When inference drops below $1, databases inherit the agent problem

BAIR argues that GPT-4-class inference has fallen from roughly $30 per million tokens to under $1. For data teams, the bottleneck moves from the model itself to memory, coordination and control around agents.

BAIR moves the argument from models to the systems around them

Berkeley AI Research published a perspective essay, Intelligence is Free, Now What? Data Systems for, of, and by Agents. The authors argue that the cost of useful AI is falling fast enough that, for much knowledge work, intelligence starts to look like a cheap commodity.

They ground that claim in numbers: GPT-4-class capabilities cost about $30 per million tokens in early 2023, while comparable capabilities now run below $1 and some providers are pushing below $0.10. They also cite inference price declines across benchmarks of roughly 9x to 900x per year, with a median near 50x.

From there, they frame three research directions: data systems for agents, data systems of agents and data systems by agents. In plain terms, systems that serve agents, systems made of agents and systems that agents help design or operate.

Engineering teams will optimize coordination, not just token spend

When models are expensive, every call is rationed. When they get cheap, it becomes rational to run more agents, try more variants and treat inference as part of the runtime. That changes the job for databases, orchestration and observability.

The practical implication is not that every workflow should become agentic overnight. It is that data systems need to handle messier workloads: many concurrent requests, long contexts, agent memory, speculative runs and auditable decisions.

For product teams, this changes the cost model. Token price falls out of the headline budget, but the cost of a wrong action, bad retrieval or agent memory that should have expired moves into view.

Cheap intelligence still does not buy truth, permissions or accountability

The BAIR post is a perspective, not a product launch or production benchmark. Its pricing numbers fit the broader API market, but the claim of near-free intelligence is a strategic interpretation, not a law of nature.

Cheap inference also does not make everything cheap. Context windows, latency, storage, privacy, governance and human sign-off remain hard limits. With agents, the expensive part is often the moment they do something convincingly wrong and someone has to reconstruct the trail.

The real test is whether databases can treat agents as normal traffic

The next signal to watch is whether these ideas become concrete architectures: transactions for agent workflows, structured memory, stronger provenance, evals for data operations and safe speculative execution.

The proof will not be another chart of token prices. It will be a data team letting a swarm of agents touch company data without forcing security to stand beside the monitor with a fire extinguisher.

Lilith's verdict

Cheap tokens are the receipt, not the victory. Once intelligence costs pocket change, the real bill arrives in memory, permissions and the person sweeping up after a bad agent move.