successful tasks are cost-effective, then companies can set a low cost cap and still capture most successes. One might be concerned that this underestimates costs, because models tend to succeed at cheap tasks. But even at the 80% time horizon, where models complete the majority of tasks, cost ratios remain well below 1 (figure B3). Including failures doesn't change this (figure B4).[4] This means companies can profitably automate a broad share of tasks, not just a narrow cheap subset.
Time horizons improvements aren’t driven by expensive long tasks
One might respond to figure 1 by arguing that even if cost ratios at each model’s 50% time horizon aren’t increasing, improvements to those 50% time horizons come from successes on long tasks with high cost ratios. If this were true, we'd expect cost ratios to rise with task length among successfully completed tasks. I show that’s false: among successful tasks, cost ratios don't rise with task length.[5]
Figure 2: Cost ratio by task duration for successful attempts only (excluding tasks under 1.5 minutes), across a selection of models. The shaded region shows the weighted interquartile range. Cost ratios do not increase with task length, they decline, though this likely reflects a selection effect where models only succeed on relatively cheap long tasks.
This doesn’t mean models are more efficient at longer tasks. There's a strong selection effect here: models pass far fewer long-horizon tasks, and likely only succeed on those that happen to be relatively cheap. But since only successes push the time horizon forward, what matters is whether longer successful tasks are expensive, and the data shows they aren’t.
Progress at a fixed cost is just as fast
Another way to examine the affordability of AI progress is to ask how much progress slows if companies cap AI spending at a fraction of human cost.
I model the affordable 50% time horizon, the maximum task length at which models have a greater than 50% chance of both completing a task and doing so under a given cost ratio. If progress is fast at a low cost ratio, then we can expect that AIs will soon be able to cheaply complete long tasks. I test a range of caps, from 1/4x to 1/32x human cost, to see whether tighter budgets substantially slow the trend.
I model affordability and passing separately. I fit a logistic curve to the pass rate as a function of task length, and a separate logistic[6] to the probability that any passing task is affordable. I do this because pass rate falls with task length, but affordability increases (see figure 2). Fitting a single logistic curve would fail to capture these opposing trends.
Multiplying the pass rate logistic and affordability logistic gives the probability that a task at some length will both pass and be affordable:
The affordable 50% time horizon is the longest task length at which p(pass & afford) is greater than or equal to 50%.
Figure 3: Factored model for Claude 4 Opus at four cost thresholds (1/32x, 1/16x, 1/8x, 1/4x human cost). Red dots show failed or over-cost tasks, while green dots show passed and affordable tasks. Each panel decomposes the affordable pass rate (red) into pass probability (blue) and affordability given passing (green dashed). The dashed vertical line marks the affordable 50% time horizon. Tighter cost constraints reduce the horizon only modestly.
At different maximum cost ratios, the doubling times are reliably similar.
Figure 4: Affordable 50% time horizon over time after 2024 at four cost thresholds: unlimited, 1/4x, 1/8x, and 1/32x human cost. Doubling times range from 3.0 to 3.3 months across thresholds, with high R² values throughout. Stricter cost constraints filter out some models but do not substantially slow the trend.
Stricter cost thresholds filter out some models, but enough recent models have cleared even the tightest thresholds to keep the doubling time around 3 months.
Limitations of my methodology
This data has some important weaknesses for the analysis I performed above:
AIs may disproportionately succeed on cheaper tasks at any given length, which would mean the cost ratios I observe reflect the cheapest subset rather than a representative workload. I minimize the effect this has on my results by focusing on trends in cost ratios rather than absolute levels, and by examining cost ratios at higher reliability thresholds where the selection effect is weakened.[7]
Models’ primary objective in these benchmarks is to complete tasks rather than minimize costs. My approach assumes these tasks were completed at the minimum cost, but some tasks may have been solvable at lower cost (e.g., with different prompting). As a result, I might significantly overestimate the costs of some tasks compared to scaffolding or prompting to optimize for cost.
METR notes that scaffolding differences make cost comparisons between models difficult. Unfortunately, I couldn’t find data that holds scaffolding fixed, so there isn’t an easy way to correct this. However, scaffolding differences would need to change costs by manyfold to produce a noticeable effect on the cost trend.
AIs might have advantages on benchmark tasks because of overfitting, which could drive down absolute costs.[8] But this would only affect the trends I observe if overfitting disproportionately reduces costs at longer or shorter time horizons, and I see no reason to expect that.
I think that these may mildly influence my results, but are unlikely to change overall trends.
Inference scaling will just make progress faster
If companies are willing to increase cost ratios (to 1, or even above), they might unlock even better capabilities. AI might be faster or higher quality than human labor in some domains, so the additional cost may be worthwhile.
UK AISI showed that newer models have larger gains from additional inference compute. In some domains, like ARC-AGI, performance improves a lot from large amounts of test-time compute, as demonstrated by o3’s performance at different cost thresholds and Ryan’s results . METR shows that additional token budget can improve a model’s 50% time horizon (note that this is an upper limit on per-task tokens, so not analogous to cost ratio), although with the Claude Code scaffold this is fairly limited.
Over time, we should expect the gains from additional inference compute to grow as models improve at using test-time compute. Already, AI providers offer increased test-time compute at a cost premium, like GPT-5.4 pro and extended thinking with Claude. As companies automate more labor, they might increasingly opt for these premium tiers over base models.
It's not yet clear how much improvement is possible by increasing test-time spending, but given the relatively low cost necessary to achieve METR's time horizons, substantial increases may be economically viable. Any gains would come on top of the trends observed by METR.
Conclusion
The data shows that observed increases in AI agentic capabilities have not come from rising cost ratios. METR's trendlines predict the length of tasks AI can complete cheaply.
I think companies will be able to affordably deploy models to complete tasks at their 50% or 80% time horizons. At these horizons, many tasks will fail, but at ~3% cost ratio companies can cap spending per attempt well below human cost, retry identified failures, and still profitably automate. And, more speculatively, if companies can predict which tasks will succeed, they can selectively automate those.
In some domains, it might be possible to increase cost ratios to make AI agents much more capable, but this possibility strictly shortens timelines to any capability milestone, rather than lengthening them. The first AIs with human-level intelligence may cost more than human labor, but cheap human-level AIs should not be far behind.
Appendix A: Why I get different results from Ord
Toby Ord analyzed the same issue and found very different results. I think this is explained by how he calculated hourly cost, which leads to significant overestimates. I also explain more minor reasons to prefer my analysis to his at the end of this appendix.
Ord’s results come from his analysis of this graph, released by METR (with his annotations).
Figure A1: Comparison of time horizon achieved by token cost used for different models. Each model is run once with a constant token budget, and the performance at each cost threshold is calculated by counting all unfinished runs as failures. Costs for o3 and GPT-5 are approximated using cost data for OpenAI o1. Toby Ord has added saturation point annotations to the figure, with diagonal lines indicating a flat hourly cost.
Ord uses this graph to find the hourly cost of AIs at their “saturation point” (where increasing the total budget doesn’t significantly increase the model’s 50% time horizon).
I don’t think Ord’s methodology calculates the real hourly AI cost for any set of tasks. Ord calculates hourly cost using the following method for the saturation point:
Where the per-task budget is the maximum that can be spent on any individual task.
For example, my data shows GPT-5 succeeds on 1-2 hour tasks at a median cost of $1.29, but Ord’s method requires a $10+ budget to reach a time horizon in that range. The reason is that the 50% time horizon at the saturation point is influenced by passing tasks significantly longer than the horizon itself (see figure A2). So the per-task budget must be large enough to pay for the most expensive task necessary to reach the saturation point, which has a high absolute cost even if its hourly cost is low. Dividing this inflated budget by the shorter 50% time horizon produces an inflated hourly cost. In effect, hourly cost is set by the price of passing a 4-hour task, divided by a 1.6-hour horizon.
Figure A2: As the cost cap rises, GPT-5's horizon grows from ~30s to ~3.5h. The bottom panel decomposes this growth by counterfactual attribution: at each cost step, I remove each duration bucket's newly-passed runs and refit the logistic to measure its share of the p50 shift. Early growth (under $1) is driven by short tasks (<4m, 4-16m). Above $10, 4-16h tasks (purple) dominate; expensive successes on very long tasks are what push the horizon from ~2h toward its ~3.5h plateau. (My updated data has a different time horizon than in Figure A1, but it matches METR’s updated time horizon.)
Figure A3: Dividing each model's required budget by its time horizon gives an implied hourly rate. Ord's method implies $5–59/hr across models (red) at their overall 50% time horizon, but the actual weighted mean and median per-task hourly costs of successful tasks near each model's horizon duration are much lower (orange, green). The overstatement ranges from 9–64x against the median. Error bars show 90% bootstrap confidence intervals. I have ignored saturation for simplicity, and simply find the cost necessary to reach 95% of a model's unrestricted 50% time horizon.
This leads to significant overestimation of absolute costs, possibly by an order of magnitude or more, and it is sufficiently noisy (overestimation from 9x to 64x) that it's also difficult to draw conclusions about trends.[9]
I think there are other, more minor reasons to prefer my analysis to Ord’s:
Even if cost ratios are high for very long tasks (far beyond models’ 50% time horizons), companies could choose not to use models for these tasks. The more relevant metric is the cost of tasks at models’ 50% or 80% time horizons, which I show are low in figure 1 and appendix B.
Ord’s finding of an upward trending in cost ratios largely relies on o3 and GPT-5 having higher hourly cost than other models. But in figure A1, the token cost of these models is approximated with the token cost of o1, a particularly expensive model.
Ord’s analysis uses fewer and older models than my analysis, and so may be less reflective of trends.
Appendix B: Additional cost at time horizon graphs
50% time horizon
I show that the overall trend of cost ratio versus time horizon doesn’t depend much on whether I include tasks with time horizons further than the model’s 50% time horizon.
Figure B1: Same as Figure 1 but with a narrower task selection band (±0.05 OOM). The trend is unchanged.
Figure B2: Same as Figure 1 but with a wider task selection band (±0.2 OOM). The trend is unchanged.
80% time horizon
Figure B3: Same as Figure 1 but measured at each model's 80% time horizon. Cost ratios remain below human cost with no upward trend, though o1-preview and o1 are notably more expensive at this threshold.
Including failures
Figure B4: Same as Figure 1 but including failed attempts in the cost calculation. Median cost ratios shift upward for reasoning models (o1-preview, o1, o3) that spend heavily even on failures, while most other models remain largely unchanged.
Appendix C: 80% affordable time horizon
Figure C1: Same as Figure 3 but at the 80% reliability threshold. Doubling times (2.8–3.1 months) are similar across cost thresholds, though stricter thresholds filter out more models, leaving fewer data points.
For some applications (e.g., physically dangerous or secretive ones), AI might be used to replace humans even if it is more expensive than them. I still think AI labor being expensive would substantially slow adoption in many areas.
For METR's data, I don't have exact inference costs for most models, so I estimate cost from each run's total token count multiplied by the simple average of that model's per-token input and output prices on OpenRouter. Note that OpenRouter prices can differ from direct API prices (e.g. Claude 3.5 Sonnet is $6/$30 per million tokens on OpenRouter vs $3/$15 on Anthropic's API). Because the data only records an aggregate token count (no input/output breakdown), this assumes a roughly equal split between input and output tokens. I tested this assumption against a few tasks for which METR recorded the exact cost and found it to be a reasonable approximation. Pricing was not available on OpenRouter for a few models: for Claude 3 Opus I used Anthropic's historical API pricing, for o1-preview and Claude 3.5 Sonnet (Old) I substituted pricing from their current-generation equivalents (o1 and Claude 3.5 Sonnet (New), respectively). Some runs for GPT-4 0314 and o1-preview lack token counts and are excluded from the cost analysis.
Failed attempts are harder to interpret, since a failure doesn't tell us how many tokens would have been needed to succeed, but the absence of any upward trend even with failures included suggests selection isn't masking rising costs.
I exclude tasks under 1 minute 30 seconds because they are predominantly run on a different scaffold (swaa/generate) that is less comparable to the agentic scaffolds used for longer tasks. I also do this for figures 3 and 4.
To avoid selection effects, one might compare models only on tasks they all complete successfully. But weak models only complete very short tasks, so this would only tell us about costs on 5-minute tasks. I'm instead interested in costs at the frontier of each model's capabilities.
Ord also calculates “sweet spots”, where models are maximally cost-effective, using a similar methodology. Although the effect might be smaller, I suspect something similar would apply to the sweet spot analysis, because Ord’s method still inflates costs at shorter time horizons.
METR's frontier time horizons are doubling every few months, providing substantial evidence that AI will soon be able to automate many tasks or even jobs. But per-task inference costs have also risen sharply, and automation requires AI labor to be affordable, not just possible.[1] Many people look at the rising compute bills behind frontier models and conclude that automation will soon become unaffordable.
I think this misreads the data. The rise in inference cost reflects models completing longer tasks, not models becoming more expensive relative to the human labor they replace. Current frontier models complete tasks at their 50% reliability horizon for roughly 3% of human cost, and this hasn’t increased as capabilities have improved. So, cost isn't an additional bottleneck beyond capability, and we should expect to see automation at roughly the time predicted by METR's capability trendlines.
I define cost ratio as the inference cost of the average AI trajectory that solves a task divided by human cost to complete the same task. Using METR’s data, I examine the trend in cost ratios over time. I show three things:
Raising the cost ratio by providing additional compute might translate into longer time horizons, but that would be in addition to, rather than causing, current trends. So, compared to extrapolating METR’s data, increased inference-time spending will strictly shorten AI timelines.
In his article on this subject, Toby Ord came to the opposite conclusion; he argued that there is moderate evidence that “the hourly costs are rising exponentially” at the frontier of AI capabilities. In appendix A, I argue that the methodology Ord used to produce this conclusion is unreliable and leads to significant overestimates of hourly model cost.
Thanks to Toby Ord, Buck Shlegeris, Agustín Covarrubias, Alex Mallen, Alexa Pan, Tim Hua, Aniket Chakravorty, Arun Jose, and Francis Rhys Ward for helpful comments on earlier drafts. Thanks especially to Ryan Greenblatt for the ideas that inspired this article.
All figures in this article were generated with code that can be found here.
Evidence from METR
In this post, I analyze publicly available time horizon data from METR. I calculate cost ratio by estimating AI inference cost from token count (using OpenRouter pricing) and dividing by the human cost for each task provided by in the dataset. All calculations are performed using METR’s task weighting. I made some modeling assumptions to fill gaps in the METR dataset (see footnote).[2] These assumptions shouldn’t affect my overall conclusions, because they shouldn’t significantly affect the trend across models.
My analysis has important limitations, which I discuss after presenting evidence.
Cost ratio at models’ 50% time horizon isn’t increasing
To investigate whether cost ratios rise as models improve, I examine the cost ratio at each model's 50% reliability time horizon. For each model, I select tasks that the model successfully completes with lengths between 0.79 times and 1.29 times the calculated time horizon (which is plus or minus 0.1 orders of magnitude).[3] I graph the median cost ratio in the selected set of tasks.
Figure 1: Median cost ratio of successful completions near each model's 50% time horizon (between 0.79x and 1.29x the time horizon), with interquartile range. As time horizons increase from minutes to hours across successive models, cost ratios remain well below 1 (dashed line), with no upward trend.
Cost ratios show no upward trend across models, and remain well below 1. As models' 50% time horizon increases exponentially, companies can cheaply use them to complete longer tasks.
I filter for passing tasks to show that cost isn't an additional limitation on top of capabilities. If successful tasks are cost-effective, then companies can set a low cost cap and still capture most successes. One might be concerned that this underestimates costs, because models tend to succeed at cheap tasks. But even at the 80% time horizon, where models complete the majority of tasks, cost ratios remain well below 1 (figure B3). Including failures doesn't change this (figure B4).[4] This means companies can profitably automate a broad share of tasks, not just a narrow cheap subset.
Time horizons improvements aren’t driven by expensive long tasks
One might respond to figure 1 by arguing that even if cost ratios at each model’s 50% time horizon aren’t increasing, improvements to those 50% time horizons come from successes on long tasks with high cost ratios. If this were true, we'd expect cost ratios to rise with task length among successfully completed tasks. I show that’s false: among successful tasks, cost ratios don't rise with task length.[5]
Figure 2: Cost ratio by task duration for successful attempts only (excluding tasks under 1.5 minutes), across a selection of models. The shaded region shows the weighted interquartile range. Cost ratios do not increase with task length, they decline, though this likely reflects a selection effect where models only succeed on relatively cheap long tasks.
This doesn’t mean models are more efficient at longer tasks. There's a strong selection effect here: models pass far fewer long-horizon tasks, and likely only succeed on those that happen to be relatively cheap. But since only successes push the time horizon forward, what matters is whether longer successful tasks are expensive, and the data shows they aren’t.
Progress at a fixed cost is just as fast
Another way to examine the affordability of AI progress is to ask how much progress slows if companies cap AI spending at a fraction of human cost.
I model the affordable 50% time horizon, the maximum task length at which models have a greater than 50% chance of both completing a task and doing so under a given cost ratio. If progress is fast at a low cost ratio, then we can expect that AIs will soon be able to cheaply complete long tasks. I test a range of caps, from 1/4x to 1/32x human cost, to see whether tighter budgets substantially slow the trend.
I model affordability and passing separately. I fit a logistic curve to the pass rate as a function of task length, and a separate logistic[6] to the probability that any passing task is affordable. I do this because pass rate falls with task length, but affordability increases (see figure 2). Fitting a single logistic curve would fail to capture these opposing trends.
Multiplying the pass rate logistic and affordability logistic gives the probability that a task at some length will both pass and be affordable:
The affordable 50% time horizon is the longest task length at which p(pass & afford) is greater than or equal to 50%.
Figure 3: Factored model for Claude 4 Opus at four cost thresholds (1/32x, 1/16x, 1/8x, 1/4x human cost). Red dots show failed or over-cost tasks, while green dots show passed and affordable tasks. Each panel decomposes the affordable pass rate (red) into pass probability (blue) and affordability given passing (green dashed). The dashed vertical line marks the affordable 50% time horizon. Tighter cost constraints reduce the horizon only modestly.
At different maximum cost ratios, the doubling times are reliably similar.
Figure 4: Affordable 50% time horizon over time after 2024 at four cost thresholds: unlimited, 1/4x, 1/8x, and 1/32x human cost. Doubling times range from 3.0 to 3.3 months across thresholds, with high R² values throughout. Stricter cost constraints filter out some models but do not substantially slow the trend.
Stricter cost thresholds filter out some models, but enough recent models have cleared even the tightest thresholds to keep the doubling time around 3 months.
Limitations of my methodology
This data has some important weaknesses for the analysis I performed above:
I think that these may mildly influence my results, but are unlikely to change overall trends.
Inference scaling will just make progress faster
If companies are willing to increase cost ratios (to 1, or even above), they might unlock even better capabilities. AI might be faster or higher quality than human labor in some domains, so the additional cost may be worthwhile.
UK AISI showed that newer models have larger gains from additional inference compute. In some domains, like ARC-AGI, performance improves a lot from large amounts of test-time compute, as demonstrated by o3’s performance at different cost thresholds and Ryan’s results . METR shows that additional token budget can improve a model’s 50% time horizon (note that this is an upper limit on per-task tokens, so not analogous to cost ratio), although with the Claude Code scaffold this is fairly limited.
Over time, we should expect the gains from additional inference compute to grow as models improve at using test-time compute. Already, AI providers offer increased test-time compute at a cost premium, like GPT-5.4 pro and extended thinking with Claude. As companies automate more labor, they might increasingly opt for these premium tiers over base models.
It's not yet clear how much improvement is possible by increasing test-time spending, but given the relatively low cost necessary to achieve METR's time horizons, substantial increases may be economically viable. Any gains would come on top of the trends observed by METR.
Conclusion
The data shows that observed increases in AI agentic capabilities have not come from rising cost ratios. METR's trendlines predict the length of tasks AI can complete cheaply.
I think companies will be able to affordably deploy models to complete tasks at their 50% or 80% time horizons. At these horizons, many tasks will fail, but at ~3% cost ratio companies can cap spending per attempt well below human cost, retry identified failures, and still profitably automate. And, more speculatively, if companies can predict which tasks will succeed, they can selectively automate those.
In some domains, it might be possible to increase cost ratios to make AI agents much more capable, but this possibility strictly shortens timelines to any capability milestone, rather than lengthening them. The first AIs with human-level intelligence may cost more than human labor, but cheap human-level AIs should not be far behind.
Appendix A: Why I get different results from Ord
Toby Ord analyzed the same issue and found very different results. I think this is explained by how he calculated hourly cost, which leads to significant overestimates. I also explain more minor reasons to prefer my analysis to his at the end of this appendix.
Ord’s results come from his analysis of this graph, released by METR (with his annotations).
Figure A1: Comparison of time horizon achieved by token cost used for different models. Each model is run once with a constant token budget, and the performance at each cost threshold is calculated by counting all unfinished runs as failures. Costs for o3 and GPT-5 are approximated using cost data for OpenAI o1. Toby Ord has added saturation point annotations to the figure, with diagonal lines indicating a flat hourly cost.
Ord uses this graph to find the hourly cost of AIs at their “saturation point” (where increasing the total budget doesn’t significantly increase the model’s 50% time horizon).
I don’t think Ord’s methodology calculates the real hourly AI cost for any set of tasks. Ord calculates hourly cost using the following method for the saturation point:
Where the per-task budget is the maximum that can be spent on any individual task.
For example, my data shows GPT-5 succeeds on 1-2 hour tasks at a median cost of $1.29, but Ord’s method requires a $10+ budget to reach a time horizon in that range. The reason is that the 50% time horizon at the saturation point is influenced by passing tasks significantly longer than the horizon itself (see figure A2). So the per-task budget must be large enough to pay for the most expensive task necessary to reach the saturation point, which has a high absolute cost even if its hourly cost is low. Dividing this inflated budget by the shorter 50% time horizon produces an inflated hourly cost. In effect, hourly cost is set by the price of passing a 4-hour task, divided by a 1.6-hour horizon.
Figure A2: As the cost cap rises, GPT-5's horizon grows from ~30s to ~3.5h. The bottom panel decomposes this growth by counterfactual attribution: at each cost step, I remove each duration bucket's newly-passed runs and refit the logistic to measure its share of the p50 shift. Early growth (under $1) is driven by short tasks (<4m, 4-16m). Above $10, 4-16h tasks (purple) dominate; expensive successes on very long tasks are what push the horizon from ~2h toward its ~3.5h plateau. (My updated data has a different time horizon than in Figure A1, but it matches METR’s updated time horizon.)
Figure A3: Dividing each model's required budget by its time horizon gives an implied hourly rate. Ord's method implies $5–59/hr across models (red) at their overall 50% time horizon, but the actual weighted mean and median per-task hourly costs of successful tasks near each model's horizon duration are much lower (orange, green). The overstatement ranges from 9–64x against the median. Error bars show 90% bootstrap confidence intervals. I have ignored saturation for simplicity, and simply find the cost necessary to reach 95% of a model's unrestricted 50% time horizon.
This leads to significant overestimation of absolute costs, possibly by an order of magnitude or more, and it is sufficiently noisy (overestimation from 9x to 64x) that it's also difficult to draw conclusions about trends.[9]
I think there are other, more minor reasons to prefer my analysis to Ord’s:
Appendix B: Additional cost at time horizon graphs
50% time horizon
I show that the overall trend of cost ratio versus time horizon doesn’t depend much on whether I include tasks with time horizons further than the model’s 50% time horizon.
Figure B1: Same as Figure 1 but with a narrower task selection band (±0.05 OOM). The trend is unchanged.
Figure B2: Same as Figure 1 but with a wider task selection band (±0.2 OOM). The trend is unchanged.
80% time horizon
Figure B3: Same as Figure 1 but measured at each model's 80% time horizon. Cost ratios remain below human cost with no upward trend, though o1-preview and o1 are notably more expensive at this threshold.
Including failures
Figure B4: Same as Figure 1 but including failed attempts in the cost calculation. Median cost ratios shift upward for reasoning models (o1-preview, o1, o3) that spend heavily even on failures, while most other models remain largely unchanged.
Appendix C: 80% affordable time horizon
For some applications (e.g., physically dangerous or secretive ones), AI might be used to replace humans even if it is more expensive than them. I still think AI labor being expensive would substantially slow adoption in many areas.
For METR's data, I don't have exact inference costs for most models, so I estimate cost from each run's total token count multiplied by the simple average of that model's per-token input and output prices on OpenRouter. Note that OpenRouter prices can differ from direct API prices (e.g. Claude 3.5 Sonnet is $6/$30 per million tokens on OpenRouter vs $3/$15 on Anthropic's API). Because the data only records an aggregate token count (no input/output breakdown), this assumes a roughly equal split between input and output tokens. I tested this assumption against a few tasks for which METR recorded the exact cost and found it to be a reasonable approximation. Pricing was not available on OpenRouter for a few models: for Claude 3 Opus I used Anthropic's historical API pricing, for o1-preview and Claude 3.5 Sonnet (Old) I substituted pricing from their current-generation equivalents (o1 and Claude 3.5 Sonnet (New), respectively). Some runs for GPT-4 0314 and o1-preview lack token counts and are excluded from the cost analysis.
The overall shape of the graph isn’t very sensitive to this parameter. I show this in Appendix B.
Failed attempts are harder to interpret, since a failure doesn't tell us how many tokens would have been needed to succeed, but the absence of any upward trend even with failures included suggests selection isn't masking rising costs.
I exclude tasks under 1 minute 30 seconds because they are predominantly run on a different scaffold (swaa/generate) that is less comparable to the agentic scaffolds used for longer tasks. I also do this for figures 3 and 4.
I tried many fits, and the logistic had the best accuracy on withheld datapoints.
To avoid selection effects, one might compare models only on tasks they all complete successfully. But weak models only complete very short tasks, so this would only tell us about costs on 5-minute tasks. I'm instead interested in costs at the frontier of each model's capabilities.
Thanks to Gordon Seidoh Worley for this observation. From their comment.
Ord also calculates “sweet spots”, where models are maximally cost-effective, using a similar methodology. Although the effect might be smaller, I suspect something similar would apply to the sweet spot analysis, because Ord’s method still inflates costs at shorter time horizons.