GPT-6.1 Terra halves Sol's API pricing, but maxed-out reasoning can cost more

Terra halves Sol's per-token rates

OpenAI's newly launched GPT-6.1 Terra undercuts its sibling GPT-6.1 Sol by exactly 50% on standard API pricing, according to a developer analysis on dev.to. For prompts under 272K tokens, the post lists Terra at $1.00 per million input tokens and $5.00 per million output tokens, with cached input at $0.05. Sol charges $2.00 and $10.00 respectively, with cached input at $0.10, while the flagship GPT-6 Astra runs $10.00 in and $50.00 out. The launch also came with rate limits the author describes as generous.

Half-price tokens invite a simple migration plan: move everyday pipelines, code generators and browser automation over to Terra, and if quality slips, push the reasoning effort slider all the way up. The post spends the rest of its length explaining why that arithmetic fails in production.

Per-task cost beats per-token price

Figures from Artificial Analysis (AA v4.3.2), cited in the dev.to writeup, weigh prompt, completion, reasoning and cached tokens across standardized developer tasks. At maximum reasoning effort, Terra scored an index of 48.11 at a weighted cost of $0.40 per task; Sol reached 51.83 for $0.72; Astra hit 52.67 for $3.26. Terra at max is roughly 44% cheaper than Sol at max, but scores notably lower.

The crossover appears at intermediate tiers. Sol run at its high tier scored 50.24 for $0.3191 per task: 2.13 index points above Terra at max, yet about 21% cheaper. Even Sol's xhigh tier, at 51.04 for $0.3929, still costs slightly less than Terra maxed out. Terra's own mid tiers are far cheaper — 41.32 for $0.0953 at medium, 43.87 for $0.1343 at high — but the score gap to Sol widens accordingly.

The explanation offered is architectural: a smaller model reasoning through layered dependencies emits long internal chains of thought, frequently down dead ends, while a stronger base model compresses the same problem and finishes with fewer tokens.

The post also relates one anonymized anecdote: a developer moved a backend authentication refactor onto Terra at max effort and ended the day more than 20% over their usual Sol bill, with the agent inventing authorization contracts the team had already settled on and introducing race conditions by the fourth prompt in the session.

Terra's sweet spot and its cliff

LiveBench 2.0 results quoted in the article, all measured at max reasoning, map where Terra holds up:

• Browser automation: Terra 93.1, Sol 93.3, Astra 91.3 — effectively a tie, with Astra occasionally overthinking simple UI steps.
• Code repair: Terra 90.8, Sol 96.0, Astra 93.7.
• Research reports: Terra 73.9, Sol 83.5, Astra 85.2.
• Long-horizon tasks: Terra 48.6 versus Sol's 68.7 and Astra's 77.5 — a 20.1-point deficit.

The pattern is consistent. Deterministic, clearly structured web workflows suit Terra fine. Multi-turn codebase maintenance, where the model must track new constraints while preserving earlier implicit contracts, is where it loses coherence and regresses previously verified logic. According to the post, OpenAI recommends medium as Terra's default tier, where average task cost stays under $0.10, making it strong value for scrapers, monitors and data pipelines.

The division of labour the author proposes is correspondingly split: Terra at medium for operational glue work, Sol for code repair and long-horizon refactors, and Astra reserved for the hardest problems despite its tenfold price tag.

Why it matters

Halving the token rate does not halve the bill if the cheaper model needs many more tokens to finish a job, or to be corrected afterwards. For teams budgeting API spend, the unit that matters is cost per completed task, not per million tokens, and reasoning effort is a pricing lever in its own right. The numbers above come from a single dev.to post drawing on Artificial Analysis and LiveBench 2.0, so they are one data point rather than settled consensus. Still, the operational rule it lands on is worth testing against your own invoices: turning a stronger model down one tier often beats turning a budget model up to max.