Frontier AI labs must recover large development costs from the money left after serving customers. They serve two markets with different sources of margin growth. In the cost-conscious market, inference innovations can lower the cost of delivering adequate results. In the quality-conscious market, margins expand principally through capital investment in better models that command a premium. Competition can erode either advantage, and a frontier lab may have to finance another release before recovering the cost of earlier models and failed experiments.
Coordinated pacing could let leading models earn premiums for longer while reducing the capital needed to keep up with rival development. Inference improvements could continue lowering serving costs during that longer earning period. A lab cannot obtain the same protection by slowing down alone. Restrictions on imitation and competitors’ access to compute could strengthen that protection, though cheaper adequate models would remain a threat. The gains must outweigh added safety costs and earnings forgone by delaying better models. That potential financial benefit helps explain why frontier labs would support government-coordinated pacing alongside its stated safety purpose.
A lab that slows development alone risks losing customers.
Dario Amodei wants frontier labs to have more time for safety work “without sacrificing commercial advantage.” He proposes government coordination and capability checkpoints, and commits to giving independent reviewers access and publication rights. Shared, verifiable limits would make it easier for a lab to slow development without losing customers to a competitor that keeps improving its models. [1]
Cheaper competitors can reduce revenue before development costs are recovered.
On September 13, GPT-6 Astra at high effort scores 51.05 at $1.72 per benchmark task; Fable 5 with fallback scores 49.70 at $8.75. Astra scores higher at roughly one-fifth the cost. Customers who can replace Fable with Astra have a reason to switch or demand a lower price. Either response can reduce the revenue available to recover Fable’s development costs. [2]
These figures use Intelligence Index v4.3. They show a cheaper substitute on the benchmark; they do not tell us how many customers switch or whether either model is profitable.
Historical price decline · Through 3 September 2026 · Index v4.1.1
Estimated half-life of the minimum price for comparable capability: 46 days.
Each line follows the cheapest model in the same five-point score band after a flagship launches. A starting price of 1 falls to 0.5 when it halves.
Method, model histories & data
The estimate uses five releases with at least 90 days of history. Each release contributes its full history, from 98 to 294 days; the cutoff does not shorten any line. Newer releases are shown with all available points, including the September 1–3 launches. GPT-5 is excluded from the estimate.
We fit an exponential decline to each older release’s daily price history, using log prices and a starting value of 1. We average the five decline rates and convert that rate to a half-life: 45.82 days. The displayed straight segments interpolate between price events.
Three releases in the estimate share the 55–60 score band, so their histories overlap. A band can include models scoring below the flagship. Requiring a model to match or beat the flagship’s exact score gives a half-life of 50 days; this is a sensitivity check, not an uncertainty interval.
The starting price can already be below the flagship’s own price. Before August 19, most prices are later observations dated back to launch, with some known price cuts added. Earlier prices and availability were not fully recorded. Scores use Artificial Analysis v4.1.1: the September 3 snapshot, plus Astra’s launch-day measurements from the September 4 morning archive. GPT-5’s price remains in the market comparison data.
| Release | Score band | Days | In fit |
|---|---|---|---|
| GPT-5.1 | 35–40 | 294 | Yes |
| GPT-5.4 | 50–55 | 182 | Yes |
| GPT-5.5 | 55–60 | 133 | Yes |
| Opus 4.7 | 55–60 | 140 | Yes |
| Opus 4.8 | 55–60 | 98 | Yes |
| GPT-5.6 Sol | 60–65 | 56 | No |
| Fable 5 | 60–65 | 86 | No |
| Opus 5 | 60–65 | 41 | No |
| Fable 5.1 | 65–70 | 2 | No |
| Gemini 3.8 Flash | 55–60 | 1 | No |
| Muse Spark 1.3 | 60–65 | 1 | No |
| GPT-6 Astra | 60–65 | 0 | No |
Download the plotted data (JSON) · Benchmark methodology ↗
The current comparisons come from the September 13 Pareto-frontier snapshot, covering 41 configurations on Index v4.3. We checked the four configurations cited here against Artificial Analysis’s embedded data. The cost frontier contains 13 configurations from 4 model families. These positions identify benchmark tradeoffs, not profitable businesses. Download that snapshot (JSON).
Fable 5.1 at max effort with fallback has the highest score in this snapshot: 53.37 at $7.63 per task. Astra max scores 52.81 at $3.26. Fable costs 2.3 times as much for 0.56 more index points. The benchmark does not establish how much customers would pay for that score difference. Fable 5 uses its reported fallback configuration.
Cost-conscious and quality-conscious markets expand margins in different ways.
The distinction is how providers increase the margin between what customers pay and what it costs to serve them.
Quality-conscious
Customers pay a premium for better results. The principal path to higher margins is capital investment in greater capability: funding research, compute, and training to build models whose results justify higher prices.
The premium must exceed any additional serving costs, and the resulting earnings must repay the development investment. As rivals improve, maintaining the premium requires further capital.
Cost-conscious
Once several models meet a task’s quality requirement, buyers compare total cost per successful task. Providers can expand margins through inference innovations that make adequate results cheaper to deliver. Better memory management, for example, lets the same hardware serve more requests. [5]
Margins widen when serving costs fall faster than prices. Rivals can adopt similar efficiencies and pass savings to customers, so the advantage may be temporary.
A single customer can buy in both markets. A use case can move into the cost-conscious market once cheaper models become adequate, while new capabilities can create new quality-conscious uses. [4]
Profitable releases must also cover losses on unsuccessful development.
Both markets can generate earnings after serving costs. For a frontier lab, those earnings must support repeated capital investment in development; high token volume alone does not establish that they are sufficient.
A lab pays for successful models, unsuccessful models, and experiments that never produce a release. If only leading models recover their development costs, their profits must also cover the unsuccessful attempts. A lab can therefore release several profitable models and still lose money overall. To remain competitive, it may have to fund another generation before recovering its earlier spending.
Slower competition can increase earnings and reduce development spending.
Coordinated pacing could reduce the rate of capital investment required to compete in the quality-conscious market and extend the period of premium pricing. Inference innovations could keep lowering serving costs for existing models, expanding margins without another frontier training run. Competition in the cost-conscious market would still pressure providers to pass those savings on through lower prices.
At unchanged monthly earnings after serving costs, selling a model for twelve months instead of six doubles its contribution toward development costs. Fewer development cycles could also reduce annual spending. These gains must exceed the earnings the lab gives up by delaying its own better models and any additional safety costs.
One option in a separate AI Futures proposal would allocate 70% of compute to serving customers, 25% to safety, and 5% to capability R&D. More capacity for serving customers produces more revenue only if there is demand for it. [3]
Restricting cheaper competitors increases the financial benefit of pacing.
Limits on new capabilities do less to protect revenue if competitors can offer existing capabilities at lower prices. Amodei also calls for restrictions on Chinese access to compute, unauthorized distillation, and model-weight theft. These measures could make some forms of competition slower or more expensive, helping existing labs maintain prices. They would not prevent competitors from independently developing cheaper models. [1]
Protecting current revenue helps explain why labs support pacing.
The quality-conscious market rewards capital investment in better models, but competition forces labs to keep reinvesting to sustain their premiums. Coordinated pacing could protect those premiums for longer, reduce the pace of new investment, and leave more time for inference innovations to improve margins on existing models. In the cost-conscious market, those innovations would still face pressure from cheaper competitors. Customers would receive some improvements and price reductions later. These financial incentives help explain the appeal of coordinating limits on capability and restricting cheaper competition.