The exam where it matches

GPT-6.1 Terra is OpenAI's cheaper tier for work that involves screens and software: reading visual material, operating applications, and writing in a natural register. It launched on October 26 through the OpenAI API, and AWS announced general availability on Amazon Bedrock the same day, with access through Bedrock's APIs or through ChatGPT Work and Codex configured to run on Bedrock.

The result OpenAI leads with is Agents' Last Exam v1, a test of professional computer tasks that OpenAI describes as turning architectural drawings into a 3D model and reviewing an animated film shot by shot. Terra scored 57.6% at medium reasoning effort, the same score OpenAI reports for GPT-6 Astra at medium effort. The estimated cost per task was $0.29 for Terra against $4.82 for Astra. Against its predecessor, GPT-5.6 Terra's 42.6% at medium effort, the gain is 15 points. OpenAI also says Terra completed these tasks more than three times faster on average than Astra and Sol.

On OSWorld, Terra also surpasses GPT-5.6 Terra's best score at medium effort, OpenAI says, and on MMMU-Pro, a visual reasoning benchmark, Terra's 85.5% at maximum effort sits just below Sol's 86.3% and Astra's 86.7%.

That claim needs its qualifications kept attached. The tie is exact only on this one exam, at matched effort, in OpenAI's own evaluation. That is a vendor-run comparison, with no independent reproduction on launch day. The speed figure describes task completion on this exam rather than a general tokens-per-second rate, and Artificial Analysis had not published output-speed or latency measurements for Terra at the launch-day check. A tie on a workload test is a real result. It is not a ranking of the two models.

The charts where it does not

On OpenAI's coding test, the order reverses. Terminal-Bench 4.0 measures coding agents solving tasks from a terminal. Terra reached 50.3% at maximum effort. Sol at maximum effort reached 55.9%, and Sol at medium effort, 52.7% — a different effort setting — already sat above Terra's maximum. Claude Sonnet 5.5's reported maximum was 70.6%. OpenAI's estimated cost for Sol at medium effort on that test was $1.02 per task, against $1.84 for Terra at maximum effort. For demanding coding, running the stronger model at a moderate setting beat running the cheaper model at full effort on both score and estimated task cost in OpenAI's own chart.

Desktop work shows the same pattern with a pricing twist. On OSWorld 2.1, Terra's maximum score of 74.2% is a clear step up from GPT-5.6 Terra, and still below Sol at 78.0% and Astra at 80.9%. At maximum effort, OpenAI's estimated cost per task was $1.19 for Terra and $1.15 for Sol. In OpenAI's reported OSWorld estimates, Terra's lower token rate did not translate into a lower cost per task.

The independent snapshot points the same way. Artificial Analysis scored Terra at maximum effort 48 on its Intelligence Index v4.3.2 on launch day, below Astra at 53 and Sol at maximum effort at 52. None of this makes Terra a weak model, and it does not overturn the exam tie: the other charts measure different capabilities. What they show is that the tie does not generalise. One exam was chosen to stand for the model, and the model is more uneven than that exam suggests.

Half price, with a threshold

Terra's standard rates are exactly half of Sol's on input and output, and lower than GPT-5.6 Terra's on both, which makes AWS's "at least 50% lower" accurate against the two models it named. The condition is the 272,000-token threshold. Past it, the higher rates apply to the full request, not just to the tokens beyond it: input costs $2 per million tokens, cached input $0.10, and output $7.50. The 1,050,000-token window is real capacity, but a prompt that uses most of it is not billed at the headline rate.

The Bedrock side of the launch has a gap in it. AWS says the model is generally available and names the access routes, but its announcement states no Bedrock prices and no list of regions, pointing readers to the Bedrock documentation instead. Nor should OpenAI's API prices be assumed to apply to Bedrock; AWS does not restate them. A buyer comparing vendors today can price the OpenAI API exactly and cannot price Bedrock from the launch material at all.

The practical trial is therefore specific. Run Terra on the visual and office work it is pitched for, at the effort setting you would actually use, and compare cost per accepted result rather than cost per token. One API detail matters for agent builders: tool calling requires the Responses API, per OpenAI's model reference. Include at least one long-context case, because the full-request surcharge above 272,000 input tokens is where the half-price pitch quietly stops applying.

Sources

  1. [1] OpenAI, "Introducing GPT-6.1 Terra" (launch announcement, October 26, 2026)Read source
  2. [2] OpenAI API, "GPT-6.1 Terra Model" (model reference and pricing, read October 26, 2026)Read source
  3. [3] AWS, "OpenAI GPT-6.1 Terra is now generally available on Amazon Bedrock" (announcement, October 26, 2026; read via full-text republication)Read source
  4. [4] Artificial Analysis, "GPT-6.1 Terra pushes the cost efficiency frontier" and the GPT-6.1 Terra model comparison page (October 26, 2026)Read source
  5. [5] Kingy AI, "GPT-6.1 Terra: Benchmarks, Pricing & Model Comparisons" (transcription of OpenAI's launch charts, including the OSWorld and Terminal-Bench cost-per-task estimates; October 26, 2026)Read source