A cheaper tier that is also a better one
Model price cuts are routine; a price cut paired with a real accuracy jump is not. GPT-6.1 Terra lands at half the per-token cost of OpenAI's own Sol and 5.6 Terra tiers, which would normally read as a pure cost play aimed at developers who had been priced out of agentic workloads. What makes this release different is that OpenAI's own stress-test numbers show the cheaper model is also dramatically more reliable, with its internal error rate dropping from 53.6 percent to 16.7 percent at low reasoning effort and down to 4.3 percent when the model is allowed to think harder.
That combination matters because computer-use and agentic workflows have been stuck in pilot purgatory at a lot of enterprises precisely because the economics and the reliability both looked bad at once. A team could get an agent that worked most of the time at a price that made production deployment hard to justify, or they could throttle usage to control cost and watch reliability suffer further. Terra attacks both constraints in the same release, which is a more useful signal than the headline price cut alone.
What the benchmark number actually means
Terminal-Bench 4.0 measures a model's ability to complete real command-line and file-system tasks end to end, which is a reasonable proxy for the kind of agentic work enterprises actually want: running scripts, managing environments, chaining tool calls without a human stepping in every few seconds. Terra solved 50.3 percent of those tasks at maximum reasoning effort, more than doubling the 21.5 percent its predecessor managed. OpenAI also claims it is more than three times faster than GPT-6 Astra, which compounds the reliability gain with a latency gain that matters for anything running in a loop.
We would caveat that internal benchmarks from the model's own maker always deserve a skeptical read, and Artificial Analysis's independent framing, that Terra is the first Terra-class model since GPT-5.6 and more intelligent for less than a third of the cost, is closer to the number worth trusting. Either way, the direction is consistent across both OpenAI's and the third party's numbers, which is the part that should change a procurement conversation.
Where this actually ships first
Terra rolled out immediately to ChatGPT Work, to Codex, and to the API under the model id gpt-6.1-terra, with a context window of 1.05 million tokens. That distribution choice is deliberate. ChatGPT Work and Codex are the two surfaces where enterprise technical staff actually run agentic workflows today, rather than the consumer chat interface, so OpenAI is putting the cheaper, more reliable model directly in front of the buyers who make the renewal decision.
The context window increase is easy to overlook next to the price and accuracy numbers, but a 1.05 million token window changes what an agent can hold in working memory across a long computer-use session: entire codebases, long log files, multi-step browser sessions. We think that window size does more for sustained agentic reliability than any single prompt-engineering trick a customer could apply on top of a smaller-context model, because it lets the model keep a coherent record of every action it has already taken rather than relying on a summarized history that can quietly drop the one detail that mattered for step forty of a fifty-step task.
The competitive pressure this creates
Anthropic and Google both have mid-tier models positioned for cost-sensitive agentic workloads, and both now face a direct price and performance comparison they did not have to answer a month ago. A halved price with a tripled reliability rate on the same workload class is the kind of release that shows up in competitive bake-offs almost immediately, because procurement teams running multi-vendor evaluations will simply rerun their test suites against the new number.
We expect at least one of OpenAI's two main competitors to respond with a comparable mid-tier release within a quarter, not because Terra is unbeatable on raw capability but because the price-per-reliable-task ratio it sets becomes the new floor that any agentic product has to clear to stay in consideration. Enterprises running active evaluations should rerun them now rather than waiting for the response, since Terra's current numbers are the ones available to negotiate against today.
What CIOs should do with this release
If your organization shelved a computer-use or agentic pilot because the unit economics did not clear a defensible ROI threshold, Terra is the release that reopens that case. The combination of a halved price and a meaningfully lower error rate changes the breakeven math on workloads like automated QA, back-office document processing, and IT operations tasks that were previously too expensive to run at the volume needed to prove value.
We would not treat the benchmark numbers as a reason to skip your own evaluation, since Terminal-Bench and internal stress tests do not map one to one onto any specific enterprise workflow. But they are a strong enough signal to justify re-running a shelved pilot now, with a tighter cost model and a shorter time-to-decision than the first attempt used, and with a specific instruction to the team to measure the new error rate against production data rather than against the vendor's own benchmark suite before any budget gets reallocated to a wider rollout.
