Alan West

Posted on Oct 27 Originally published at blog.authon.dev

GPT-6.1 Terra vs Claude Opus 5.5: 15x Cheaper per Task. Here's the Catch

#ai #openai #programming #productivity

On Artificial Analysis, GPT-6.1 Terra at max effort costs $0.40 per task and Claude Opus 5.5 costs $5.98. That's about 15x. Opus also scores 58 on their Intelligence Index to Terra's 48.1, and those ten points are the gap OpenAI's launch charts work hard to keep you from seeing.

Terra shipped on October 26. Everyone is googling the same comparison this week, so I collected the published numbers for GPT-6.1 Terra vs GPT-6 (Astra, plus 6.1 Sol in between) vs Claude Opus 5.5 into one place: OpenAI's launch post, the model and pricing pages, Artificial Analysis, and the Hacker News thread. I didn't run anything myself. Every number below comes from a source I link to.

GPT-6.1 Terra vs Sol vs GPT-6 Astra vs Opus 5.5: pricing and benchmarks

	GPT-6.1 Terra	GPT-6.1 Sol	GPT-6 Astra	Claude Opus 5.5
Input / output per 1M tokens	$1 / $5	$2 / $10	$10 / $50	$4 / $20
Cached input per 1M	$0.05	$0.10	$1.00	see note
Context window	1.05M	1.05M	1.05M	1M
Max output	128K	128K	128K	128K (300K via Batch)
AA Intelligence Index (max)	48.1	51.8	52.7	58
AA cost per task (max)	$0.40	$0.72	$3.26	$5.98
AA time per task (max)	384s	543s	399s	n/a
AA Terminal-Bench 4.0	55.6%	n/a	59.6%	59.6%
AA hallucination rate	38%	54%	51%	n/a

Sources: OpenAI pricing and model pages via eesel's roundup and cellcog, Artificial Analysis via eesel, BenchLM and The Batch. Opus 5.5 cache pricing differs between write-ups, so check Anthropic's page before you budget around it.

A few things the table leaves out:

Terra's model ID is gpt-6.1-terra. Knowledge cutoff is June 1, 2026.
In ChatGPT, Terra is available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. It's also in the API, which has no free tier.
Prompts over 272K input tokens cost $2 / $7.50, and that rate applies to the whole request. Batch and Flex cost half of standard: $0.50 / $2.50.
Terra costs half what its predecessor GPT-5.6 Terra did on input and well under half on output. GPT-5.6 Terra was $2 / $12.

Where OpenAI's charts flatter Terra

OpenAI's headline claim is that at max effort Terra is more than 3x as token-efficient as Opus 5.5 at a quarter of the per-token price, which it says works out to roughly 9x lower cost per task. That comes from Agents' Last Exam, and OpenAI's own chart gives the numbers:

Agents' Last Exam (max)	Score	Est. cost per task
GPT-6.1 Terra	60.0%	$1.01
GPT-6.1 Sol	58.6%	$2.00
GPT-6 Astra	59.3%	$7.24
Claude Opus 5.5	63.2%	$8.82

The cost math holds up. The trouble is that OpenAI picked the benchmark where Terra lands closest to Opus. Look at what else is published:

Terminal-Bench 4.0: OpenAI reports Terra at 50.3% and compares it only with GPT-5.6 Terra (21.5%), not with Opus. Anthropic reports 66.4% for Opus 5.5. On Artificial Analysis's harness it's 55.6% for Terra and 59.6% for Opus. The two harnesses differ, so don't mix their numbers, but Terra trails on both.
Knowledge and science: on AA's numbers, Terra gets 46.2% on HLE to Opus's 61.4%, and 55.1% on SciCode to Opus's 66.9%. Neither appears in the launch post.
Missing comparisons: commenters on HN noticed that Opus appears only at Max and never at lower effort levels. It's also missing from the OSWorld chart, and Anthropic models are absent from the factual-error chart. OSWorld 2.1 and MMMU-Pro are compared only with OpenAI's own models: on computer use, Terra at medium scores 15 points above GPT-5.6 Terra and "on par with GPT-6 Astra", and on MMMU-Pro it beats GPT-5.6 Terra at every reasoning setting and "nearly matches" Astra and Sol.
Terminal-Bench-Science: OpenAI says Terra comes within five points of Opus 5.5 at less than a fifth of the estimated cost. That's a loss with a coupon attached.
More than three times faster than Astra and Sol: AA's max-effort timing has Terra at 384s per task and Astra at 399s. That's 4% faster. OpenAI's 3x figure comes from its own task runs.

In the thread, OpenAI's tedsanders pointed out that eval error bars are "commonly like ±5%." Fair enough, but by that measure Terra's 3.2-point ALE deficit to Opus is inside the noise too.

Terra's effort levels: where cost per task flips the ranking

Here Terra earns its spot. If you sort by score, Opus wins. If you sort by score per dollar, Terra's lower effort levels win, and by a lot. Artificial Analysis breaks Terra down by effort:

Terra effort	AA Index	Cost per task	Time per task
none	33.5	$0.11	n/a
low	36.2	$0.05	29s
medium	41.3	$0.10	63s
high	43.9	$0.13	96s
xhigh	45.4	$0.19	160s
max	48.1	$0.40	384s

AA calls low and medium Pareto-optimal. Going from xhigh to max roughly doubles the cost for 2.7 points. Opus 5.5's cheapest setting, low effort, is listed at $0.55 per task and scores 42 on the index. Terra at medium gets 41.3 for $0.10. In the other direction, Opus at medium already scores 51, above Terra at max, but AA's Opus page doesn't publish a cost for that setting, so I can't tell you what it costs per task.

Third-party evals point the same way. chriddyp's Plotly data-analytics benchmark has Terra joining "the Pareto Frontier in between GPT-6 Luna and GPT-6.1 Sol," and of 33 models tested, only Haiku 5.5 and Nemotron 3.5 Super VL finished faster (both scored lower). thefourthchime's PacBench gave Terra 84 points for $0.19 in 5.4 minutes, against Grok 4.7's 83 for $1.69 in 36.5 minutes.

Terra at max is harder to defend. zozbot234 argued in the thread that Terra Max is beaten by Sol High on both price per task and intelligence. Sol at max already scores 51.8 for $0.72. So if you're about to run Terra at max, try Sol first, and if you want Terra, run it at medium.

What people using it report

jl runs a production app almost entirely on the OpenAI API and called Terra "an AWESOME release" for cost-sensitive agentic work. They still use Claude for coding.
agluszak said Terra, like Luna, "will stop at an earliest excuse." That's the lazy-agent problem again.
mapontosevenths ran one prompt through Codex (on Sol, not Terra) and Claude. Claude used 3.01x the tokens (649,147 vs 141,160) and wrote 2.25x more code. The same user said OpenAI had cut usage on its $200 plan by about 50% and called "cost per task" a "weaselly way" to hide that.
On the other side, hansonkd wrote: "I hate Opus 5.5, constant overreach and over-engineering for any project." If that's how Opus behaves for you, Terra's lower token count looks like a feature.

One number I'd watch: OpenAI's system card addendum has a test where the model, mid-task, finds a message board where other agents post. Terra took the specified unauthorized action in 13.75% of samples (11 of 80). Sol did it in 1.25%. If you're wiring Terra into a multi-agent setup with shell access, keep that in mind.

Should you switch to GPT-6.1 Terra?

My opinion: Terra is the new default for high-volume agent loops, classification, ticket triage and tool-calling pipelines, where you pay per call and 41 on the AA index at medium is plenty. For hard coding work, long refactors and anything where a wrong answer costs more than $5.98, it doesn't replace Opus 5.5.

If you want to try it without committing, pin the effort level explicitly. Medium is the API default today, but defaults change and max costs four times as much:

curl https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6.1-terra",
    "reasoning": { "effort": "medium" },
    "input": "Classify this support ticket: ..."
  }'

In Codex:

codex -m gpt-6.1-terra -c model_reasoning_effort="medium"

Before you switch, check these:

Run your own 20-50 real tasks at Terra medium, Terra max and Sol high. Compare cost per successful task, since that's the figure that shows up on your bill.
Log input sizes. If your prompts regularly go past 272K tokens, the whole request is billed at $2 / $7.50. Sol has the same surcharge, so moving up a tier won't dodge it.
Use Batch/Flex ($0.50 / $2.50) for anything that doesn't need an answer within seconds.
Keep Opus 5.5 (claude-opus-5-5) behind a fallback for the tasks Terra gives up on early.

So: are you switching your default model to Terra, or are a few points on OpenAI's chosen benchmarks too small to matter for your work? And if you've run Terra on a real codebase, did it finish the job or "stop at an earliest excuse"? Tell me in the comments.

DEV Community