- The short version
- How were GPT-6.1 Terra and Claude Haiku 4.5 tested?
- What are GPT-6.1 Terra and Claude Haiku 4.5 API Standard prices?
- How did GPT-6.1 Terra and Claude Haiku 4.5 compare on the easy kit?
- How did GPT-6.1 Terra and Claude Haiku 4.5 compare on the hard kit?
- Why did GPT-6.1 Terra and Claude Haiku 4.5 cost different amounts per task?
- What did these tests leave unmeasured?
- FAQs
The short version
- Easy and hard rounds cost GPT-6.1 Terra $0.0630 and $0.0689, versus Claude Haiku 4.5 $0.1393 and $0.2127; hard-kit runs fully right were 9 of 9 versus 3 of 9.
- Both models got 9 of 9 easy-kit runs fully right. Median round times were 31.5 seconds for Terra and 105.7 seconds for Haiku 4.5.
- Both list $1.00 per million input tokens and $5.00 per million output tokens, but Haiku 4.5 generated much more output in these runs.
On the easy kit, one three-task round cost $0.0630 with GPT-6.1 Terra and $0.1393 with Claude Haiku 4.5; both got 9 of 9 runs fully right, with median round times of 31.5 and 105.7 seconds. On the hard kit, a round cost $0.0689 and $0.2127, with 9 of 9 and 3 of 9 runs fully right; median times were 44.3 and 178.4 seconds. We used subscription sign-ins and paid nothing per token, so those dollar figures are what the recorded tokens would cost at API Standard prices.
How were GPT-6.1 Terra and Claude Haiku 4.5 tested?
On October 28, 2026, we ran three rounds on a Linux server, using a fresh folder for each run and medium effort: GPT-6.1 Terra ran in Codex CLI 0.159.0 through codex exec, while Claude Haiku 4.5 ran in Claude Code 2.1.312 through claude -p. Each model received the same tasks and prompt text: a Python code fix, questions about a technical drawing we made, and a supplier form completed from an email. The hard kit used the same task types, with hidden code tests, a two-view drawing with calculations, and an email thread that corrected earlier details. Because Codex and Claude Code supply different instructions and tool definitions, part of the cost gap comes from the agents as well as the models.
Our Codex setup supplied an additional instruction file of about 7,000 characters; Claude Code had no instruction files from us. Claude Code could inspect and change task-folder files and run the test command, with web tools disabled.
Terra became available in Codex following its October 26, 2026 launch. Our earlier Terra task-cost test covers a separate set of Codex runs.
What are GPT-6.1 Terra and Claude Haiku 4.5 API Standard prices?
Both models list $1.00 per million input tokens and $5.00 per million output tokens. Their cache-read rates differ. These US dollar Standard API prices were read on October 28, 2026.
Standard prices per million tokens
| Model | Input | Cache read | Cache write | Output |
|---|---|---|---|---|
| GPT-6.1 Terra | $1.00 | $0.05 | $1.25, 30-minute retention | $5.00 |
| Claude Haiku 4.5 | $1.00 | $0.10 | $1.25, 5-minute retention; $2.00, 1-hour retention | $5.00 |
How did GPT-6.1 Terra and Claude Haiku 4.5 compare on the easy kit?
Both models completed all 9 easy-kit runs correctly, while Terra’s median three-task round cost $0.0630 and took 31.5 seconds, against $0.1393 and 105.7 seconds for Haiku 4.5.
Easy-kit cost, accuracy and time
| Measure | GPT-6.1 Terra | Claude Haiku 4.5 |
|---|---|---|
| Runs fully right | 9 of 9 | 9 of 9 |
| Code fix, median cost | $0.0233 | $0.0607 |
| Technical drawing, median cost | $0.0194 | $0.0255 |
| Supplier form, median cost | $0.0204 | $0.0367 |
| One round, median cost | $0.0630 | $0.1393 |
| All nine runs together | $0.2027 | $0.3900 |
| Rounds for $1, about | 16 | 7 |
| Median time for one round | 31.5 seconds | 105.7 seconds |
| Median input tokens per round | 197,078 | 145,634 |
| Share of input read from cache | 77.5% | 81.2% |
| Output share of cost | 6.5% | 47.9% |
Haiku 4.5’s median round cost 2.21 times Terra’s despite their matching base input and output rates. Every easy-kit code run passed its 8 tests, every drawing run answered 10 questions correctly, and every supplier form had all 15 fields right. Haiku 4.5 took 31.5 to 60.4 seconds per code fix, while Terra took 15.3 to 19.5 seconds.
How did GPT-6.1 Terra and Claude Haiku 4.5 compare on the hard kit?
Terra got 9 of 9 hard-kit runs fully right, compared with 3 of 9 for Haiku 4.5. Median rounds cost $0.0689 and $0.2127, and took 44.3 and 178.4 seconds, respectively.
The hard-kit code task contained a small Python package with seven distinct bugs. A code run counted as fully right only if all 24 tests passed without changes to the visible tests.
The two-view drawing posed 12 questions. Four required calculations covering net face area, revised-flange batch mass, spacing between bolt-circle holes and counterbore volume.
Hard-kit results across three rounds
| Measure | GPT-6.1 Terra | Claude Haiku 4.5 |
|---|---|---|
| Runs fully right | 9 of 9 | 3 of 9 |
| Code fix, hidden tests passed | 8 of 8 in every run | 3 of 8 in every run |
| Drawing runs fully right | 3 of 3 | 1 of 3 |
| Document runs fully right | 3 of 3 | 2 of 3 |
| Code fix, median cost | $0.0265 | $0.0729 |
| Drawing, median cost | $0.0195 | $0.0693 |
| Document, median cost | $0.0219 | $0.0609 |
| One round, median cost | $0.0689 | $0.2127 |
| Rounds for $1, about | 15 | 5 |
| Median time per round | 44.3 seconds | 178.4 seconds |
| Median input tokens per round | 224,860 | 307,457 |
| Median output tokens per round | 1,395 | 21,598 |
| Output share of cost | 10.1% | 53.1% |
Haiku 4.5’s median hard-kit round cost 3.09 times Terra’s. Haiku 4.5 passed all 16 visible code tests in each run, but passed only 3 of 8 hidden tests each time, leaving the bugs checked by the other hidden tests unfixed in the code. In two drawing runs, it calculated batch mass as 1,061.1 g instead of 2,113.4 g. In one document run, it entered the discounted price, 17.58, where the tier unit price was 18.50, affecting the later money fields; it also gave the contract end as 2028-06-10 instead of 2028-07-10.
Why did GPT-6.1 Terra and Claude Haiku 4.5 cost different amounts per task?
On the easy kit, Haiku 4.5 produced 1,226 to 7,337 output tokens per run, mostly thinking tokens (1,089 to 5,727); Terra produced 129 to 457 output tokens, including 9 to 59 reasoning tokens. Output made up 47.9% of Haiku 4.5’s easy-kit cost and 6.5% of Terra’s, even though Haiku 4.5 used fewer median input tokens per round. On the hard kit, Haiku 4.5 used a median 307,457 input and 21,598 output tokens per round, versus Terra’s 224,860 input and 1,395 output tokens; output accounted for 53.1% and 10.1% of their costs.
Codex recorded no cache-write tokens, so we charged Terra’s uncached input at the $1.25 rate for a 30-minute cache write, following OpenAI’s caching guidance for GPT-5.6 and later models. Claude Code classified its cache writes as one-hour writes; our calculated costs matched the amounts Claude Code reported for every run.
The figures above price Terra’s uncached input at its 30-minute cache-write rate and Haiku 4.5’s reported cache writes at its 1-hour rate; using $1.00 for Terra’s uncached input and Haiku 4.5’s 5-minute write rate instead gives median rounds of $0.0529 versus $0.1154 on the easy kit, and $0.0583 versus $0.1871 on the hard kit.
What did these tests leave unmeasured?
- Direct API runs; the tests used each maker’s coding agent with a subscription sign-in.
- Effort settings other than medium.
- Batch, Flex and Fast mode pricing.
FAQs
Is GPT-6.1 Terra cheaper per task than Claude Haiku 4.5?
In these tests, yes. A median three-task round cost $0.0630 on Terra versus $0.1393 on Haiku 4.5 for the easy kit, and $0.0689 versus $0.2127 for the hard kit, when priced at API Standard rates.
Which got more hard-kit runs fully right, GPT-6.1 Terra or Claude Haiku 4.5?
GPT-6.1 Terra got 9 of 9 hard-kit runs fully right; Claude Haiku 4.5 got 3 of 9. Terra passed all 8 hidden code tests in every run, while Haiku 4.5 passed 3 of 8 in every run.
Which is faster, GPT-6.1 Terra or Claude Haiku 4.5?
GPT-6.1 Terra finished a median three-task round in 31.5 seconds on the easy kit and 44.3 seconds on the hard kit. Claude Haiku 4.5 took 105.7 and 178.4 seconds, respectively.
Did the GPT-6.1 Terra and Claude Haiku 4.5 subscription tests incur per-token charges?
No. The runs used ChatGPT and Claude subscription sign-ins, and we paid nothing per token. The reported costs price the recorded token usage at each maker’s API Standard rates.
What did the GPT-6.1 Terra vs Claude Haiku 4.5 tests leave unmeasured?
They did not run the APIs directly, test other effort levels, or measure Batch, Flex or Fast mode prices.
Why compare GPT-6.1 Terra with the older Claude Haiku 4.5?
Claude Haiku 4.5 is an older Claude model. We chose it because its listed $1.00 input and $5.00 output prices per million tokens match GPT-6.1 Terra’s.
Related on NerdsChalk
Leave a Reply