GPT-6.1 Terra has lower published Standard API rates than GPT-6 Astra for matching requests, while GPT-6 Astra scored higher across three synthetic tasks reported by The New Stack’s author. In those single runs, Terra finished sooner and had lower estimated token costs. The results point to a trade-off, not a universal winner.
API rates depend on context and processing mode
For Standard requests with up to 272,000 input tokens, OpenAI lists Terra at $1 per million input tokens and $5 per million output tokens. Astra’s rates are $10 and $50, respectively. The short-context rates are one-tenth as high for Terra. OpenAI also lists separate cached-input and long-context rates in its API pricing documentation.
| Model | Short-context input | Cached input | Short-context output | Long-context input | Cached input | Long-context output |
| GPT-6.1 Terra | $1 | $0.05 | $5 | $2 | $0.10 | $7.50 |
| GPT-6 Astra | $10 | $1 | $50 | $20 | $2 | $75 |
Standard rates in USD per million tokens. Short context means up to and including 272,000 input tokens; long context means more than 272,000 input tokens.
OpenAI lists Batch and Flex rates at half the Standard price, Fast mode at twice the Standard price, and a 10% premium for eligible regional-processing endpoints. Those conditions change the applicable rate. Total task cost also depends on token use and any tool charges, so the per-token price alone does not predict a bill.
Both models have the same published limits: a 1,050,000-token context window, up to 922,000 input tokens, and up to 128,000 output tokens. Both accept text and images and return text. Terra’s documented reasoning settings run from none through max; Astra’s start at low. OpenAI describes Terra for visual reasoning, computer use, and writing, while positioning Astra for demanding reasoning, coding, research, and computer use. See the GPT-6.1 Terra and GPT-6 Astra model pages for their specifications.
How the models scored on three synthetic tasks
The New Stack’s author ran each final task once per model at high reasoning effort, using synthetic data. Astra scored higher overall, with 148 points out of 150 to Terra’s 121. The task scores were:
| Task | GPT-6.1 Terra | GPT-6 Astra |
| SQL reasoning | 26/50 | 48/50 |
| Data validation | 45/50 | 50/50 |
| Incident-log triage | 50/50 | 50/50 |
| Total | 121/150 | 148/150 |
In the validation run, Terra flagged 35 of 40 invalid records and did not flag any of the 10 valid records that looked suspicious. Astra found all 40 invalid records. Both models received full marks for triaging a fabricated outage from about 7,000 log lines across more than 20 services.
Across the three runs, Terra took 258.8 seconds in total, compared with 709.2 seconds for Astra. The author estimated token costs of $0.61 for Terra and $7.88 for Astra, using Standard uncached rates and excluding tools and billing adjustments. These are estimates for those runs, not invoices or general cost predictions.
What the results can tell you about choosing a model
Terra’s lower published rates matter most when the requests fit the relevant context and processing conditions. Its shorter total time and lower estimated cost also appeared in these three runs. Astra’s higher total score came from stronger results on SQL reasoning and validation; the models tied on incident triage.
The tests do not settle which model is better for software engineering or everyday work. The tasks used synthetic data, each final version was run once per model, and the author says the results are not representative of ordinary work. The same high-effort setting also did not ensure equal reasoning-token budgets. For a coding decision, the task scores here are narrow evidence, not a general coding ranking.
As a practical starting point, compare Terra when lower published token rates or OpenAI’s stated focus on visual reasoning and writing fit your workload. Consider Astra when its published focus on demanding reasoning, coding, or research better matches the job. Then account for the context band, processing mode, token use, and tool charges that apply to your API requests.
A separate benchmark snapshot
Lifey Guy’s English-language comparison video reports separate figures attributed to Artificial Analysis: Intelligence Index v4.3 scores of 48 for Terra and 53 for Astra, Terminal-Bench 4.0 results of 56% and 59%, and SciCode results of 55% and 56%, respectively. These figures offer another benchmark perspective, distinct from the three synthetic task runs.
For background on Astra’s agent-oriented positioning, see NeoTeo’s earlier coverage of GPT-6 Astra.
