Why AI Got Cheap and Fast in September
Jev costs $42 per billion tokens, HydraFusion claims lower bills, Unity shipped 31 agent skills. Why AI got cheap and fast, and how startups should budget.
On this page
Same smarts, smaller bill, shorter wait. September stacked three cost moves: Jev at $42 per billion input tokens, GitHub HydraFusion preview claiming 36 to 67 percent lower spend, and Unity shipping 31 maintained skills so agents stop burning tokens on outdated forum code. Here is what got cheap, what got fast, and a budget any startup can run.
Takeaways
Jev resets the floor for decisions: $42 per B input, output free, 70 to 500ms, typed answers with confidence.
HydraFusion routes models per task: comparable or better than Claude Opus 5 on TerminalBench 2.1, DeepSWE, and CheckpointBench in preview, via /experimental in Copilot CLI.
Unity cuts wasted tokens: 31 engine skills for Unity 6 plus, from URP migration to multiplayer, maintained by engine teams.
Startup budget: cap cost per completed run, cache repeats, charge for outcomes, not chat.
What actually got cheaper?
Input tokens for judgment calls. A support triage that cost $0.013880 with an LLM demoed at $0.000081 with Jev on the TypeSafe side by side, before prompt tuning. HydraFusion attacks the same bill from above: pick a cheap model for easy steps, a strong model for hard steps, across three execution patterns, billed at standard model rates with no preview surcharge. Unity attacks waste: one correct URP migration skill beats ten forum retries that compile but fail.
old bill: one flagship model x every step x retries x forum fixes
new bill: routed models x cached catalogs x maintained skills x one pass
guardrail: budget per run -> alert at 2x -> kill at 3x -> review prompt on killSimple startup math
Pick 100 runs as the unit. Multiply tokens per run by price, add tool fees, divide by completed outcomes. A cheap model needing three retries loses to a dear one passing first time: track cost per merged result, not per call.
How should a small team budget agents?
Set per run caps
Tokens, tool calls, and wall time per run with auto stop. Long loops need compaction or checkpoints, not larger limits.
Cache what repeats
Tool catalogs with TTL, prompt caches across reconnects, skill files for repo knowledge. HydraFusion style routing only pays when repeats stay cached.
Buy maintained knowledge once
Unity style skills, versioned docs, and eval pinned prompts beat web search per run. Pay an engineer once to encode the path, save tokens forever.
Price the outcome
Charge per resolved ticket, migrated project, or shipped workflow. Token bills swing monthly, outcome prices hold: margin lives in the gap.
What stays expensive?
Taste, liability, and novel reasoning. Frontier chat, proofs, and large refactors still need strong models with human review. Real time classification, routing, scoring, extraction, and guardrails move to fast cheap calls. Split the workload on purpose: Jev style decisions inside code, LLMs for words and hard leaps, maintained skills bridging both.
Should you chase every new discount?
Should I switch models every week?
No. Pin a stack that passes your evals, re benchmark monthly across cheap plus flagship plus local options. Switch only when cost per completed outcome drops for three runs in a row.
What is the one skill to add first?
A project setup skill: editor config, version control, packages, migration notes. Unity put setup first for the same reason: correct starts prevent the most expensive retries.
As of September 25, 2026: intelligence per dollar became the feature. Route models, cache repeats, install maintained skills, and cap every run. Close the series with typed decisions to wire the cheapest reliable call into your own code.