On October 1, Earendil tagged Pi 1.0, the first stable release of its MIT-licensed terminal coding agent, and the Hacker News thread that followed ran to 1,367 points. The release adds deferred tool loading, virtual model extensions, a full-screen TUI now switched on by default, an experimental “Pi Durable” for agent sessions meant to outlive the tab they were born in, and — as the release coverage summarized it — roughly 40 percent fewer codemode prompt tokens. Earendil says hundreds of thousands of weekly users weighed in on the road to 1.0.

Most of the thread argued about taste. Fullscreen against native scrollback. Vanilla against the oh-my-pi battery pack. Local Qwen as the principled alternative to renting inference. Almost nobody argued about the number that reaches an actual budget owner, and the release notes quietly buried the tell: cache warming for Anthropic models. You do not build cache warming unless you have already learned that prompt length is a second-order term and cache hit rate is the first-order one.

Brevity Is Not a Line Item

Under Anthropic’s pricing — the write-premium structure the company invented in August 2024 — a cache write costs 1.25x the base input rate at the default five-minute TTL and 2x at the one-hour tier, while a cache read costs 0.1x. Flexera’s walkthrough of the arithmetic is worth reading twice. At the standard $3 per million input tokens (the introductory $2 ran through August 31), that means $3.75 per million to write a five-minute cache, $6 per million for the hour-long tier, and $0.30 per million to read. In their worked example — 250 writes and 4,750 reads of a 10,000-token prefix — writes come to about $9.38 a day and reads to about $14.25.

Notice the shape of that. Reads outnumber writes nineteen to one and still cost more in total. The bill is not a function of how much text you send. It is a function of how often you send the same text.

Which means a harness that rewrites its system prompt, reorders its tool definitions, or splices a session-specific block onto the front of the context every turn is paying the 1.25x or the 2x and never collecting the 0.1x. It is possible to trim 40 percent of a prompt and raise costs, and if you have ever watched a cache-creation line item appear in a monthly invoice after a “cleanup” sprint, you know it is more than possible. A shorter prompt that stays put is a saving. A longer prompt that stays put is usually also a saving. A shorter prompt that changes is a donation.

There is a floor under all of this, too. Per-model minimum cacheable token counts mean a sufficiently austere context simply cannot be cached at all — you pay the full input rate on every turn, forever. Minimal and expensive is the worst quadrant, and it is the one a certain kind of engineering culture walks straight into.

The Lock-In Was Never in the License

A staff engineer who runs the agent fleet at a mid-size logistics software company put it plainly in a Slack DM during a Thursday deploy freeze: “We cut our system prompt by a third in July and our Anthropic line went up eleven percent. The cache stopped hitting. Nobody on my team had ever seen a cache-write line item before.”

The MIT license frees the code. It does not free the invoice. This is the part the open-harness crowd skips. Anthropic built the write premium in 2024; AWS Bedrock adopted it; OpenAI held out, marketing free automatic cache writes as a differentiator, and then dropped that with GPT-5.6; Azure is closing the loop now, per Finout’s write-up of the shift. Four vendors, one multiplier. A forkable harness is not a forkable cost model, and the thing that actually locks a team in is not the API surface — it is the pricing structure every major lab has now converged on, voluntarily, without a regulator in the room.

So the reassuring story about small open-source tools as a check on platform power is half right. They check the interface. They do not check the meter.

What a Budget Owner Should Actually Instrument

The metric to watch is cache-read ratio per session, not tokens per request. Prefix stability across turns. How often a routing decision, a timestamp, or a freshly serialized tool schema busts the prefix a hundred times a day. That is a boring dashboard. It is also the only one that predicts the invoice.

And there is a perverse consequence worth naming: the incentives push harness authors toward visible frugality — smaller prompts, prettier TUIs, a leaner diff in the changelog — while the charges are governed by a multiplier that never appears in the release notes. Pi 1.0 may well be the best terminal agent shipping today. Its authors, to their credit, shipped cache warming alongside the token cut; they know where the money is. The rest of us are still applauding the wrong number.

The right’s instinct here — small tools, exit rights, no bloat — is correct about software and wrong about invoices. Bloat you can fork. A write premium you can only pay.

Sources