Two price-performance moves in one week: Qwen's 3.8-Max positioned as better and cheaper, DeepSeek-V4-Flash in public beta with upgrades.

The practical consequence

Jobs you priced out six months ago may now be affordable. Embedding a full corpus, summarising every abstract in a subfield, running an LLM judge over an evaluation set — these were budget decisions at last year's rates.

Costs falling repeatedly means the assumption underneath a "we can't afford that" decision expires quickly. Re-check the ones that shaped your architecture.

What does not fall

Storage, and human review. A cheaper model that lets you process ten times more text produces ten times more output somebody has to read — and ten times more rows somebody has to store.

We hit exactly that wall recently: a vector index that was cheap to build sat at 99% of a database quota, serving a feature that had been removed from the UI.

The rule of thumb

Falling inference costs shift the bottleneck downstream. Budget for the review and the storage, not the tokens.

Re-price the decisions, not just the jobs

The important consequence of falling costs is not that existing jobs get cheaper. It is that decisions you made because of cost are now wrong.

"We sampled 500 papers because embedding the full corpus was too expensive" is a methodological compromise driven by a price that has since fallen by an order of magnitude. That compromise is still sitting in your pipeline, and nobody revisits it, because it was never written down as a cost decision.

Go find those. They are usually in the methods section, phrased as if they were principled.

What does not get cheaper

Storage. Vectors and extracted text accumulate and do not compress well. On a fixed-quota database this becomes the binding constraint long before compute does.

Human review. Ten times the output requires ten times the checking, and that cost is flat. A pipeline that produces more than you can review produces unreviewed output, which is worse than producing less.

Maintenance. Every job you add is a job that breaks when an API changes.

The concrete example from this codebase

An embedding index built over 25 days of local GPU time occupied 216 MB and drove a 500 MB database to 99% capacity. It served a similarity feature that had been removed from the UI months earlier.

Nothing about that was a compute problem. The compute was nearly free. The failure was not tracking whether the artefact was still used.

A budgeting heuristic

Price the whole lifecycle: generate, store, review, maintain. Token cost is usually the smallest of the four, and it is the only one anyone estimates.

What to actually do this quarter

Pick the three largest cost-driven compromises in your current pipeline and re-cost them at today's prices. At least one will have flipped.