OpenAI’s API changelog lists GPT-6.1 Terra as a model released on October 26, 2026, for visual reasoning and computer use. Its model card gives it a 1.05-million-token context window and a maximum output of 128,000 tokens. Standard API rates are $1 per million input tokens and $5 per million output tokens, with cached input at $0.05 per million. The key budgeting detail is a higher-cost band for very long prompts.
OpenAI’s pricing page defines short context as 272,000 input tokens or fewer and long context as anything above that. Once a prompt crosses the line, the input and cached-input rates double and the output rate rises by 1.5 times, and the higher rates apply to the whole request rather than only the tokens past the threshold. The worked examples below show what that does to a single call.
Two requests 10,000 tokens apart
Take two requests that each return 8,000 output tokens. Request A sends 270,000 input tokens, just under the line. Request B sends 280,000, just over it. On Standard rates, request A costs 0.27 million input tokens at $1 ($0.27) plus 0.008 million output tokens at $5 ($0.04), or $0.31. Request B costs 0.28 million input tokens at the long-context $2 rate ($0.56) plus 0.008 million output tokens at $7.50 ($0.06), or $0.62. Adding 10,000 input tokens, about 3.7 percent more prompt, doubles the bill for the call. Across 1,000 such calls the gap is $310 against $620.
The jump is all-or-nothing, so trimming retrieved documents or older conversation turns to stay at or below 272,000 input tokens saves far more than the trimmed tokens themselves cost. The 128,000-token output cap is a separate ceiling; it does not offset input usage or turn the 1.05-million-token context into a combined allowance.
The same two requests in each service mode
OpenAI prices the same model differently by service mode. The pricing page lists Batch and Flex at half the Standard rates in both bands, Fast at twice Standard, and a 10 percent uplift for regional processing (data residency) endpoints on models released on or after March 5, 2026. The pricing page lists no Ultrafast rate for GPT-6.1 Terra. The table applies each mode’s published short- and long-context rates to requests A and B.
| Service mode | Rates used (input / output per million, short then long) | Request A: 270,000 input | Request B: 280,000 input |
|---|---|---|---|
| Standard | $1 / $5; $2 / $7.50 | $0.27 + $0.04 = $0.31 | $0.56 + $0.06 = $0.62 |
| Batch | $0.50 / $2.50; $1 / $3.75 | $0.135 + $0.02 = $0.155 | $0.28 + $0.03 = $0.31 |
| Flex | $0.50 / $2.50; $1 / $3.75 | $0.135 + $0.02 = $0.155 | $0.28 + $0.03 = $0.31 |
| Fast | $2 / $10; $4 / $15 | $0.54 + $0.08 = $0.62 | $1.12 + $0.12 = $1.24 |
| Regional, on Standard | Standard plus 10 percent | $0.31 x 1.1 = $0.341 | $0.62 x 1.1 = $0.682 |
Two results stand out. A long-context request sent through Batch or Flex, at $0.31, costs the same as a short-context request on Standard. A short request on Fast, at $0.62, costs the same as a long request on Standard. For work that can wait, the mode choice moves the bill as much as the threshold does.
Cached input softens the jump but does not remove it
Agents that resend the same long prefix can lean on cached input. Suppose 250,000 of each prompt’s input tokens are cache hits and the cache is already written. Request A then costs 0.25 million cached tokens at $0.05 ($0.0125), 0.02 million uncached tokens at $1 ($0.02) and the $0.04 of output, or $0.0725. Request B costs 0.25 million cached tokens at the long-context $0.10 ($0.025), 0.03 million uncached tokens at $2 ($0.06) and $0.06 of output, or $0.145. The call is far cheaper than the uncached version, but crossing 272,000 tokens still doubles it. The first request that writes the cache pays the cache-write rate instead, which the model card puts at $1.25 per million for 30-minute retention before the long-context adjustment.
Computer-use calls need the Responses API
The model card describes GPT-6.1 Terra as supporting computer use and lists the API identifier gpt-6.1-terra for both the Responses and Chat Completions endpoints. OpenAI’s GPT-6 guide says tool calling with Terra requires Responses, so applications that call tools through Chat Completions must move those requests. Changing only the model identifier is not enough.
Token charges are not the whole agent bill. The model card notes a per-call fee for tool-specific models such as search and computer use, and the pricing page lists web search at $10 per 1,000 calls plus the retrieved content billed as input tokens. Search results that land in the prompt also count toward the 272,000-token line, so a research agent can cross into the long-context band through tool output alone.
For readers comparing API inference with a separate local model, NeonControl’s local-AI hardware-fit calculator estimates memory needs, but not throughput or total cost. Plan-based access is a separate question again; NeonControl’s AI-provider comparison tracks plan terms apart from metered API rates.
