GPT-6.1 Terra costs half as much. I checked which halves OpenAI meant.

OpenAI launched GPT-6.1 Terra on Monday (October 26) with a pricing headline: standard API prices 50% lower than GPT-5.6 Terra and GPT-6.1 Sol — $1 per million input tokens, $0.05 cached, $5 output. Pricing headlines are arithmetic, and arithmetic can be checked. I ran every number in the announcement through a calculator and two workload models. Most of it holds up. One claim doesn't multiply.

The workload math

First, what the family actually costs per task. Prices per million tokens (input / cached / output), from OpenAI's Terra post and its model pages as reported by ComputerWorld and The Register:

GPT-6.1 Terra      $1.00 / $0.05 / $5.00
GPT-6.1 Sol        $2.00 / $0.10 / $10.00
GPT-6 Astra        $10.00 / $1.00 / $50.00
GPT-6 Luna         $0.10 / $0.01 / $0.50
Claude Opus 5.5    $4.00 / $0.10 / $20.00

Workload A is a computer-use agent task — the thing Terra is sold for: 180k fresh input tokens of screen and context, 60k cached, 12k output. Workload B is OpenAI's own example, invoice processing: 20k in, 2k out, scaled to 10,000 documents.

Workload A (one computer-use task):
  Terra $0.243   Sol $0.486   Astra $2.460   Luna $0.025   Opus 5.5 $0.966

Workload B (invoices, per 10,000 docs):
  Terra $300     Sol $600     Astra $3,000   Luna $30      Opus 5.5 $1,200

The ratios behave exactly as advertised at the list-price level: Terra is precisely half of Sol on both input and output, precisely one-tenth of Astra per token (matching the "on par with Astra at one-tenth its price" line), and precisely one-quarter of Opus 5.5 per token. No rounding games in the rate card itself.

Checking the compound claims

The interesting claims are the compound ones — price × token efficiency = cost per task — because that's where launch posts usually smuggle a factor of two.

Terminal-Bench 4.0: checks out. Terra solves 50.3% of tasks at max effort versus GPT-5.6 Terra's 21.5%. That's 50.3 / 21.5 = 2.34× — "more than double" is accurate, even slightly understated.

Factuality: checks out. The internal error rate at low effort fell from 53.6% to 16.7%, a drop of 36.9 points — the post says "nearly 37." At max effort it "fell by 18 points to 4.3%," which quietly tells you the predecessor's max-effort rate was 22.3%. Consistent.

Agents' Last Exam: implies a hidden 1.5×. Terra scores 15 points higher than its predecessor "at about one-third the cost," while its list price only halved. For both to be true, Terra must also be using about 1.5× fewer tokens per task (0.5 ÷ ⅓ = 1.5). Plausible for a model tuned for computer use, but the token saving is doing a third of the work in that sentence and never gets stated.

The Opus comparison: doesn't multiply. At max effort, the post says Terra is "over three times as token efficient as Claude Opus 5.5, and one-fourth its price per token, resulting in a ~9x lower cost per task." But 3 × 4 = 12, not 9. If the cost ratio really is ~9× and the price ratio is exactly 4×, the implied token efficiency is 9 / 4 = 2.25× — not "over three times." The two efficiency figures presumably come from different evaluation slices or effort settings, but as printed, a reader who multiplies the two headline numbers gets 12× and the post claims 9×. To OpenAI's credit, the claimed total is lower than the naive product — this is sloppy composition, not inflation. It's still the one number in the post you cannot reproduce from the other numbers in the post.

What this means if you're buying

Terra's real pitch isn't the benchmarks; it's the shape of the price curve. The family now spans two orders of magnitude — Luna at $0.10/$0.50 to Astra at $10/$50 — with Terra sitting at the point where a computer-use task costs about 24 cents instead of Sol's 49 or Astra's $2.46. At invoice scale, the Terra-vs-Sol decision is $300 versus $600 per 10,000 documents: real money, but small next to the cost of reviewing the output. That cuts both ways. As one analyst quoted by ComputerWorld put it, in an illustrative coding workload about 5.4 seconds of extra human review erases the inference saving entirely. A model that is half the price and needs one more minute of checking per task is not half the cost.

So my reading: the rate card is honest, the benchmark deltas I could verify are honest, and the one compound claim that fails arithmetic fails in the conservative direction. Buy Terra for high-volume, low-review-cost work — forms, invoices, screen scraping with deterministic checks. Keep Sol or Astra for the tasks where a wrong answer is expensive to catch. And whenever a launch post gives you an efficiency factor and a price factor, multiply them before you repeat the total. It takes ten seconds, and about one time in five, the total is a different number.

Sources: OpenAI, "Introducing GPT-6.1 Terra" (Oct 26, 2026) — all benchmark figures and Terra pricing are OpenAI's own; family pricing corroborated by ComputerWorld and The Register (Oct 26–27, 2026). Workload costs are my calculations from those list prices.