GPT-6 Astra costs 2.5× more than GPT-5.6 Sol — but most of the upgrade looks agentic, not reasoning

I compared GPT-6 Astra with GPT-5.6 Sol across the benchmarks where I could find reasonably comparable results.

The headline result is obviously that Astra wins most of them.

But I think the more interesting story is where the improvement actually happens.

GPT-6 Astra costs roughly 2.5× more at API list prices:

- Input: $10 vs $4 / 1M tokens

- Output: $50 vs $20 / 1M tokens

You might expect a similarly large jump in general reasoning capability, but that’s not really what the benchmark pattern shows.

https://preview.redd.it/j07aopmgpsnh1.png?width=1388&format=png&auto=webp&s=16b96ca9b24b1c3e353f7cb2e21e671250176a76

Some more traditional reasoning / knowledge evaluations move only modestly:

- GPQA Diamond: 96.0 vs 94.6

- LifeSciBench: 60.3 vs 59.9

- BrowseComp: 91.5 vs 90.4

The much larger gains appear on agentic and long-horizon workloads:

- MRCR 512K–1M: 96.3 vs 73.8

- Terminal-Bench 4.0: 57.9 vs 37.3

- SRE-Bench: 88.0 vs 55.9

The context result is especially interesting because both models advertise roughly the same 1.05M context window.

So Astra didn’t really get a larger context window — it seems to have become much better at actually using the far end of it.

There’s another weird result in coding agents. On Artificial Analysis Coding Agent v1.4:

- Coding Agent Index: 67 vs 65

- Cost per task: $4.72 vs $5.00

- Terminal-Bench v2.1: 83 vs 83

- DeepSWE: 67 vs 69

- Time per task: 26.8 min vs 10.2 min