When AI Gets Cheap Enough, Volume Work Stops Being a Sample

A research team can now run a frontier-capability language model over ten thousand candidate neuroscience papers for a little over a hundred dollars. At March 2026 prices, the same pass would have cost several thousand dollars. A year ago, the capability did not exist at any price.

The economics come from a blog post by Catalystneuro, a small team that builds neuroscience infrastructure. Their argument is not that AI is getting smarter; it is that AI is getting cheap enough to run at project scale. The dominant coverage of capability milestones, they write, misses the more economically consequential shift.

The ceiling is the newest capability a frontier model can reach. The floor is the cost of running a model that is good enough at a given task. Catalystneuro argues the floor has fallen roughly two orders of magnitude; the recent drop between March and August 2026 is closer to 20 to 50 times for the same benchmark. The post's plots are derived from public data on Artificial Analysis's intelligence-versus-price benchmarks, which track hundreds of models over time and remain the more durable anchor for the cost-versus-capability claim.

Per-token price falls for a few reasons. Distillation compresses a large model into a smaller one that costs less to run. Quantization reduces the precision of the weights, cutting memory and compute. Specialized inference chips, including the custom silicon that has filled out hyperscaler fleets through 2025 and 2026, push the unit economics down further. Catalystneuro does not separate these effects; the floor is the net result.

Catalystneuro is using the cost drop to measure dataset reuse across the entire DANDI Archive, a public neuroscience repository. Corpus-scale analysis of that archive was a sample-only exercise before. The worked example is one team's project, not a general benchmark, so the cost numbers are the authors' own; a robust version of the story needs the same shape outside neuroscience.

Digital Applied's Q2 2026 Efficient Frontier analysis shows the same performance-versus-price curve from a third-party view, useful as corroboration that the floor is moving faster than the ceiling. Independent benchmark data anchors the cost-versus-capability claim.

What changes when the floor drops? Categories of work that used to be sample-sized become project-sized: literature review at corpus scale, contract review at archive scale, summarization across large forums, high-frequency monitoring. A team that was never going to train a frontier model can still run a good-enough one ten thousand times on a real budget.

The HN discussion of the Catalystneuro post takes the same lens to other domains. One commenter cites a 50 to 500 times cost reduction in robotics from specialized chips and improved distillation, and projects that by 2031, Fable-5-or-greater intelligence could run on smartphones. Both figures are community speculation, not forecasts; the robotics number and the on-device projection are not anchored to a published benchmark.

The thesis has a falsifier. The floor-drop story only unlocks volume work if per-token price is the actual bottleneck. Latency, error rates, integration cost, and review overhead can still keep a corpus-scale project out of reach even when the model call is cheap. A defensible test is whether the cost of running the model drops faster than the cost of validating its output.

For now, the ten-thousand-papers pass is a real, dated example: a hundred dollars, recent frontier models, corpus-scale coverage. The next test is whether the same shape holds outside neuroscience, on a contract archive or a large public forum. If it does, the cheap-models story stops being a take and becomes a budget line.