Auditing GPT-6.1 Terra: A Complete Release Sheet and a 9× Cost-per-Task Claim

Key takeaways

• GPT-6.1 Terra launched Oct 26 with its full disclosure printed at once: standard API pricing of $1 per million input tokens, $0.05 cached, $5 output, benchmark results, and a safety evaluation on the same launch page (OpenAI's launch post).
• The $1/$0.05/$5 price band is 50% below the standard API prices of GPT-6.1 Sol and GPT-5.6 Terra, and the cached-input rate is a published 95% discount — Kimi K3.1 Pro and MiniMax M3.1 Flash print 90% cache discounts at their sizes (launch post; Kimi pricing; M3.1 Flash pricing).
• At maximum reasoning effort, OpenAI measures Terra at over 3× the token efficiency of Claude Opus 5.5 at one-fourth the per-token price — a claimed ~9× lower cost per task on the vendor's own harness (launch post).
• Terminal-Bench 4.0 reads 50.3% at max effort, more than double GPT-5.6 Terra's published 21.5%, while the sheet itself routes the hardest coding to GPT-6.1 Sol and GPT-6 Astra (launch post).
• The factuality row is the largest single-column move: a 53.6% error rate on error-inducing prompts at low effort fell to 16.7%, and 4.3% at maximum effort (launch post).

Two launches this week ran the same buyer test in opposite directions. On Oct 23, Moonshot published Kimi K3.1 Pro's release sheet — benchmark scores and a line-by-line checkable pricing page — but routed the artifacts that settle self-hosting to a vendor-confirmed Nov 4 weights date, with the license still unnamed at launch (the sheet-first audit). A day later, OpenAI's GPT-6.1 Terra put everything on the launch page before a token was billed: standard API pricing of $1/$0.05-cached/$5 per MTok, results across five agentic benchmarks, and a safety evaluation on the same page. The genre's expectation inverted: the closed model is the one whose sheet ships complete, and it is the open-weight launches that now make buyers wait.

This article reads Terra's sheet the way a procurement review should. It itemizes the pricing and cache rows a cost model can use today, separates the benchmark rows OpenAI measures itself from what an independent column would settle, reads the safety disclosure that now ships as part of the sheet, and names the two things even a complete sheet cannot print — a third-party calibration column, and weights that will never arrive. It builds on the Kimi K3.1 Pro release-sheet audit, which covered the five-check weights-day method for a launch that arrives sheet-first; this is the opposite corner of the same problem — what to buy on when there is no weights day at all.

The sheet, line by line

The pricing page is short and complete. Standard API pricing is $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens, with the launch post stating these standard prices are 50% lower than GPT-5.6 Terra's and GPT-6.1 Sol's (launch post). Three rows, no tiering asterisks: a cost model can be built on the page the day it appears. That is more than a convenience. When an aggregator's pre-launch Terra card — updated Oct 24, before the model existed — carried the model at a different price point with a "ChatGPT default" status the launch page contradicts, the garbled copy circulated for less than a day and died on contact with the page. The primary sheet is the audit artifact, and Terra's was complete enough to correct the copies immediately.

The cache row is the deepest published discount in this set. Terra's cached input is 5% of its list input rate — $0.05 against $1. The 90%-discount class is what the flash-tier and flagship launches printed this month: Kimi K3.1 Pro's $0.20 cached rate against a $2.00 input price (the sheet-first audit) and MiniMax M3.1 Flash's $0.015 against $0.15 (the launch-day read). A 95% discount does not mean cheaper work by itself — DeepSeek's 1017 snapshot showed cache-hit share, not the list, sets effective cost — but for a routing model whose inputs repeat (the same system prompt, the same tool schemas, the same screen context across form fields), the rate at which reuse is billed is the row that decides where the workload lands.

Availability is stated, not promised. Terra is available to all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex starting Oct 26, and to developers through the OpenAI API as gpt-6.1-terra — with the launch post noting the same usage allowance covers more work than the allowance did under GPT-5.6 Terra, and substantially more than under Sol or GPT-6 Astra (launch post). The rollout status also closes a genre question: there is no "weights arriving" line for this model, ever. A closed vendor's completeness costs nothing at weights time because there is no weights time.

The benchmark rows, and what each is measured on

The computer-use headline row is measured on Agents' Last Exam v1 — tasks the sheet describes as turning architectural drawings into 3D models and reviewing film shot by shot into time-coded logs. At medium reasoning effort Terra scores 15 percentage points above GPT-5.6 Terra at about one-third the cost; it performs on par with GPT-6 Astra at about one-tenth Astra's per-token price; and at maximum effort OpenAI measures over 3× the token efficiency of Claude Opus 5.5 at one-fourth the per-token price, which the sheet rounds to a ~9× lower cost per task (launch post). Every number in that sentence is a vendor measurement on OpenAI's own harness and effort settings. The claim is checkable in structure — the price is published, so the per-task figure reduces to a token-count claim — but the token counts are not independently reproduced anywhere yet.

The coding row carries its own boundary. On Terminal-Bench 4.0, Terra solves 50.3% of tasks at maximum reasoning effort, more than double GPT-5.6 Terra's published 21.5% — and the same paragraph states that for the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice (launch post). The sheet prints the workload where its model loses. That is the same disclosure shape the genre recorded on Kimi's sheet, where the launch led with the bench it lost by 0.7 points, and it is the honest label a routing table needs: Terra is the workday computer-use tier, not a frontier-coding tier. On Terminal-Bench-Science 0.1, Terra surpasses Sol at lower cost and comes within five points of Opus 5.5 at less than one-fifth the estimated cost (launch post).

The factuality row is the largest single-column move on the sheet. On OpenAI's internal benchmark of error-inducing prompts, the error rate fell from 53.6% for GPT-5.6 Terra to 16.7% at low effort, and to 4.3% at maximum effort — a drop the sheet describes as nearly matching the rates for GPT-6 Astra and Sol (launch post). For an always-on fleet the operational read is specific: a model that can be run at low effort with a 16.7% error rate on adversarial prompts is priced and profiled for volume work, while the 4.3%-at-max column is the one to cite in a verification-heavy workflow.

What no sheet prints about itself is the third-party column. Artificial Analysis' snapshot used in the Kimi audit's cross-check predates Terra (updated Oct 24), and no independent index row for gpt-6.1-terra exists yet in the surfaces this audit uses — neither an Intelligence Index composite nor an AA-class calibration column. Until one lands, the ~9× cost-per-task figure and the ALE score stand only on OpenAI's own measurements. That is not a defect of this sheet; it is the normal state of a launch, and it is precisely the gap the always-on economics frame tells buyers to fill with their own token logs after the first bill.

The safety disclosure now ships as part of the sheet

The Preparedness row is on the launch page, not in a footnote. The sheet states that GPT-6.1 Terra meets the Critical threshold in cybersecurity under OpenAI's Preparedness Framework, maintains the protections developed for GPT-6 Astra, and carries the GPT-6 series' alignment improvements — with the launch post reporting Terra is more likely than GPT-5.6 Terra to stay in scope and honestly disclose limitations, with strong resistance to prompt injection in OpenAI's evaluations, and the full detail in the system card addendum (launch post).

Two reads follow. First, "Critical in cybersecurity" is a capability disclosure, not a deployment blocker — it tells a buyer the model is capable enough in cyber tasks to warrant the framework's highest labeling tier, which is exactly the information a risk review needs before wiring the model into tool-attached workflows. Second, the disclosure being part of the launch-day sheet means a buyer's governance review has something dated to attach to its file: the system card addendum, published with the model. The release-sheet-audit genre exists because October's launches varied in which artifacts landed first; Terra's variation is that nothing was deferred — and simultaneously, because the model is closed, nothing can be verified beyond the vendor's artifacts at all.

The contrast case: what the open-weight launches still route to a later date

October's three audits now read as three disclosure patterns. MiniMax M3.1 Flash shipped weights four days after its API launch and let the license file carry the buying conditions — a $20 million revenue line with attribution (the weights-day audit). Kimi K3.1 Pro launched sheet-first and complete on capability and price, with the weights and the license unnamed until Nov 4. GPT-6.1 Terra published everything at once — and offers no weights, ever. The interesting consequence is asymmetric: open-weight launches expose buyers to a license-reading event that can rewrite self-hosting economics after purchase (the $20 million line), while the closed launch exposes buyers to a single-vendor dependency that can reprice at renewal. Neither is avoidable; both are auditable, and a sheet that arrives complete on day one — with prices, scores, and safety rows checkable against the page — is the disclosure that makes the comparison honest. This is the argument's cost-side sibling to the staged-release precedent: what varies across launches is not the capability claim but which artifacts land first, and the buyer's diligence has to follow the artifacts.

One more market texture worth filing: the aggregator catalogs around a launch are not sheets. A pre-launch Terra card that carried the wrong price and the wrong status survived less than 48 hours against the primary, but long enough to end up in a routing memo. A buyer who quotes a model's numbers from a tracker gets the tracker's error bars; the launch page is the only surface that carries the launch's own commitments.

One-minute map of the sheet and its gaps:

The two panels are the split the buyer's diligence draws: the teal side is what the launch made checkable on Oct 26, the violet side is what only time or third parties can settle, and the coral strip is October's full disclosure-pattern ledger. The bottom card is the rule this genre has landed on: buy on the sheet, audit on the artifacts.

What changes in routing this week

The cheap tier's ledger gains its OpenAI-native row. The subagent-and-computer-use cost-collapse arc now runs: Haiku 5.5 → GPT-6.1 Sol/Luna → GLM-5.3 Flash → DeepSeek's 1017 snapshot → MiniMax M3.1 Flash → Terra (the tier's ledger article). What distinguishes Terra in that ledger is not a lower number — Kimi's and MiniMax's flash tiers list lower absolute prices — but the class of work it claims: precise computer use against desktop applications, the OSWorld 2.1 workload, at a $1/$5 band with a published 95% cache discount. A closed vendor joining the cost-collapse tier collapses the last excuse for treating cheap tiers as open-model-only: the tier is now priced from both directions.

Cost per completed task, not per token, is still the decision frame (the per-task frame). Terra's price band says nothing until multiplied by task runtime and hit share. The sheet gives the inputs to compute it — token price, cached rate, and two benchmark scores that proxy tokens-per-task at two effort settings. Effort is the bill's other axis: the 16.7% factuality row is the volume column at low effort, while the 4.3% rate lives behind maximum effort, and volume work rarely needs it.

The workload boundary the sheet prints is the routing boundary. Terra is the workday tier — proposals, technical drawings, forms, video-editing logs (launch post) — with harder coding explicitly routed to Sol and Astra by OpenAI's own text. The cheapest error a buyer can make with a complete sheet is to read its winning rows as global: 50.3% on Terminal-Bench 4.0 does not compete with the frontier's hardest-coding tier, and the sheet says so.

Related reading

• Kimi K3.1 Pro and the 0.7-Point Gap: Reading the Release Sheet Before the Nov 4 Weights — the parent audit in this genre: the sheet-first launch this article contrasts, and the five-check method to run on its Nov 4 weights
• The $20 Million Line in the License: Weights Day Audits MiniMax M3.1 Flash — the weights-day audit that opened the genre: the license event that can rewrite self-hosting economics after purchase
• Inference Economics: Why Always-On Production Agents Are Now Affordable — the per-task cost frame that turns published price bands and cache rows into a routing decision

A representative build

A 300-person industrial-equipment distributor's operations desk runs about 2,500 computer-use agent tasks a week — invoice detail entry into NetSuite, pricing lookups across four supplier portals, and technical-drawing interpretation for equipment quotes. Terra's sheet moves two routings this week: the high-volume visual form class (the invoice-entry and portal-fill work OSWorld 2.1 proxies) moves to Terra at medium reasoning effort, where the +15-point Agents' Last Exam row pairs with the $1/$5 band and a $0.05 cached-input rate that pays out against their repeated-prefix traffic; terminal-and-code maintenance jobs stay on Sol and Astra, per the sheet's own boundary. The document-audit fleet stays on Kimi K3.1 Pro pending the Nov 4 checks, and the routing table carries a column the sheet cannot answer — independently measured cost per task — which the team fills from its own token logs after the first billing cycle, with Terra's max-effort economics re-priced against the 16.7%/4.3% factuality columns before any verification-heavy workflow lands there.

Request a scoped build. One-week discovery. You get a system inventory, workflow map, and fixed scope — whether or not you build with us.