What changed in Pi 1.0, and what does it cost?
Pi 1.0, released October 1, 2026, is still free: you pay only for the model provider behind it. The change that moves your bill is a leaner codemode prompt, about 40% smaller, so a GPT-5.6 request drops from about 5,300 to 3,300 tokens. Pi 1.1 added Claude Haiku 5.5 and fixed session costs that undercounted long prompts.
The changes that affect your bill
| Release | Change | Cost effect |
|---|---|---|
| 1.0 (Oct 1) | Leaner codemode | About 2,000 fewer prompt tokens on every request with the default tools and codemode on. |
| 1.0 (Oct 1) | Image generation in codemode | Scripts can call image models with your session's credentials, and that usage counts toward the session cost. |
| 1.0 (Oct 1) | Anthropic copy-code login | Sign in when the browser runs on another machine, so a headless box can use your Anthropic account. |
| 1.1 (Oct 7) | Claude Haiku 5.5 and a cost fix | Haiku 5.5 is $0.10 / $0.50 per million tokens, five times that on prompts over 100K. Session costs no longer undercount long prompts on models priced by prompt length. |
| 1.2 (Oct 12) | Per-key OpenRouter model lists | With an OpenRouter key, Pi lists only the models that key may use, with their prices. |
Source: Pi releases 1.0.0 to 1.2.0, checked October 14, 2026.
What the smaller prompt saves
Pi measured the cut on GPT-5.6; the same 2,000 tokens come off whichever model you run. They matter most on cache misses, at the start of a session or after a pause.
| 2,000 tokens on Claude Sonnet 5.5 | Saved |
|---|---|
| Per request, cache miss ($2 per million) | $0.004 |
| Per request, cached ($0.10 per million) | $0.0002 |
| Per 1,000 requests, all cache misses | $4.00 |
Our estimate at Anthropic's list prices, from Pi's own token counts.
Upgrade notes that can bite
- The terminal now runs fullscreen. Set
tuiModeto"regular"to keep normal scrollback. --providerwithout--modelnow fails with an error instead of quietly running another provider's default model.- Codemode scripts that checked for a tool with
typeof tools.namemust use"name" in tools. - Pi 1.0.3 renamed the
azure-openai-responsesprovider toazure. Rename it inauth.json,models.jsonandsettings.json; old sessions on it resume on another model and lose their prompt cache.
The catch
A smaller prompt trims every request, but it doesn't cap anything. A long codemode session on a frontier model still bills per token, and Haiku 5.5 is cheap only while prompts stay under 100,000 tokens.
Where Standard Compute fits
Pi takes an OpenAI-format provider in ~/.pi/agent/models.json, so it runs on Standard Compute with base URL https://api.stdcmpt.com/v1 and model standardcompute. Plans start at $19 a month with a $20 compute budget, and requests stop at the budget instead of billing overage.
FAQ
Is Pi 1.0 free?
Yes. The Pi coding agent is free; you pay the model provider you connect, by API key or by signing in to an account such as Anthropic.
How much does Pi 1.0's leaner codemode save?
About 2,000 prompt tokens per request: Pi's own example drops from about 5,300 to 3,300 tokens. On Claude Sonnet 5.5 that is $0.004 per uncached request, or $4 per 1,000 requests.
Does Pi support Claude Haiku 5.5?
Yes, since Pi 1.1.0 on October 7, 2026, as anthropic/claude-haiku-5-5 with adaptive thinking. It is $0.10 / $0.50 per million tokens up to 100,000-token prompts and $0.50 / $2.50 above that.