New model releases: OpenAI Terra 6.1, Kimi K3.1 Pro, and Kimi K3.1 Pro Highspeed
Summary
GPT-6.1 Terra performed roughly as expected. This might be my new default model for coding tasks. Kimi K3.1 Pro and Pro Highspeed have not impressed me; in my opinion, they are not worth using.
Terra 6.1: an upgrade over Terra 5.6
Terra 6.1 performed roughly as expected. It is an upgrade over Terra 5.6: better quality overall and significantly cheaper, with particularly good results in coding tasks. The only weak spot I observed was inconsistent results when doing extraction from unstructured data. This might be my new default model for coding tasks.
All five workloads, current run
| model | mean quality | cost / task | latency |
|---|---|---|---|
| GPT-6.1 Terra | 0.845 | $0.017 | 16.9 s |
| GPT-5.6 Terra | 0.815 | $0.042 | 18.5 s |
| GPT-6 Luna | 0.804 | $0.0019 | 16.8 s |
Terra 6.1 scored better than Terra 5.6 on repository implementation (0.85 versus 0.79), and much better on architecture (0.60 versus 0.46). Its extraction score was slightly lower (fell from 0.96 to 0.88). On repository implementation, it was cheaper but slower: 40.1 seconds versus 28.7 seconds.
Kimi K3.1 Pro: hard to justify for these tasks
In my opinion, Kimi K3.1 Pro and Pro Highspeed are not worth using. Their quality is slightly better than Terra’s in some tasks and far below Astra’s. They cost a lot and are significantly slower than OpenAI or Anthropic models.
All five workloads, current run
| model | mean quality | cost / task | latency |
|---|---|---|---|
| GPT-6.1 Terra | 0.845 | $0.017 | 16.9 s |
| Kimi K3.1 Pro | 0.836 | $0.077 | 164.8 s |
| Kimi K3.1 Pro Highspeed | 0.836 | $0.233 | 121.9 s |
Kimi Pro scored higher than Terra 6.1 on extraction and matching, tied it on research, and fell behind on architecture and repository implementation. On repository implementation, regular Pro averaged 363.8 seconds, compared with Terra 6.1’s 40.1 seconds.
Pro Highspeed: a small improvement at a higher cost
Pro Highspeed is slightly better than Pro in tasks like extraction, at 3x higher costs.
On extraction, Highspeed scored 0.969 versus Pro’s 0.949, at $0.0885 versus $0.0253 per task. It averaged 45.5 seconds versus 49.2 seconds. Highspeed did not improve quality on architecture or repository implementation.
All five workloads; Astra is historical context from the previous run
| model | mean quality | cost / task | latency |
|---|---|---|---|
| GPT-6.1 Terra | 0.845 | $0.017 | 16.9 s |
| Kimi K3.1 Pro | 0.836 | $0.077 | 164.8 s |
| Kimi K3.1 Pro Highspeed | 0.836 | $0.233 | 121.9 s |
| Astra 6 (previous run) | 0.892 | $0.172 | 28.0 s |
Astra was not rerun in this edition. Its row is historical context from the previous Release Radar, not a fresh head-to-head result. Both Kimi variants tied Astra on research; Astra’s previous scores were higher on the other four workloads.
Testing results
Full results
| model | mean quality | cost / task | latency | Architecture | Repository implementation | Noisy extraction | Semantic matching | Grounded research |
|---|---|---|---|---|---|---|---|---|
| GPT-6.1 Terra | 0.845 | $0.017 | 16.9 s | 0.604 | 0.850 | 0.883 | 0.889 | 1.000 |
| GPT-5.6 Terra | 0.815 | $0.042 | 18.5 s | 0.462 | 0.791 | 0.961 | 0.883 | 0.977 |
| GPT-6 Luna | 0.804 | $0.0019 | 16.8 s | 0.493 | 0.711 | 0.934 | 0.883 | 1.000 |
| Kimi K3.1 Pro | 0.836 | $0.077 | 164.8 s | 0.523 | 0.796 | 0.949 | 0.912 | 1.000 |
| Kimi K3.1 Pro Highspeed | 0.836 | $0.233 | 121.9 s | 0.514 | 0.786 | 0.969 | 0.912 | 1.000 |
| Kimi K2.7 Code | — | — | — | 0.397 | — | 0.938 | 0.854 | 0.941 |
| Claude Fable 5.1 | 0.770 | $0.314 | 44.6 s | 0.407 | 0.686 | 0.922 | 0.941 | 0.895 |
| Claude Opus 5.5 | 0.741 | $0.135 | 31.6 s | 0.439 | 0.744 | 0.934 | 0.754 | 0.834 |
Methodology
We tested five workloads with eight cases per workload, repeated twice, and graded responses against task-specific criteria. Quality scores measure rubric performance, not task pass rates.
Summary charts compare all five workloads, excluding Kimi K2.7 Code because its repository implementation results were incomplete; its available scores remain in the full table. Costs exclude judging and setup, latency measures response time, and Astra results come from the previous run.