GPT-6.1 Terra Costs Half of Sol Per Token. It Burns 5.5x the Tokens, So Sol Is Still Cheaper.

What OpenAI shipped

GPT-6.1 Terra is OpenAI’s new mid-tier model. Its model page calls it “a fast, efficient model for visual reasoning, precise computer use, and clear, natural writing.” The API ID is gpt-6.1-terra, with a 1,050,000-token context window, 128,000 max output tokens and a June 1, 2026 knowledge cutoff.

The price is the headline: $1 per million input tokens and $5 per million output tokens. That is exactly half of GPT-6.1 Sol ($2 / $10) and a tenth of GPT-6 Astra ($10 / $50). The previous Terra, GPT-5.6 Terra, cost $2 / $12.

According to BleepingComputer’s report of the launch, OpenAI says Terra performs on par with GPT-6 Astra at one-tenth the price per token, finishes tasks more than 3 times faster than Astra and Sol, and scores 50.3% on Terminal-Bench 4.0 at maximum reasoning effort. The same report notes OpenAI still suggests Sol or Astra for the most complex coding tasks. Terra ships in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users.

Vals.ai added Terra to its Terminal-Bench 4.0 leaderboard the same day. It scored 49.49%, within a point of OpenAI’s own 50.3%. The score holds up. The cost story is where it gets interesting.

The OpenAI lineup on Terminal-Bench 4.0

Terminal-Bench 4.0 is 66 long, real-deliverable tasks across software, science, ML, operations, hardware, security and media. Vals.ai runs every model through the same minimal harness, mini-swe-agent, with a single bash tool and “no step or cost limit.” Each score is the mean of 3 full runs.

Source: Vals.ai Terminal-Bench 4.0 leaderboard, updated October 26, 2026 (score, cost, tokens). Prices: OpenAI API pricing and model pages, standard tier. GPT-6 Sol’s price is not on OpenAI’s current pricing page, so it is left blank.

Where the extra tokens go

Vals.ai publishes token totals for each model. Terra and Sol produce similar amounts of output: 7.10 million output tokens for Terra, 6.31 million for Sol, about 12% apart. The gap is on the input side: 1.02 billion input tokens for Terra against 182 million for Sol.

Almost all of that is cache reads. By our math on Vals.ai’s totals, 98.8% of Terra’s input tokens were read from cache, against 94.7% for Sol. Cached input on Terra costs $0.05 per million, so 5.5 times the tokens only becomes an 18% bigger bill. Without caching it would be far worse.

Sol and Astra are outliers in the other direction: around 3 million tokens per test, the leanest of any model near the top of the board. Terra behaves like most of the field, not like its two bigger siblings.

Long contexts carry a second price trap. OpenAI’s Terra page says prompts with more than 272K input tokens are “priced at 2x input and cache rates and 1.5x output for the full request.” An agent that keeps re-reading a growing context can cross that line, and then the half-price rate no longer applies to that request.

Is Terra really “on par” with Astra?

Not on this benchmark. Astra scores 59.60%, Terra 49.49%: 10.1 points apart. Terminal-Bench 4.0 has 66 tasks, so that is roughly 6.7 tasks per run. Vals.ai lists Terra’s standard error at 1.821 points, far smaller than the gap.

The price claim holds better than the parity claim. Terra’s per-token rate is a tenth of Astra’s, and its cost per test is $2.04 against $9.58, about a fifth. That is a real saving, just not the 10x the rate card suggests.

Speed is the same story. OpenAI says “more than 3x faster” than Astra and Sol. On Vals.ai, Terra averages 32m45s per test, Astra 35m42s and Sol 45m12s. That duration includes the time commands take to run in the sandbox, so it is not a pure model-speed number. Still, on full agent tasks, the gap is 8% against Astra and 28% against Sol, not 3x.

One caveat in Terra’s favor: Vals.ai doesn’t state which reasoning effort it used for Terra, and the model supports six levels, from none to max, with medium as the default. A different setting would move both the score and the token count.

Terra still made a huge jump

Compare Terra with the model it replaces and the release looks great. GPT-5.6 Terra scored 22.73% at $5.60 per test. GPT-6.1 Terra scores 49.49% at $2.04: 26.76 more points for 64% less per test.

If you were running GPT-5.6 Terra in a coding agent, switching is an easy call. The question is only whether Terra is the right new default, and on these numbers Sol is the better buy for agent work.

Cost per point = Vals.ai cost per test ÷ Vals.ai score, rounded to the cent. A rough efficiency read, not a Vals.ai metric.

Across the whole board, Terra lands 10th of 56 models and is one of the cheapest in the top 10. Only Sol is cheaper per test among the 10 highest scores. That’s a strong result for a mid-tier model. It just isn’t the cheapest way to run OpenAI on agent tasks.

What this means if you code with AI

For agent runs, try GPT-6.1 Sol before Terra. On Terminal-Bench 4.0, Sol scores 5.56 points higher, costs 16% less per test and uses about a fifth of the tokens. If you pay OpenAI per token for long agent loops, Sol is the default these numbers point to.

Use Terra where its rate can win. Short requests that don’t loop, like chat, single-file edits, screenshots of a UI or computer-use steps, are billed close to the rate card. There, half the rate means close to half the bill, and OpenAI pitches Terra for exactly that work.

On a subscription, watch limits, not prices. In Codex or ChatGPT you don’t see a per-token bill. You see usage limits, and a model that reads 5.5 times the tokens can use them up faster, whatever its rate card says.

Measure cost per finished task on your own repo. Terminal-Bench is long, strict and roughly three-quarters outside traditional software engineering. Run a week of your own tickets on 2 models and compare the bills and the pass rate before switching.

Cap your agent’s spend. Vals.ai runs agents with no step or cost limit for up to 8 hours per task. Your own setup may not stop sooner unless you tell it to.

Where MegaBrain fits

Full disclosure: we sell an AI coding agent. MegaBrain’s terminal coding agent runs GLM, Kimi, MiniMax, DeepSeek, Qwen, Grok and Auto on the paid plans, Claude, GPT and Gemini from credits from Pro, and any model through your own API key or local models, free on every plan. So you can run the same task on two models and compare the bills yourself. Plans are on the pricing page. MegaBrain is not affiliated with Vals.ai or OpenAI.

Sources

All pages opened October 27, 2026.

• Vals.ai — Terminal-Bench 4.0 leaderboard and methodology (scores, cost, tokens and duration per test; updated October 26, 2026)
• OpenAI — GPT-6.1 Terra model page (price, context window, reasoning effort levels, long-context pricing)
• OpenAI — API pricing (GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna)
• OpenAI — GPT-5.6 Terra model page (previous Terra price)
• BleepingComputer — “OpenAI launches GPT-6.1 Terra with 3x faster AI agents, 50% lower pricing” (OpenAI’s launch claims)