GPT-6.1 Terra API Task Costs in Our Codex Test With a ChatGPT Sign-in

The short version

• One three-task round cost about 8 cents for GPT-6.1 Terra, about 18 cents for GPT-6.1 Sol and about 1 cent for GPT-6 Luna, at API-equivalent rates.
• GPT-6.1 Terra and GPT-6.1 Sol each got all nine runs fully right.
• The runs used a ChatGPT plan’s Codex allowance. These dollar estimates were not per-token charges.

What did one round of GPT-6.1 Terra, GPT-6.1 Sol and GPT-6 Luna tasks cost?

We priced the tokens from our own Codex test at OpenAI’s API Standard rates: a round covering a code fix, a technical drawing and a supplier form came to about 8 cents for GPT-6.1 Terra, about 18 cents for GPT-6.1 Sol and about 1 cent for GPT-6 Luna. GPT-6.1 Terra answered correctly in all nine runs. Our ChatGPT sign-in used the plan’s Codex allowance, so we paid no per-token bill; the dollar amounts represent the API equivalent of those tokens.

How were GPT-6.1 Terra, GPT-6.1 Sol and GPT-6 Luna tested in Codex?

We ran our earlier GPT-6.1 Terra test on a Linux server with Codex CLI 0.159.0 and a ChatGPT sign-in, from October 26, 2026, to October 27, 2026. All three models used medium reasoning effort, completing three runs per task with a fresh folder for each run.

We created the tasks so each had one exact right answer:

• A Python code repair, checked against eight tests.
• Ten questions about a mounting-plate technical drawing we made.
• A supplier form with 15 fields, completed using an email.

What are GPT-6.1 Terra’s API input, cached input, cache-write and output prices?

GPT-6.1 Terra’s Standard API rates are $1.00 for input, $0.05 for cached input, $1.25 for 30-minute cache writes and $5.00 for output, per 1 million tokens. We used the following US dollar prices, read on October 28, 2026.

Standard API prices per million tokens

Codex recorded zero cache-write tokens in every run with our ChatGPT sign-in. OpenAI’s prompt caching guide sets the default 30-minute write rate for GPT-5.6 and later at 1.25 times ordinary input, so we priced all uncached input at that write rate, replacing the ordinary input rate rather than adding a surcharge.

Using ordinary input pricing for those uncached tokens would make every total about 15% to 17% lower. For Terra, retaining a cache write for 12 hours costs $2.50 per million tokens, compared with $1.25 for 30 minutes, before any long-context adjustment.

When a prompt exceeds 272K input tokens, the entire request uses double the input and cache rates and 1.5 times the output rate. None of our requests approached that threshold: even the largest complete run contained 236,046 input tokens.

Other API pricing adjustments apply to these models:

• Batch and Flex each reduce Standard prices by 50%.
• Fast mode doubles the Standard rates.
• Where regional processing is available, it adds a 10% premium.

The task figures are medians from three runs each. The round figure is the median of three rounds, with each round containing all three tasks.

API-equivalent cost for each task

Terra’s round cost 44% of Sol’s, below the 50% that its exactly half-price token rates alone would produce. Sol processed more uncached input in some runs: its third form run had 37,720 uncached tokens, compared with 21,723 to 33,875 across Terra’s form runs. Luna’s round cost about one-seventh of Terra’s despite using more input tokens: a median of 349,343 per round, against Terra’s 261,086 and Sol’s 258,326.

Luna used 165,457 to 236,046 input tokens for each code-fix run; Terra and Sol used about 121,000.

How much did prompt caching affect GPT-6.1 Terra task costs?

GPT-6.1 Terra’s API-equivalent round cost was $0.0784 with caching; pricing the same tokens without caching gave $0.2665, or 3.4 times more. Cached reads accounted for 76.4% of Terra’s input tokens.

Input dominated the estimated bill:

• Output contributed 5% to 7% of each model’s round cost: Terra 6.6%, Sol 5.3% and Luna 6.7%.
• Each run generated 75 to 979 output tokens, including 0 to 187 reasoning tokens at medium effort.
• Codex’s input includes its instructions, tool definitions and file contents, with the conversation sent again at each task step.
• The drawing task alone consumed 25,649 to 26,377 input tokens per run.
• Supplier-form runs consumed between 71,499 and 134,889 input tokens each.
• Code-fix runs consumed between 120,749 and 236,046 input tokens each.

The no-cache comparison applies ordinary input prices to the same tokens, with output included in both estimates.

With and without caching

Terra’s first drawing run had a cold cache, with none of its 26,299 input tokens cached and a cost of $0.0336; the next two each reused 12,160 cached tokens and cost $0.0189 and $0.0190.

What happened in GPT-6 Luna’s supplier-form runs?

GPT-6 Luna completed one supplier-form run correctly out of three; the other two returned entirely empty forms without reading the email. Across all tasks, Luna’s fully correct runs were code fix 3 of 3, technical drawing 2 of 3 and supplier form 1 of 3.

Correct runs, elapsed time and cost

Luna’s incorrect drawing run gave a net area of 9,285 mm² and a mass of 301.1 g; the correct values were 8,999 mm² and 291.6 g. In one empty-form run, Luna reported that Codex’s command tool on our server could not start, returning Failed RTM_NEWADDR: Operation not permitted; the event log for the other contains no command. We therefore do not treat those two empty forms as a measure of Luna’s form-filling ability or assign Luna a cost per correct form.

What we did not measure

• Fast mode, Batch or Flex pricing in actual runs.
• The API itself: the runs used Codex with a ChatGPT sign-in.
• Low or high reasoning effort. Codex defaults to low for Terra; all three models used medium in this test.

FAQs

How much does a GPT-6.1 Terra task cost in Codex?

Our median API-equivalent costs were $0.0250 for the code fix, $0.0190 for the drawing task and $0.0344 for the supplier form, using Standard prices read on October 28, 2026. These were estimates from Codex token usage; our ChatGPT sign-in consumed the plan’s Codex allowance without a per-token bill.

Was GPT-6.1 Terra cheaper per round than GPT-6.1 Sol?

Yes. Terra’s median three-task round was $0.0784 at API-equivalent rates, compared with Sol’s $0.1779. Terra cost 44% of Sol’s round, with both models getting all nine runs right.

Does prompt caching matter for Codex costs?

It substantially changed the API equivalent of our Codex token usage. Terra’s round was $0.0784 with caching, compared with $0.2665 for identical tokens priced without caching, or 3.4 times more. Our ChatGPT sign-in used the plan’s allowance rather than charging those dollar amounts.

Was GPT-6 Luna cheaper per round than GPT-6.1 Terra?

Yes. Luna’s median round cost $0.0115 at API-equivalent rates, about one-seventh of Terra’s $0.0784. Luna fully completed the code fix in 3 of 3 runs, the drawing in 2 of 3 and the form in 1 of 3. The two unsuccessful form runs returned empty fields without reading the email: one reported a command-tool startup failure, while the other run’s event log contains no command.

Did Codex charge per token for these runs signed in with ChatGPT?

No. The runs consumed our ChatGPT plan’s Codex allowance, and we paid nothing per token. The dollar figures estimate what those tokens would cost under OpenAI’s API Standard pricing read on October 28, 2026.

Did GPT-6.1 Terra and GPT-6.1 Sol get every run right?

Yes. GPT-6.1 Terra and GPT-6.1 Sol each finished 9 of 9 runs fully correctly, covering three runs apiece of the code fix, technical drawing and supplier form.

What values did GPT-6 Luna get wrong in the drawing task?

One drawing run returned 9,285 mm² for net area and 301.1 g for mass. The correct answers were 8,999 mm² and 291.6 g. Luna’s other two drawing runs were fully right.

Where does Codex report a task’s token usage?

Codex records each run’s usage in its turn.completed event. It reports input, cached input, cache-write, output and reasoning token counts. Cached input belongs within the input total, while reasoning tokens belong within the output total.

Related on NerdsChalk

• GPT-6.1 Terra Brings Screen Actions Into Its Visual Toolkit