GPT-6.1 Terra halves API prices and closes the gap on OpenAI's biggest models

OpenAI has released GPT-6.1 Terra, a mid-range model priced at half the API rates of its predecessor GPT-5.6 Terra — $1 per million input tokens and $5 per million output tokens, down from $2 and $12. Cached inputs drop to $0.05 per million tokens, a quarter of the previous rate, putting Terra at half the price of Anthropic's mid-tier Claude Sonnet 5.5, which charges $2 and $10 per million input and output tokens respectively.

The pricing move matters beyond one product refresh. Mid-tier models carry the bulk of enterprise AI workloads, and the gap between cheap-but-weak and expensive-but-capable models has been the main cost bottleneck for companies deploying AI agents at scale. A model that combines frontier-adjacent performance with half the token price changes the economics of automation — though, as always, OpenAI's own benchmark claims deserve scrutiny until third parties replicate them.

What does the new model actually do?

OpenAI says GPT-6.1 Terra combines fast visual reasoning with precise computer control and a more natural writing style, targeting everyday work tasks: drafting proposals, interpreting technical drawings, filling out forms. The model is available now for Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and through the API.

According to OpenAI, the model pushes the cost-to-performance boundary further than any of its predecessors. That is a vendor's claim, not a verified fact — but independent evaluator Artificial Analysis broadly backs it up, calling GPT-6.1 Terra a new benchmark for intelligence per dollar.

How big are the performance gains?

At maximum reasoning effort, Terra gains 6 points on Artificial Analysis's Intelligence Index over GPT-5.6 Terra while costing less than a third per task — $0.40 versus $1.40. At lower effort levels the gains run from 7 to 13 points, placing the model on the Pareto frontier, the point where cost and performance cannot simultaneously improve.

In the new Coding Agent Index v2.0, Terra scores 48 — six points below GPT-6.1 Sol at 54 and 17 points above GPT-6 Luna at 31. The pattern repeats across coding benchmarks:

• Terminal-Bench 4.0: Terra solves 50.3 percent of tasks at maximum reasoning effort, up from 21.5 percent for GPT-5.6 Terra. Artificial Analysis says that roughly 29-point jump puts Terra in the same tier as GPT-6.1 Sol.
• Terminal-Bench-Science 0.1: Terra outperforms GPT-6.1 Sol at lower cost and comes within five points of Claude Opus 5.5 at less than a fifth of the estimated cost.
• MMMU-Pro (visual reasoning): Terra outperforms its predecessor at every reasoning level and nearly matches GPT-6 Astra and GPT-6.1 Sol.
• OSWorld 2.1 (desktop agent benchmark): Terra beats GPT-5.6 Terra's best result at medium effort for a fraction of the cost.

OpenAI still recommends GPT-6.1 Sol and GPT-6 Astra for the most demanding coding tasks — a caveat worth noting, since it caps how far the mid-range model actually replaces the flagship tier.

The takeaway: Terra's coding and reasoning gains are large by any standard, but the company itself draws a line at the hardest workloads.

Why does the hallucination number stand out?

The most striking figure in the release concerns honesty under uncertainty. On the private AA-Omniscience benchmark from Artificial Analysis, GPT-6.1 Terra resorts to guessing only 38 percent of the time when it lacks the answer, down from 88 percent for GPT-5.6 Terra. Accuracy also rose by two points.

That halving-plus of guesswork addresses one of the most persistent complaints about deploying language models in business settings: fabricated answers delivered with confidence. A model that abstains when it does not know is materially safer for document-heavy office work than one that guesses. The benchmark is private, however, so independent verification of that 38 percent figure is not currently possible.

How fast is computer control?

OpenAI positions GPT-6.1 Terra as particularly strong at automated software control — the use case where agents operate applications and desktop environments directly. The numbers the company publishes are aggressive:

• On the Agents' Last Exam benchmark, Terra matches GPT-6 Astra's performance at one-tenth the token price.
• Compared to Claude Opus 5.5 at maximum effort, Terra is more than three times as token-efficient at a quarter of the token price — roughly nine times lower cost per task.
• Terra finishes more than three times faster than both GPT-6 Astra and GPT-6.1 Sol, according to OpenAI.

Speed and cost per completed task, not raw intelligence scores, determine whether AI agents are commercially viable for routine work. If OpenAI's figures hold up, Terra makes agentic automation roughly an order of magnitude cheaper against the strongest competing model.

What about safety?

GPT-6.1 Terra meets OpenAI's "Critical" cybersecurity threshold as defined by the company's Preparedness Framework, and inherits the safeguards of GPT-6 Astra. Compared to its predecessor, the model is more likely to stay within its assigned boundaries and disclose restrictions to users.

According to the system card, GPT-6.1 Terra holds up about as well as GPT-6 Astra and GPT-6.1 Sol against direct attacks on the instruction hierarchy (the rules governing which instructions a model must follow), with those models hitting 99.99 percent on such tests. For indirect prompt injection attacks — where malicious instructions are buried in third-party content that the model processes — Terra's defense rate sits at 98.8 percent, up from 96.2 percent for GPT-5.6 Terra.

A 1.6-point improvement in injection defense is real progress, but it also means roughly one in eighty injections still gets through — a number that matters for anyone pointing this model at untrusted web content.

What happens next?

GPT-6.1 Terra is available now across ChatGPT Work, Codex, and the API, and its pricing puts direct pressure on Anthropic's Sonnet line, which has anchored the mid-tier market for enterprise agent deployments. The open question is whether Anthropic responds with its own price cut or performance jump — and whether OpenAI's benchmark claims, especially on the private AA-Omniscience test, survive contact with independent evaluation.