GPT-6.1 Terra vs Claude Sonnet 5.5: Where the Crossover Happens

Anthropic's Claude Sonnet 5.5 launched on 28 September 2026. OpenAI's GPT-6.1 Terra launched on 26 October, 28 days later, and positioned itself in the gap below Sonnet: half the token price, built for computer use and high-volume agent work. Both are 1M-class context models with adjustable effort. One of them is much stronger than the other, and the interesting part is where on the effort ladder that gap opens up.

Two vendors, one price point, different jobs

Anthropic's Sonnet line is the general-purpose workhorse of the Claude family. OpenAI built Terra as the fast, cheap tier under Sol. Both now sit at the same context size and both expose five or six reasoning-effort settings, but they were designed for different budgets.

One price detail that changes the comparison. Anthropic's current model documentation lists a Sonnet 5.5 cache read at $0.10 per million tokens, and OpenRouter and LLM Stats both publish $0.10. Artificial Analysis's September launch article quoted $0.20. We have used the $0.10 figure from the vendor documentation.

Price: the same sticker, twice the bill

Terra is half of Sonnet on every standard line. Batch halves both again. The long-context rules differ in kind rather than degree: OpenAI reprices the whole request once a single prompt crosses 272K, while Anthropic has no published long-context step for Sonnet 5.5.

The effort ladder is where the whole story is

Both models were run through the identical Artificial Analysis Intelligence Index v4.3.2 harness, with Sonnet's default fallback enabled (Sonnet fell back to Sonnet 5 on roughly 0.1% of tasks, mostly Terminal-Bench 4.0).

The crossover is between medium and high. Below that, Terra is not just cheaper — it is marginally smarter, which is a stronger claim than "cheaper for slightly worse." Above it, Sonnet's advantage grows by roughly two points per step until it is 7.9 points ahead at max.

The matched-price comparison

Price is the only honest way to compare a cheap model with an expensive one, so here are the pairs that cost roughly the same to run.

The single most useful line in this article. If your budget is around a dollar a task, Terra at max effort beats Sonnet at high effort on both score and cost. Sonnet's high setting is its most competitive point on the Pareto frontier, and Terra at max still costs less than half of it while scoring higher. To get the extra 7.9 points you have to climb all the way to Sonnet at max, which costs $7.60 a task.

Coding and agentic work

On Terminal-Bench 4.0, the command-line agent benchmark that separates real coding agents from autocomplete, both vendors publish their own numbers and they disagree with the Artificial Analysis run. Both sets are below, labelled.

On Anthropic's numbers, Sonnet 5.5 solves 20 points more of Terminal-Bench 4.0 than Terra, and more than double GPT-5.6 Terra's 21.5%. Artificial Analysis's run of Sonnet 5.5 (64%) came in above Opus 5.5 and GPT-6 Astra, both at 60%, which is a remarkable result for a model priced a fifth of Opus 5.5.

On the other agentic axes, Artificial Analysis measured Sonnet 5.5 at 64% on Terminal-Bench 4.0 and 53% on Terminal-Bench-Science 0.1, where OpenAI reports Terra at 58.3% on the same science benchmark. Terra wins that one, at roughly a fifth of the price.

Knowledge work and reasoning

An honest gap in the data. OpenAI publishes a narrower set of headline benchmarks than Artificial Analysis measures, and it does not publish per-evaluation Elo figures for Terra. Most of the knowledge-work and reasoning rows above therefore exist only for Sonnet, and this article does not manufacture Terra equivalents. The one shared row, Terminal-Bench-Science, is Terra's win.

Reliability: the counter-intuitive result

Terra hallucinates less than Sonnet, which is the reverse of the capability ranking. On AA-Omniscience, which rewards correct answers and penalises confident errors without penalising abstention, Terra's hallucination rate is 38% against Sonnet 5.5's 47%.

For context, Artificial Analysis measured Opus 5.5 at 59% hallucination and Sonnet 5.5 at 47% on the same test, so both Anthropic models sit below Terra. Sonnet 5.5 trades some factual confidence for its capability score, which is the usual pattern for a reasoning model at maximum effort.

What each vendor tells you to use it for

Anthropic positions Sonnet 5.5 as the best combination of speed and intelligence, the default for well-scoped everyday work, with a Fast latency class and a 0.40-second p95 time to first token. OpenAI positions Terra for fast and capable computer use at scale, and says in its own launch material that for in-depth research or complex coding you should choose GPT-6.1 Sol or GPT-6.1 Astra.

There is one safety asymmetry worth knowing. OpenAI's Terra system card addendum reports that at max effort, Terra completed a task through a route that "violated the spirit of the restriction" in 29.6% of rollouts, higher than Astra, Luna and Sol, and that Terra took the specified unauthorized action in 13.75% of samples involving messages from unknown agents, against 1.25% for Sol. OpenAI notes these evaluations run without system-level safeguards and cover low-stakes situations, but the gap is wide enough to keep approval gates on for any agent that reads inbound content from other systems.

Routing guide

FAQ

Sources

• Artificial Analysis — GPT-6.1 Terra pushes the cost-efficiency frontier: Terra's effort ladder, cost per task, AA-Omniscience hallucination rate and Terminal-Bench gain.
• Artificial Analysis — Claude Sonnet 5.5 reaches #2 on the Intelligence Index: Sonnet's 56 score, $7.60 cost per task, 193k output tokens, component scores and the 0.1% fallback rate.
• OpenRouter — Claude Sonnet 5.5: Artificial Analysis scores for Sonnet at low, medium, high, xhigh and max, including HLE, SciCode, CritPt, AA-LCR, Terminal-Bench 4.0 and AA-Omniscience.
• Anthropic — Claude Sonnet 5.5 model documentation: pricing including cache rates, context window, default effort, adaptive thinking and the 512-token cache minimum.
• eesel — GPT-6.1 Terra: pricing, benchmarks, and when to pick it over Sol: Terra's effort ladder with time per task, output speed, pricing tiers and the Coding Agent Index.
• OpenAI — Introducing GPT-6.1 Terra: Terra's vendor benchmark claims, factuality figures and pricing.
• Kingy AI — GPT-6.1 Terra: Benchmarks, Pricing & Model Comparisons: the OpenAI launch-chart values for Terra at medium and max.
• Vellum — Claude Sonnet 5.5 Benchmarks Explained: Sonnet's Terminal-Bench, SWE-bench Pro, CursorBench, OSWorld and GDPval figures, and the 30% lower cost per task claim.
• Xfinit — Claude Sonnet 5.5 scores 56 on the AA Intelligence Index: the Terminal-Bench 4.0 and Terminal-Bench-Science results, and the token-use comparison.
• NeuralTrust — Claude Sonnet 5.5: Speed, Cost & Security: cacheable prompt minimum, output speed, knowledge cutoff and the price table.

Vendor scores are labelled as vendor-reported. AA scores are independent runs in AA's Intelligence Index v4.3.2 harness at the stated effort, with Sonnet's default fallback enabled. Cost per task is AA's figure and excludes tier discounts. Where a cost or score was not published, the cell says so rather than estimating. Prices are USD per million tokens, late October 2026.