One easy-kit round cost $0.0630 on GPT-6.1 Terra and $0.0847 on Claude Sonnet 5.5; hard-kit rounds cost $0.0689 and $0.1137 respectively, with each model getting 9 of 9 hard-kit runs fully right. Both also scored 9 of 9 on the easy kit, where median round times were 31.5 s for GPT-6.1 Terra and 47.1 s for Claude Sonnet 5.5, compared with 44.3 s and 44.8 s on the hard kit. These US dollar amounts price the recorded tokens at API Standard rates: subscription sign-ins covered the runs, and we paid no per-token charges.
- How were GPT-6.1 Terra and Claude Sonnet 5.5 tested?
- How do GPT-6.1 Terra and Claude Sonnet 5.5’s Standard API token prices compare?
- What did a round of easy-kit tasks cost on GPT-6.1 Terra and Claude Sonnet 5.5?
- What did a round of hard-kit tasks cost on GPT-6.1 Terra and Claude Sonnet 5.5?
- Why were the per-round costs closer than the token prices?
- What did the GPT-6.1 Terra and Claude Sonnet 5.5 tests leave unmeasured?
- FAQs
How were GPT-6.1 Terra and Claude Sonnet 5.5 tested?
We ran both kits on a Linux server on October 28, 2026, using GPT-6.1 Terra in Codex CLI 0.159.0 and Claude Sonnet 5.5 in Claude Code 2.1.312, both at medium effort. We made three easy-kit tasks: a Python repair assessed by eight tests, ten questions about a mounting-plate drawing, and an email-based supplier form containing 15 fields. Each model completed three rounds with identical prompts and task files, using fresh folders and running sequentially.
Claude Code had permission to read or change task-folder files and execute the test command only. Its web tools were disabled.
Some of the cost difference comes from the agents’ instructions and tool definitions: Codex also loaded our instruction file of about 7,000 characters, while Claude Code had no instruction files from us.
How do GPT-6.1 Terra and Claude Sonnet 5.5’s Standard API token prices compare?
Claude Sonnet 5.5 charges twice GPT-6.1 Terra’s Standard rates for input, cache reads and output. The following prices are in US dollars for each 1 million tokens, read on October 28, 2026; cache-write rates also depend on retention time.
Standard token rates and cache retention
| Model | Input | Cache read | Cache write | Output |
|---|---|---|---|---|
| GPT-6.1 Terra | $1.00 | $0.05 | $1.25 (30-minute) | $5.00 |
| Claude Sonnet 5.5 | $2.00 | $0.10 | $2.50 (5-minute), $4.00 (1-hour) | $10.00 |
GPT-6.1 Terra also offers 12-hour cache retention at $2.50 for each million tokens written, before any long-context adjustment. Its 30-minute option costs $1.25 for the same token quantity.
GPT-6.1 Terra applies higher rates when a prompt exceeds 272K input tokens: input and caching rates double, while output costs rise to 1.5 times Standard across the entire request. Claude Sonnet 5.5 retains Standard token rates throughout its 1M-token context window.
Other processing options change the listed rates:
- GPT-6.1 Terra’s Batch and Flex options each reduce Standard prices by 50%.
- Claude Sonnet 5.5’s Batch API handles requests asynchronously, reducing both input and output rates by 50%.
- GPT-6.1 Terra’s Fast mode doubles Standard prices.
- GPT-6.1 Terra adds 10% for regional processing wherever that option is available.
These rates price tokens, rather than completed tasks. GPT-6.1 Terra’s rollout across ChatGPT Work, Codex and the API began on October 26, 2026; the runs here used subscription access through the coding agents.
What did a round of easy-kit tasks cost on GPT-6.1 Terra and Claude Sonnet 5.5?
The median easy-kit round cost $0.0630 for GPT-6.1 Terra and $0.0847 for Claude Sonnet 5.5, or about 6 cents and about 8 cents. Claude Sonnet 5.5’s round cost was 1.34 times GPT-6.1 Terra’s, despite its input and output token rates being twice as high.
Easy-kit cost, speed and accuracy
| Measure | GPT-6.1 Terra | Claude Sonnet 5.5 |
|---|---|---|
| Runs fully right | 9 of 9 | 9 of 9 |
| Code fix median cost | $0.0233 | $0.0386 |
| Technical drawing median cost | $0.0194 | $0.0272 |
| Supplier form median cost | $0.0204 | $0.0181 |
| One round median cost | $0.0630 | $0.0847 |
| Rounds for $1, about | 16 | 12 |
| Median time for one round | 31.5 s | 47.1 s |
| Median input tokens per round | 197,078 | 75,415 |
| Input tokens retrieved from cache | 77.5% | 79.8% |
| Output share of the cost | 6.5% | 24.1% |
GPT-6.1 Terra had lower costs for the code repair and drawing, while Claude Sonnet 5.5 cost less for the supplier form. Every run passed the eight code tests, answered the ten drawing questions correctly and completed all 15 form fields correctly.
What did a round of hard-kit tasks cost on GPT-6.1 Terra and Claude Sonnet 5.5?
A hard-kit round had a median API-equivalent cost of $0.0689 on GPT-6.1 Terra and $0.1137 on Claude Sonnet 5.5, a 1.65-times difference. Each model completed all nine runs correctly, with median round times of 44.3 s and 44.8 s respectively.
The harder code task contained seven bugs, with 16 visible tests and eight hidden checks; success required passing every check without modifying the visible tests. The drawing task asked 12 questions about a bracket shown in two views, including four calculations involving area, mass, hole spacing and counterbore volume. The document task required 20 fields from a five-message supplier thread and CSV price list, accounting for later corrections, discounts, invoices, VAT, currency conversion and an inclusive contract end.
The code exercise used a small Python package with these seven faults:
- Handling dates at month boundaries.
- Money rounding using the half-up rule.
- Calculating elapsed hours over daylight-saving transitions.
- Sharing a list used as a default value.
- Resolving ties when ordering jobs.
- Rounding a fraction downward.
- Handling an empty list.
Four visible checks failed before repairs. The eight unseen tests targeted bugs missed by the visible checks, and the prompt disclosed their existence without showing them.
Hard-kit cost, speed and accuracy
| Measure | GPT-6.1 Terra | Claude Sonnet 5.5 |
|---|---|---|
| Runs fully right | 9 of 9 | 9 of 9 |
| Code fix: hidden tests passed | 8 of 8 in every run | 8 of 8 in every run |
| Drawing runs fully right | 3 of 3 | 3 of 3 |
| Document runs fully right | 3 of 3 | 3 of 3 |
| Code fix median cost | $0.0265 | $0.0603 |
| Drawing median cost | $0.0195 | $0.0294 |
| Document median cost | $0.0219 | $0.0249 |
| One round median cost | $0.0689 | $0.1137 |
| Rounds for $1, about | 15 | 9 |
| Median time per round | 44.3 s | 44.8 s |
| Median input tokens per round | 224,860 | 91,854 |
| Median output tokens per round | 1,395 | 3,478 |
| Output share of the cost | 10.1% | 31.0% |
Neither GPT-6.1 Terra nor Claude Sonnet 5.5 had an incorrect hard-kit run. Both cleared the visible and hidden code checks, and every drawing and document run was fully correct.
Why were the per-round costs closer than the token prices?
Claude Code sent Claude Sonnet 5.5 fewer input tokens: the easy-kit median was 75,415 per round, compared with 197,078 for GPT-6.1 Terra in Codex. For the easy drawing alone, Claude Sonnet 5.5 used 10,788 to 10,791 input tokens per run, compared with 26,660 for GPT-6.1 Terra. The estimates also reflect caching: Codex reported no cache-write tokens, so GPT-6.1 Terra’s uncached input was priced at $1.25 per million, while Claude Code identified its writes as 1-hour entries, priced at $4.00 per million for Claude Sonnet 5.5.
For each Claude Sonnet 5.5 run, our calculated amount agreed with Claude Code’s own cost report.
Using the alternative method, with GPT-6.1 Terra’s uncached input at $1.00 per million and Claude Sonnet 5.5’s writes at the 5-minute rate, median easy-kit rounds cost $0.0529 and $0.0628 respectively, while hard-kit rounds cost $0.0583 and $0.0866.
Our earlier cost test likewise expressed subscription-based Codex usage as API-equivalent dollar amounts.
What did the GPT-6.1 Terra and Claude Sonnet 5.5 tests leave unmeasured?
The tests measured the models through their makers’ coding agents with subscription sign-ins. They did not cover:
- Direct API runs.
- Effort settings other than medium.
- Batch, Flex or Fast mode pricing.
FAQs
Is GPT-6.1 Terra cheaper than Claude Sonnet 5.5 per round?
GPT-6.1 Terra had lower median API-equivalent round costs in both kits. The easy-kit figures were $0.0630 for GPT-6.1 Terra and $0.0847 for Claude Sonnet 5.5; the hard-kit figures were $0.0689 and $0.1137 respectively.
Did Claude Sonnet 5.5 or GPT-6.1 Terra get more hard-kit runs fully right?
They tied: GPT-6.1 Terra and Claude Sonnet 5.5 each got 9 of 9 hard-kit runs fully right. Both passed all eight hidden code tests in every code run and completed every drawing and document run correctly.
Which was faster per easy-kit round, GPT-6.1 Terra or Claude Sonnet 5.5?
GPT-6.1 Terra had the shorter median easy-kit round time at 31.5 s, compared with 47.1 s for Claude Sonnet 5.5. Hard-kit median times were 44.3 s for GPT-6.1 Terra and 44.8 s for Claude Sonnet 5.5.
Why did Claude Sonnet 5.5’s per-task cost stay closer to GPT-6.1 Terra’s than their token prices suggest?
Claude Sonnet 5.5 processed fewer input tokens through Claude Code. Its easy-kit median was 75,415 input tokens per round, against 197,078 for GPT-6.1 Terra through Codex, narrowing the cost difference despite higher per-token rates; the agents’ different instructions and tool definitions also contributed to the gap.
Did the ChatGPT and Claude subscription sign-ins incur per-token charges for these tests?
No. The ChatGPT and Claude subscription sign-ins covered these runs without per-token charges. The dollar figures estimate API Standard costs for the recorded token usage; they were not billed amounts.
Related on NerdsChalk
Leave a Reply