Factr-I vs Prime
Factr-I scored higher than both Prime builds on five of the six benchmarks and tied the Rust build on the sixth. It finished the same 1,067 tasks on 68.1M tokens for $3.81, against 135.2M for $4.61 from Prime v0.9.8 and 159.7M for $13.49 from the Rust build. Prime's Rust build holds lower peak memory on four of the six benchmarks.
Every harness ran the same 1,067 tasks from six public benchmarks on GPT-6 Luna at medium reasoning effort. Results as of 2026-10-12.
Scores
| Benchmark | Factr-I | Prime (Rust) | Prime |
|---|---|---|---|
| Long-context pairsOOLONG-Pairs · 20 tasks | 73.8%F1 | 23.4%F1 | 32.7%F1 |
| Long-context reasoningOOLONG · 50 tasks | 90.0%45/50 | 48.0%24/50 | 60.0%30/50 |
| General assistantGAIA · 165 tasks | 40.0%66/165 | 39.4%65/165 | 37.6%62/165 |
| Terminal tasksTBLite · 100 tasks | 68%68/100 | 59%59/100 | 64%64/100 |
| Code editingPolyglot · 225 tasks | 86.2%194/225 | 70.2%158/225 | 73.3%165/225 |
| Long reasoningLongCoT-Mini · 507 tasks | 19.3%98/507 | 19.3%98/507 | 17.6%89/507 |
Peak memory
| Benchmark | Factr-I | Prime (Rust) | Prime |
|---|---|---|---|
| Long-context pairsOOLONG-Pairs · 20 tasks | 84MiB | 72MiB | 627MiB |
| Long-context reasoningOOLONG · 50 tasks | 96MiB | 71MiB | 618MiB |
| General assistantGAIA · 165 tasks | 53MiB | 103MiB | 637MiB |
| Terminal tasksTBLite · 100 tasks | 27MiB | 46MiB | 778MiB |
| Code editingPolyglot · 225 tasks | 390MiB | 255MiB | 826MiB |
| Long reasoningLongCoT-Mini · 507 tasks | 68MiB | 60MiB | 609MiB |
Tokens and cost
| Benchmark | Factr-I | Prime (Rust) | Prime |
|---|---|---|---|
| Total tokensAll 1,067 tasks | 68.1M | 159.7M | 135.2M |
| Total costAll 1,067 tasks | $3.81 | $13.49 | $4.61 |
Where each one leads
Where Factr-I leads
- Higher score on five of six benchmarks, and a tie with the Rust build on long reasoning.
- Widest gaps on long context: 73.8% against 32.7% on OOLONG-Pairs, and 90.0% against 60.0% on OOLONG.
- Fewest tokens of any harness in the report, and the lower dollar cost against both Prime builds.
Where Prime leads
- Prime's Rust build uses less peak memory on four of six benchmarks, including 255 MiB against 390 MiB on code editing.
- On the general assistant benchmark the gap is 0.6 of a point, 40.0% against 39.4%, which is close to a tie.
Questions
Is Factr-I more accurate than Prime?
On these benchmarks, yes. It scored higher on five of six and tied the Rust build on long reasoning.
Which costs less to run?
Factr-I, on this report. It used 68.1M tokens for $3.81, against $4.61 for Prime v0.9.8 and $13.49 for the Rust build.
Why are there two Prime columns?
Prime rewrote itself in Rust on October 9, 2026. The earlier build ran first, then the Rust build once it was released.
Does Prime beat Factr-I anywhere?
Yes. The Rust build holds lower peak memory per task on four of the six benchmarks.