GPT-6.1 Terra pricing,
and the new 12-hour cache write.

What does GPT-6.1 Terra cost?

At Standard short-context rates, GPT-6.1 Terra costs $1 for uncached input, $0.05 for cached input, $1.25 for a 30-minute cache write, $2.50 for a 12-hour cache write, and $5 for output per million tokens — half of GPT-6.1 Sol in every bucket. OpenAI model reference.

A single request above 272,000 input tokens — counting every input bucket — is priced at 2× the input and cache rates and 1.5× the output rate for the whole request. Terra also publishes a 922,000-token maximum input within its 1,050,000-token context window, and a 128,000-token output cap.

Try a one-request or authored-workload comparison with the browser-based API cost calculator.

30-minute or 12-hour cache writes

Since October 20, OpenAI models from GPT-5.6 onward accept an optional 12-hour minimum cache lifetime in the Responses API. A cache write replaces the ordinary input price for the written tokens: 1.25× input for the default 30-minute lifetime, 2.5× for 12 hours. The cache-read rate is the same at either lifetime, and reads keep refreshing the entry’s eligibility.

Writing 100,000 tokens on Terra costs $0.125 at 30 minutes or $0.25 at 12 hours. Cache reads cost the same at either lifetime, so the write alone decides: the 12-hour write pays off only when the same stable prefix — instructions, reference material — is re-written at least twice. A prefix hit at 9 am and again at noon, 3 pm and 5 pm needs four 30-minute writes ($0.50) versus one 12-hour write ($0.25). A single expiry is already break-even on write cost, so occasional reuse stays cheapest on the default lifetime. Usage reports split written tokens by lifetime, so each written token is billed once at the rate of the lifetime it was written under.

Claude’s explicit writes use different durations instead: 5 minutes by default or 1 hour. Gemini 4 Argon has no explicit cache-write price at all. In the calculator’s Tokens mode, choosing a duration a provider does not offer marks that model not supported with the reason — it is never silently priced at another provider’s rate.

A worked cost estimate

Consider one Standard short-context request: 100,000 input tokens of which 80,000 are cached reads, plus 10,000 output tokens, with no cache writes. These are example counts, not a benchmark.

Uncached input: 0.02M × $1

$0.02

Cached input: 0.08M × $0.05

$0.004

Output: 0.01M × $5

$0.05

GPT-6.1 Terra estimate

$0.074

The same token counts on GPT-6.1 Sol produce $0.148 — exactly double, because every bucket halves. Had this request exceeded 272,000 input tokens, every rate in the Terra column would move to the long-context card instead.

Thinking tokens count toward billed output. RepoSpend separates visible output and reasoning when the local source exposes that split, then prices each once. Read the token accounting methodology.

Track GPT-6.1 Terra usage by repository

Added in RepoSpend 0.1.8, the explicit GPT-6.1 Terra card means supported local usage is priced at Terra’s own rates instead of an inherited card. RepoSpend reads existing local Codex files and groups usage by Git repository.

1. Run npx repospend with Node.js 22.12 or later. The dashboard usually opens at localhost:2005.
2. Choose your date range and Codex source, then inspect the Models view for gpt-6.1-terra.
3. Open Repos or Sessions to find where the usage came from. Compare token mix and cache reuse alongside the cost estimate.
4. Use Data Doctor or repospend doctor to check missing token detail, unpriced models, and source warnings. Review local pricing overrides in Settings if your rates differ.

Your local prompts, transcripts, and repository usage stay on your machine. This public website receives none of that data. Cloudflare Web Analytics measures pageviews across the production website, including the calculator, without receiving calculator inputs or results.

API-equivalent cost is an estimate

A ChatGPT or Codex subscription follows its own limits and credit rules. Multiplying local tokens by API rates does not reconstruct your subscription bill, remaining allowance, or invoice.

RepoSpend uses Standard short-context rates by default. OpenAI charges higher rates above 272K input tokens per request, twice Standard for Fast, and half Standard for Batch and Flex; regional processing can add 10%. Tool fees and account-specific terms can also change the bill. Check the current pricing reference.

Local session totals cannot reliably reveal every request’s pricing tier. Compare models on the same task, count retries, and verify success with your own tests and review.