$1 per Token, 3x Faster—OpenAI Puts Its "Cheap Sidekick" GPT-6.1 Terra into Codex, While Advising You Send "the Hardest Code" to Another Model
OpenAI released GPT-6.1 Terra. At $1 input / $5 output—about half the previous generation—it runs 3x faster, and its CLI coding accuracy jumped from 21.5% to 50.3%. It's usable in Codex and Copilot, though OpenAI still recommends higher-tier models for the hardest code.
On October 26, 2026, OpenAI released a new model called "GPT-6.1 Terra." It's positioned not as the top-tier brain, but as an "everyday" tier that handles routine tasks cheaply and quickly. OpenAI halved the API pricing compared to the previous generation while more than doubling the accuracy of coding that runs in the terminal, and made it instantly usable from ChatGPT's Codex. For developers, this changes the "per-step cost and wait time" when running agents.
A half-price, 3x-faster "everyday" tier
Standard API pricing is $1 per million input tokens and $5 per million output tokens. OpenAI describes this as "about half the price of GPT-6.1 Sol and GPT-5.6 Terra." Cached input is even cheaper, at $0.05 per million tokens.
| Type (standard API, per million tokens) | GPT-6.1 Terra |
|---|---|
| Input | $1 |
| Input (cached) | $0.05 |
| Output | $5 |
Compiled from source pricing information (BleepingComputer and others). Roughly half the price of GPT-6.1 Sol / GPT-5.6 Terra.
Speed is another selling point. According to OpenAI, task completion is more than 3x faster than GPT-6 Astra and GPT-6.1 Sol. On cost efficiency, it's positioned as delivering "performance on par with GPT-6 Astra at roughly one-tenth the per-token price." On Agents' Last Exam, which measures reasoning ability, it reportedly beat GPT-5.6 Terra by 15 points at medium reasoning effort while costing about one-third as much.
CLI coding accuracy more than doubled
What developers will want to watch is the jump in how well the agent works at the command line. On Terminal-Bench 4.0, which measures tasks involving terminal operations, GPT-6.1 Terra solved 50.3% at maximum reasoning effort. The previous-generation GPT-5.6 Terra scored 21.5%, so within the same "Terra" line this is more than a twofold improvement.
| Terminal-Bench 4.0 (maximum reasoning effort) | Accuracy |
|---|---|
| GPT-6.1 Terra | 50.3% |
| GPT-5.6 Terra | 21.5% |
Sources: DataCamp / MarkTechPost. A difficulty test for coding agents that run at the CLI.
A drop in error rates was also reported. At low reasoning effort it fell from 53.6% to 16.7%, and at maximum effort down to 4.3%—roughly matching the level of the higher-tier GPT-6 Astra and GPT-6.1 Sol. The design idea is that even a cheap model can curb "misses" if you turn up the effort.
Available from ChatGPT, Codex, Copilot, and Bedrock
Access was broad from day one. Here's where you can use it:
- ChatGPT: Plus / Pro / Enterprise / Edu users can use it from ChatGPT Work and Codex.
- API: The model ID is
gpt-6.1-terra. - GitHub Copilot: Generally available on Pro / Pro+ / Max / Business / Enterprise (phased rollout).
- Amazon Bedrock: Now generally available.
The fact that it landed simultaneously in Codex and Copilot—the two major coding environments—is significant in practice: you can try it by swapping in just the model without changing your existing workflow.
The cost of letting an agent run keeps changing
The price cut and speed boost pay off less in a single response than in usage where you "let an agent run for a long time." A coding agent consumes huge numbers of tokens by repeatedly reading files, running commands, looking at the results, and running more. When each step gets cheaper and faster, you can run more attempts on the same budget, and automated loops like CI and review integration become more likely to fit within realistic costs. The cheap input cache also pairs well with work that repeatedly reads the same repository.
But "the hardest code" isn't entrusted to it
At the same time, it's premature to see Terra as an all-purpose replacement. OpenAI itself still recommends GPT-6.1 Sol or GPT-6 Astra for the most complex coding. Even with benchmark numbers doubling, the lead on the hardest tasks remains with the higher-tier models.
There are operational caveats to note as well. "Cheap per token" doesn't necessarily mean "cheap per task." If you throw a cheap model at a hard problem and then crank up the reasoning effort or rack up more attempts, it can end up more expensive overall. Now that the fast-and-cheap tier has grown, the design choice of "which model to assign to which task" looks set to matter more than ever. A realistic starting point: lean on Terra for everyday fixes, tests, and routine automation, and route intricate design and tricky spots to the higher-tier models.
References: BleepingComputer, "OpenAI launches GPT-6.1 Terra" / MarkTechPost, "OpenAI Launches GPT-6.1 Terra" / DataCamp, "GPT-6.1 Terra: Benchmarks, Pricing and Access" / Nerdschalk, "GPT-6.1 Terra ChatGPT & Codex availability" / Developers Digest, "GPT-6.1 Terra release guide"
