What shipped, and when. Built in the open, mostly by one person. When we get something wrong we fix it in public, so this list has the corrections too.
Nov 2, 2026
New Terms of Service on November 16
We have published new Terms of Service that take effect on November 16, 2026. Prices move out of the Terms and onto the pricing page; a backup host may serve a request at its own price unless you turn backup rollover off in Settings; purchases are final; credits expire after 12 months with no account use (we email you first); and the limits on our liability and your indemnity are updated. We will ask you to accept them the next time you use Tīrtha on or after November 16. Every account holder gets an email about this.
Oct 31, 2026
No more free monthly requests
The free allowance of 100 requests a month per key has ended. Every request now uses prepaid credits: buy a pack in your dashboard, or use credit from a coupon we give you. Credits you already hold are unchanged. The terms (sections 4.2 and 4.5) and /pricing now say so.
Oct 26, 2026
Pinned levels, and agent sessions that hold up new
Tīrtha is still in internal testing, so keys, limits and prices may change. You can now pin a key or a request to one level: value (Base), quality (Plus) or max (Max). A pinned request is billed at the price of the host that serves it, shown on your invoice. auto is unchanged: Base $0.56 / $2.22, Plus $1.10 / $4.40, Max $1.90 / $5.69. Cached input is now billed at a lower rate, never above the input price. Max is for single accurate calls: a long agent session (4 or more tool steps, or 24k or more input tokens) is served and billed as Plus, and the response says max_redirect.
For coding agents: tool replies are no longer cut at 1,024 tokens (the default is now 8,192, or 16,384 on Plus), slow streamed turns send a keepalive every 5 to 10 seconds, and the OpenCode tool-call bug from October 18 is fixed. An auto-mode session stops cleanly at 80 requests or $2 and asks the agent to report. If one host runs out of credit we move to the next; you only get a 503 if every host for that level is out. Reasoning, when the model returns it, comes back as reasoning_content. See /pricing, /api and /agents.
Oct 20, 2026
Max now costs less: $1.90 / $5.69 new
We are always testing and making Tīrtha better for our customers. With this update we are able to drop the price of Max to $1.90 / $5.69 per million tokens in and out, one price on every host that serves it. Max used to be billed at a frontier list price plus a fee. It now runs on a model we tested on Tīrtha Bench, and like Plus its tool calls are checked against your schema before they are served. Base ($0.56 / $2.22) and Plus ($1.10 / $4.40) are unchanged. See /pricing and Tīrtha Bench.
Oct 18, 2026
New pricing: pay per level new
Pay as you go moves from one flat rate ($2 / $8 per million tokens) to a price per level: Base $0.56 / $2.22, Plus $1.10 / $4.40, and Max at a frontier list price plus a small fee. Each request is billed at the level that served it, so most work is billed at Base. See /pricing and Tīrtha Bench.
Oct 2, 2026
The open beta: sign in and make your own key new
Anyone can now sign in with Google at tirtha.ai/dashboard and create an API key, with no invite, card or sales call. Free keys carry 100 requests a month; if you need more during the beta, email us and we will raise it.
Oct 2, 2026
A real pricing page new
The free beta and pay as you go at $2 / $8 per million input / output tokens, on /pricing. Prepaid credits are coming; the terms now describe them, and refund terms will be published before they launch.
Oct 2, 2026
A new dashboard
Your keys, usage and settings in one place, with a first-run flow that walks you to your first key and a layout that works on a phone.
Oct 2, 2026
One design across every page
Every public page now shares one theme, one navigation bar and the fire logo, so the docs, policies and tools read as one site.
Oct 2, 2026
Data handling wording corrected across the site correction
Several pages still said your prompts stay private or are never retained. That was wrong: when an answer passes verification we store it together with the prompt that produced it, in a cache that is shared by default, and tirtha.no_cache skips it for a request. The server info, terms, privacy policy, enterprise and privacy pages now all say this. The browser demos that showed the old wording are retired. We also narrowed our tool-call claim to what we measure: on the primary path we do not serve a tool call that fails our checks, and the failover result is published on /capabilities.
Aug 6, 2026
The shared cache stores your prompt, and we said otherwise correction
Our privacy, security and terms pages said prompts were never placed in the shared cache. We read the running system and every cached entry carries the prompt that produced it. We corrected those pages the same day and pointed to the per-request opt out, tirtha.no_cache.
Jul 7, 2026
A wall of hostile tests for the verification endpoint hardening
Fifteen adversarial fixtures now gate every deploy of /v1/verify: instructions hidden in a source document can't steer the verdict, forged or inconsistent evidence can never open a cache gate, and a downed detector fails closed instead of guessing.
Jul 6, 2026
Verify an output before you act on it new
POST /v1/verify checks whether a claim is grounded in its source, then answers whether it is safe to serve, cache, or publish. Composite claims are split into atoms and grounded one at a time, so a failing claim comes back with the exact sentence anchored. Every verdict carries its evidence and provenance, never a bare yes/no.
Jul 6, 2026
A live playground: watch it catch a lie new
A no-key playground ran the verification engine against the real API: paste a source and a claim and see the decision, the reasoning, and the flagged sentence. It was removed in November 2026.
Jul 5, 2026
Support and feedback, in the product
Ask a question or flag an issue from your account and it reaches a real person. A second lower-cost provider was added as a failover, so one provider having a bad day no longer means your traffic pays frontier prices.
Jul 3, 2026
A plain-language security page
What we hold, what we never do, and how to check us. Corrected August 6, 2026: this entry previously said we cache verified answers and not your prompts. That was wrong. We read the database: every cached entry carries the prompt that produced it, and the shared store is the default. Send tirtha.no_cache to skip the cache for a request.
Jun 2026
Spend caps and routing dials, on by default
Every key carries a spending cap so a leaked key can't quietly become a large bill, and a dial that lets you choose how much routing you want, from everyday calls to frontier-every-time.
This is a young product and the list is short on purpose. We would rather ship a little that works than a lot that doesn't. · Docs
Bench · Back to overview