Richard Thomas

This is a "let's connect" post wrapped in a physics simulation so it doesn't feel like one. Crossed 200 followers. Every coin is one of you. The jar holds 500. I build things in public and occasionally finish them. If you do too, say hi.

on my coding benchmark sol is already close to that: 74.1 vs opus 5.5's 75.5 on the same 3 seeds, at $0.47 a run vs $5.35. that's 11x the cost for 1.4 points caveat is that it's only in my little benchmark and not on your real codebase

the "half the cost" part holds up in codex too: across 3 runs terra used almost the exact same token count as sol (1.96m vs 1.94m) so it landed at $0.22 vs $0.47, and ran 2.5x faster wall-clock on isolated algorithms like maze solving it's right on sol's heels. the only place it really drops points against sol is long-horizon state where multiple scripts have to feed each other instead of running in isolation people are being way too critical imo

Outperforming on what?

ran gpt-6.1 terra through three 6-hour seeds in codex (medium effort) on farmerbench to see how it holds up on longer coding loops. it averaged 62.16 (58.47 / 66.34 / 61.68) at $0.22 and 2m 56s a run across 147 tool calls, opening all 11 resources on every seed. when a locked hat crashed its script on seed 3, it read the docs, wrote a one-line fix and recovered in a minute. the catch vs 6.1 sol (74.08 at $0.47): terra wrote 17 to 18 separate small scripts per seed and only edited a saved program twice across 54 writes. great for fast bounded tasks, less keen on refactoring when the grid scales to 22×22.

seeing the exact same ordering in codex on farmerbench (three 6-hour seeds each at medium effort): gpt-6.1 sol: 74.1 mean at $0.47 a run (159 points/$) gpt-6 astra: 66.6 mean at $1.79 a run (37 points/$) gpt-6.1 terra: 62.2 mean at $0.22 a run (286 points/$) gpt-6 luna: 15.7 mean at $0.08 a run (189 points/$) at 286 points per dollar and 2m 56s a run, terra sits right on the pareto frontier between deepseek v4.1 flash ($0.13/run, 46.8 mean) and 6.1 sol ($0.47/run, 74.1 mean).

the 2x usage math vs 6.1 sol holds up almost to the cent in codex. across three 6-hour seeds on farmerbench (medium effort), terra used 1.96m tokens a run vs sol's 1.94m, so the run cost dropped from $0.47 to $0.22 (286 points per dollar) and wall time dropped from 7m 35s to 2m 56s, the fastest on the board. it averaged 62.2 (58.5 / 66.3 / 61.7) and beat gpt-6 astra on seed 2 (66.3 vs 63.2 at $1.79), though sol still leads on all three seeds (74.1 mean) once the codebase scales up to a 32×32 grid.

GPT-6.1 Terra played six hours of a coding game in about 3 minutes. Same game, same three seeds: Terra: 62 pts · 3 min · $0.22 a run GPT-6.1 Sol: 74 pts · 8 min · $0.47 GPT-6 Astra: 67 pts · 5 min · $1.79 Very fast. Not quite Sol. 🎥

daily reminder that you can now just think of stuff and then type stuff and then have stuff and that's pretty cool

fwiw the gap between token price and what a finished agent run actually costs is still wild once the harness gets involved. across three 6-hour seeds each on farmerbench, points per dollar spans 78x: deepseek v4.1 flash in opencode hits 368 points per dollar ($0.13/run, 46.84 mean) and gpt-6.1 sol in codex hits 159 ($0.47/run, 74.08 mean over 134 tool calls). claude opus 5.5 in claude code lands at 14 points per dollar ($5.35/run, 75.54 mean, 248 calls) and fable 5.1 drops to 4.7 ($14.68/run, 69.17 mean). cheap tokens help, but loop length and cache reads are what dictate a lot of the cost

that 2/20 where opus had the right answer in context and still got it wrong shows up constantly in my farmerbench logs too. across three 6-hour seeds in claude code, haiku 5.5 read the docs 47 times a run vs 31 for opus 5.5, opened the exact page explaining how fertilizer makes weird substance, and still never used fertilizer on any seed (36.23 mean at $0.81/run vs opus at 75.54 and $5.35/run). fable 5.1 did the same with companions on seed 3: 60 doc reads, bought every upgrade level, planted zero companions

the funny thing is the bar for 6.1 terra and fable 5.5 on my coding benchmark isn't even opus right now, it's gpt-6.1 sol. over three 6-hour seeds on farmerbench, sol in codex scored a 74.08 mean (71.82 / 75.87 / 74.56) at $0.47 a run, beating both gpt-6 astra at 66.65 ($1.79/run) and fable 5.1 at 69.17 ($14.68/run). fable 5.1 cost nearly 3x opus 5.5 ($5.35/run, 75.54 mean) and still trailed sol on two of the three seeds

Seed 1 thought of the straw hat twice and never tried it. It also bought four drones and farmed with one. Caveats: one benchmark, three seeds. High effort (Kimi has no medium), on a Kimi Code subscription; $2.23/run is at list prices. farmerbench.dev/models/kimi-…