AI Tool Scout › AI Models Review

AI Models · Frontier LLMs · Updated October 2026

GPT-6.1 Terra Review (2026): $1/$5, Computer Use, 1.05M Context

Item: GPT-6.1 Terra
Rating: 4.0
Author: AI Tool Scout

OpenAI's new middle tier is the cheapest way to buy an agent that can actually click through a browser. It is also unavailable in the one place most people would look for it, and one billing cliff away from costing twice what the price card says.

By AI Tool Scout · Published October 28, 2026 · 17 min read

⚡ TL;DR — Is $1 / $5 for a computer-use model the bargain it looks like?

For agent builders, yes. For anyone else, not yet. OpenAI released GPT-6.1 Terra on October 26, 2026, replacing GPT-5.6 Terra and slotting between GPT-6 Luna and GPT-6.1 Sol. The API rate is $1.00 per million input, $0.05 cached and $5.00 output — half of Sol's $2/$10, a tenth of GPT-6 Astra's $10/$50, and about 58% below the $12 output rate of the model it replaces — with cache writes at $1.25 (30 min) or $2.50 (12 h), Batch and Flex at 50%, Fast at 2×, regional at +10%. Specification is 1,050,000-token context, 922,000 max input, 128,000 max output, knowledge cutoff June 1, 2026, text and image in with text out, reasoning effort none → max (default medium), and the full Responses API tool set including computer use, hosted shell, code interpreter, web search, file search and MCP. No fine-tuning, no weights, no audio, no video. Independent: Artificial Analysis 48 on Intelligence Index v4.3.2, 25th of 221 models, at $0.40 per index task — against Sol's 52 at $0.72 and GPT-5.6 Terra's 42 at $1.40. Vendor, so treat it as a claim: +15 points on Agents' Last Exam over GPT-5.6 Terra at medium effort for a third of the cost, parity with Astra at a tenth the price per token, and ~9× lower cost per task than Claude Opus 5.5. Two hard catches: anything over 272,000 input tokens bills the whole request at 2× input and 1.5× output, and it is not in the standard ChatGPT chat at all — Work and Codex on paid plans only. Costed honestly: 1,000 support calls a day at 20K in / 2K out is $30 a day, about $900 a month; the same task at 300K input is $0.62 instead of $0.31.

⚠️ Read the family ladder before the benchmarks. GPT-6 is no longer one model with one price. Luna is $0.10/$0.50, Terra is $1/$5, Sol is $2/$10 and Astra is $10/$50 — a hundredfold spread across four models that all carry the same 1.05M-token window. You are not buying context as you move up, and you are not buying speed; you are buying reasoning quality at a price per token that doubles or quintuples at every step. That makes the only question that matters: is the work you are doing hard enough that Terra's 48 beats Luna's 38 on your task, because if it is not, you are paying ten times the rate for a difference you will not see. The flagship line is in our GPT-6 Astra review, and cross-vendor rates are in the LLM API cost guide.

What Is GPT-6.1 Terra?

GPT-6.1 Terra (gpt-6.1-terra) is OpenAI's October 26, 2026 mid-tier model and the direct replacement for GPT-5.6 Terra. OpenAI's positioning is explicit and unusually modest for a launch: it is a fast, cheap model for visual reasoning, computer use and writing, not a new frontier system. The word "fast" is doing real work here — OpenAI says Terra finishes tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol, which matters when the task is a fifty-step browser session that a user is watching.

The family structure is the whole story of the release. Astra shipped first as the flagship; Luna and Sol followed in September to fill in the cheap and middle tiers; and Terra now occupies the gap between them, at a price that undercuts the model it replaces on every line. Read as a ladder:

Tier	Model	Input / 1M	Output / 1M	Cached / 1M	What it is for
Cheapest	GPT-6 Luna	$0.10	$0.50	$0.01	Classification, extraction, summarisation at volume
Middle	GPT-6.1 Terra	$1.00	$5.00	$0.05	Computer use, visual reasoning, agentic office work
Upper middle	GPT-6.1 Sol	$2.00	$10.00	$0.10	Complex coding and professional work
Flagship	GPT-6 Astra	$10.00	$50.00	$1.00	Hardest long-horizon reasoning and agentic work

Two things follow from that table and neither is in the launch copy. First, the price gaps are enormous relative to the score gaps. Terra is 10× Luna, Sol is 2× Terra, Astra is 5× Sol — while the same index puts them at 38, 48, 52 and 53. Going from Luna to Terra buys ten points and costs ten times as much; going from Sol to Astra buys one point and costs five times as much. Second, the previous generation's mid-tier is now the expensive option. GPT-5.6 Terra was $2 in and $12 out; GPT-6.1 Terra is $1 and $5. If you are still pinned to the old model ID for stability reasons, you are paying double for six fewer index points.

Availability is where the release gets strange. Terra is live in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu, and on the API at the Build tier or above, with a gradual GitHub Copilot rollout across Pro, Pro+, Max, Business and Enterprise plans. It is not in the ordinary ChatGPT chat interface, and Free and Go users cannot reach it at all. On Enterprise and Edu workspaces an administrator may have to enable it before members see it — which means a rollout that "did not work" is frequently a toggle. For a model whose entire pitch is high-volume everyday work, launching it everywhere except the everyday chat surface is a decision worth questioning, and it is the first item in our cons list.

Key Features: Seven Things Worth Knowing

Computer use is the product, not a side feature. Terra reads a screen, identifies interface elements, and clicks and types through multi-step tasks across browsers, desktop software and other applications — the capability that separates an agent from a chatbot. It runs through the computer_use tool on the Responses API, so the model is looking at rendered pixels and emitting actions rather than parsing an API. That is what makes it usable on software that has no API at all, and it is also why the privacy section of this review matters more than usual.
A 1,050,000-token window — with a 922,000-token input ceiling. The window is a shared budget: 922,000 maximum input plus 128,000 maximum output equals exactly 1,050,000, so a long generation and a full document compete for the same space. Size your prompts against 922,000, not 1,050,000. The knowledge cutoff is June 1, 2026, which is four months stale at launch and a real constraint for anything time-sensitive.
The 272,000-token billing cliff is the single most important number on the card. Prompts above 272,000 input tokens bill at 2× input and cache rates and 1.5× output for the entire request. Not the excess — the whole thing. This is a pricing model that punishes exactly the long-document work a million-token window advertises, and it is the trap most launch coverage skipped.
Reasoning effort runs none → max, defaulting to medium. That is a genuine cost lever that earlier OpenAI reasoning models did not give you: a classification task can run at none and a hard agentic recovery can run at max, on the same model ID. Note that Artificial Analysis reports the largest gains over GPT-5.6 Terra at the lower effort levels — seven to thirteen points at low and medium against six at max — which is another way of saying the cheap settings are where this model earns its price.
The full Responses API tool set ships with it: computer use, a hosted shell, code interpreter, web search, file search and MCP. Tool calling requires the Responses API; Chat Completions is available without tools. For an agent that has to reach a real system, that combination — screen control plus a shell plus retrieval plus MCP connectors — is the difference between a demo and a deployment.
Text and image in, text out. No audio, no video. For an agent that watches a screen this is a defensible limit, since computer use is a stream of screenshots. For anything involving a call recording, a screen recording or a video walkthrough, Terra cannot be the model, and no amount of context window changes that.
No fine-tuning and no weights. Terra is closed and OpenAI-hosted only. There is no checkpoint to download, no licence to read, and no way to specialise it on your own data beyond prompting, retrieval and tools. If your compliance position requires self-hosting, this model is simply not an option and the open-weight families are the only route.

What OpenAI Claims, and What the Independent Index Says

OpenAI's launch numbers are specific, which is to its credit, and vendor-run, which is the caveat. The claims, all as stated:

Agents' Last Exam: 15 percentage points above GPT-5.6 Terra at medium effort, at about one-third the cost. This is the benchmark OpenAI leads with — professional computer-use tasks such as turning architectural drawings into a 3D model — and it is the one that maps to the model's stated purpose.
"Performs on par with GPT-6 Astra at only one-tenth its price per token." True as a statement about the price ratio, and worth reading carefully: Astra costs 10× per token, so parity at a tenth of the price would be an extraordinary result if it held across workloads. It is OpenAI's own harness on a benchmark OpenAI chose.
Against Claude Opus 5.5 at max effort: over three times the token efficiency, a quarter of the price per token, and roughly 9× lower cost per task. The token-efficiency leg is doing most of the work in that arithmetic, and token efficiency is the one thing you cannot verify from a rate card.
At medium effort, a single task costs $0.29 and succeeds 57.6% of the time, against 42.6% and $0.84 for GPT-5.6 Terra. That is the most useful number in the announcement, because it prices a completed task rather than a token.

Artificial Analysis, which runs its own harness, is less flattering and more useful. Terra scores 48 on Intelligence Index v4.3.2 — 25th of 221 models, at $0.40 per index task. The comparison that matters is not against Astra but against the model it replaces: GPT-5.6 Terra scores 42 at $1.40 per task, so Terra buys six index points for 29% of the cost. Against GPT-6.1 Sol at 52 and $0.72, it gives up four points and saves 44%. Artificial Analysis also puts its Coding Agent Index v2.0 score at 48, six below Sol's 54 and seventeen above Luna's 31, and describes the low and medium effort settings as Pareto-optimal for cost efficiency while the higher effort levels sit behind the frontier.

Read together, the vendor and independent pictures agree on the shape and disagree on the framing. Nobody credible is claiming Terra is a frontier-intelligence model. It is a model that finishes a cheap task for less money than anything else in its class, and the honest test of whether that matters is your own workload — because a 25th-place composite can still be first place on the narrow task you actually run.

Pricing

Figures below are from OpenAI's developer documentation and pricing page, checked October 28, 2026, and cross-read against independent model catalogues. All amounts are US dollars before tax. Confirm on OpenAI's pricing page before you build — this family has changed rates twice in six weeks.

API — standard short-context rate

gpt-6.1-terra: $1.00 per 1M input · $0.05 per 1M cached input · $5.00 per 1M output

Cache writes: $1.25 per 1M at 30 minutes · $2.50 per 1M at 12 hours

Batch and Flex: 50% of standard ($0.50 / $0.025 / $2.50) · Fast mode: 2× standard · Regional processing: +10% where offered

Long context: any request above 272,000 input tokens bills at 2× input and cache, 1.5× output, for the whole request — i.e. $2.00 / $0.10 / $7.50 per 1M

Reasoning tokens bill at the output rate. No subscription on the API. Cache reads at 5% of input are the single largest lever OpenAI gives you on this model.

What it replaces, and what it undercuts

GPT-5.6 Terra: $2.00 / $0.20 / $12.00 → GPT-6.1 Terra: $1.00 / $0.05 / $5.00. That is 50% off input, 75% off cache reads and 58% off output at a generational replacement, while scoring six index points higher. OpenAI frames the cut as "50% lower than GPT-5.6 Terra and GPT-6.1 Sol", which is accurate on input and understates the output reduction against the older model.

Against its siblings: exactly half of GPT-6.1 Sol ($2/$10) on both legs, a tenth of GPT-6 Astra ($10/$50), and ten times the price of GPT-6 Luna ($0.10/$0.50). Against Claude Opus 5.5 at $4/$20 it is a quarter of the token price.

In ChatGPT: plan-based, not per token. Terra is included in ChatGPT Work and Codex on Plus and above, and OpenAI says the same allowance covers more tasks with Terra than with GPT-5.6 Terra, and "substantially more" than with Sol or Astra — direction, not a multiplier, since no exact ratio was published.

Cost example 1 — a support agent at 1,000 calls a day

The representative workload for this model is not a chat message; it is an agent that answers something, calls a tool, and checks the result. Take 1,000 calls a day, each sending 20,000 input tokens — system prompt, tool schema, retrieved policy documents and conversation history — and receiving 2,000 output tokens of answer plus reasoning.

The arithmetic, written out (standard rate, no caching)

Input: 1,000 × 20,000 = 20,000,000 tokens a day × $1.00 ÷ 1M = $20.00
Output: 1,000 × 2,000 = 2,000,000 tokens a day × $5.00 ÷ 1M = $10.00
Total: $30.00 a day → about $900 a month (30 days). Per call: $0.03.

With 80% of input served from cache — a stable system prompt and a pinned policy corpus, which is exactly what a support agent has: 16M × $0.05 ÷ 1M = $0.80, plus 4M × $1.00 ÷ 1M = $4.00, so input falls to $4.80 and the day to $14.80, about $444 a month. At a 50% hit rate it is $615 a month. Caching is worth more here than any vendor choice on this page.

The same 1,000 calls a day	Input / 1M	Output / 1M	Per day	Per month
GPT-6.1 Terra, no cache	$1.00	$5.00	$30	~$900
GPT-6.1 Terra, 80% cache hits	$0.05	$5.00	$14.80	~$444
GPT-5.6 Terra ($2 / $12)	$2.00	$12.00	$64	~$1,920
GPT-6.1 Sol ($2 / $10)	$2.00	$10.00	$60	~$1,800
GPT-6 Astra ($10 / $50)	$10.00	$50.00	$300	~$9,000
GPT-6 Luna ($0.10 / $0.50)	$0.10	$0.50	$3	~$90
MiniMax M3.1 Flash (~$0.15 / ~$0.60)	$0.15	$0.60	~$4.20	~$126

Our arithmetic at published list prices, with the token assumptions stated above. Your real prompts will be different — measure them. Competitor rates approximate, checked October 28, 2026.

Read the Luna line twice. The same workload costs $90 a month on Luna against $900 on Terra, a tenfold difference, and the honest question is whether a support triage agent needs a 48 on the intelligence index or whether 38 is enough. For classification, retrieval-grounded answering and structured extraction, it very often is. Terra's case is that its best tasks — multi-step screen operation, visual reasoning, long tool chains — are precisely the tasks where a cheaper model degrades badly rather than gracefully. But if your agent is doing sentiment routing, you are buying a computer-use model to do a job Luna does for a tenth of the price.

Cost example 2 — the same task at 300,000 input tokens

Now take one task with a long document in context: 300,000 input tokens and 2,000 output tokens. Nothing else changes. This crosses the 272,000-token threshold, so the whole request is repriced.

The arithmetic, written out

At standard rates (if the threshold did not exist): 300,000 × $1.00 ÷ 1M = $0.30 input, plus 2,000 × $5.00 ÷ 1M = $0.01 output → $0.31.

As actually billed: input at 2× → 300,000 × $2.00 ÷ 1M = $0.60; output at 1.5× → 2,000 × $7.50 ÷ 1M = $0.015 → $0.62.

That is 1.98× — about 98% more for the same work, because the multiplier applies to every token in the request. Run a thousand of those a day and the bill is $615 a day, about $18,450 a month, against $900 for the 20K version. Cached aggressively (270,000 of the 300,000 tokens pinned) the same task lands near $0.10, which is the only reason long-context work is affordable on this model at all.

The cliff is sharper than the multiplier suggests. A 272,000-token prompt bills $0.272 of input; a 272,001-token prompt bills $0.544. One token more than doubles the input charge for the entire request, because the threshold reprices the request rather than the excess. That is a discontinuity you cannot smooth out with engineering, only with a token counter — so if you are near the line, cut a document, summarise a section, or chunk it. Note also that the window allows 922,000 input tokens, which is more than three times the threshold: the model is advertised for a load it is priced to discourage.

Is It Free?

No — there is no free path, on any surface, in any configuration.

Not on Free or Go. Neither plan includes GPT-6.1 Terra. The Free and Go desktop tier runs GPT-6 Luna, which is a different model at a tenth of the price.
Not in ordinary ChatGPT on paid plans either. Plus, Pro, Business, Enterprise and Edu reach Terra through ChatGPT Work and Codex only; the standard chat interface does not carry it. So the ceiling on "free" is not a quota, it is a surface that does not exist.
No free API tier. Every token bills from the first one, and API access requires a Build-tier organisation or above — which is a billing relationship, not a trial. There are no launch credits documented.
No open-weights fallback and no distillation route. The model is closed and OpenAI-hosted only, and fine-tuning is unsupported, so there is not even a smaller fine-tuned variant you could host. Compare Kimi K3.1 Pro, where weights are at least promised, or MiniMax M3.1 Flash at a seventh of the price.
What is genuinely free: the free chat tiers at DeepSeek, Qwen, Gemini and Kimi, and Luna in the ChatGPT desktop app on Free and Go. None of them will operate a browser the way Terra does, and all of them are free.

Practical read: evaluate Terra on a small paid API budget or inside an existing Work/Codex plan, then budget at $1.00 / $5.00 per million for anything you keep running — and add the 2× long-context line if your prompts are large.

Pros & Cons

✅ Pros

$1.00 / $0.05 / $5.00 is a genuinely aggressive rate card. Half of GPT-6.1 Sol, a tenth of GPT-6 Astra, and cheaper than the GPT-5.6 Terra it replaces on all three legs.
Six index points for 29% of the cost. $0.40 per index task against $1.40 for GPT-5.6 Terra, independently measured.
Computer use is a real API capability, with the complete Responses API tool set — hosted shell, code interpreter, web search, file search, MCP — available alongside it.
The full 1,050,000-token window at the same price. No separate long-context model to buy into, even if the 272K line makes large requests expensive.
Reasoning effort from none to max, defaulting to medium. A real cost dial, and the low settings are where the biggest gains over the previous generation land.
Fast, as advertised. OpenAI claims more than 3× the task completion speed of Astra and Sol, which matters for interactive agents; Artificial Analysis measures about 91 output tokens per second at max effort.
$0.03 per 20K-input task at list price — cheap enough that a thousand-call-a-day agent is a $900 line item rather than a budget conversation.
Cache reads at 5% of input are the best lever on the card: an 80% hit rate takes our modelled month from $900 to $444.
An answer format built for agents: OpenAI says Terra writes the result first and the supporting detail second, which is the right shape for a status report a human skims.

❌ Cons

You cannot buy it in the ordinary ChatGPT chat. Work and Codex only, on paid plans, sometimes behind an admin toggle. Anyone who wanted a cheaper everyday assistant does not get one.
The 272,000-token cliff reprices the whole request at 2× input and 1.5× output. A 300K task costs $0.62 instead of $0.31 — the long-document workload this window advertises is the one it taxes.
25th of 221 models on the independent index. At 48 it is a value pick, not an intelligence pick; anyone buying it for the hardest reasoning will be disappointed.
"On par with Astra" is a vendor comparison that OpenAI designed, on a benchmark OpenAI chose, against a model priced ten times higher. It is the most favourable possible frame.
No audio and no video input, and text-only output. Screen streams are images, so computer use works — call recordings and video walkthroughs do not.
No fine-tuning and no weights. Closed, OpenAI-hosted, no licence, no self-hosting, no distillation route.
GPT-6 Luna is a tenth of the price at $0.10/$0.50, and adequate for the classification and extraction work that makes up most agent traffic.
Four-month-stale knowledge cutoff (June 1, 2026) on a model pitched at professional document work.
1,050,000 context is really 922,000 input once the 128,000 output reservation is taken out, and the marketing never says so.

5 Best GPT-6.1 Terra Alternatives

GPT-6 Astra

Best if five index points and the hardest reasoning are worth ten times the token price

53 on the same index against Terra's 48, at $10.00 / $50.00 per 1M — the flagship of the same family, and the model Terra's launch claims are drawn against.
Built for long-horizon autonomous work: the strongest agentic and long-context retrieval results in the family, which is where the price difference actually shows up.
Ten times the price on both legs — $9,000 a month against $900 on our 1,000-call model. That is a decision, not a rounding difference.
Note that the comparison cuts both ways: if Terra really is parity at a tenth the price on your workload, you should measure it, because the person claiming that is the vendor.
Read the GPT-6 Astra review for the launch detail, the benchmark caveats and the plan requirements.

Kimi K3.1 Pro

Best if you want more intelligence per token and can live without OpenAI's computer-use tooling

55 on Artificial Analysis Intelligence Index v4.3.2 against Terra's 48, at $2.00 / $0.20 / $10.00 — twice Terra's rate for seven more points, and flat across a 1,048,576-token context with no long-context surcharge.
That last point is the one that matters for document work: a 300,000-token request costs Kimi $0.60 in input with no 2× repricing, where Terra charges $0.60 for the same tokens only because it doubled the rate from $0.30.
Terminal-Bench 4.0 at 65.7% and a strong agentic profile, though its Terminal-Bench Science row at 38.1% is a genuine weak spot.
Weights are promised for November 4, 2026 with the licence unpublished — a promise, not a checkpoint.
Our Kimi K3.1 Pro review has the full harness-by-harness breakdown.

MiniMax M3.1 Flash

Best if your bottleneck is volume and you can accept a capability drop on long tool chains

~$0.15 input / ~$0.60 output / ~$0.015 cached per 1M — about $126 a month on our 1,000-call model against Terra's $900, and a 1,048,576-token context.
428B total / ~23B active MoE with a native vision encoder and five thinking levels from Low to Max.
The honest trade: it will hold up on extraction, summarisation and classification, and degrade on the multi-step screen operation Terra was built for. If your agent needs to click through an application, this is not the substitute.
No free API tier, and weights are promised without a date.
See the MiniMax M3.1 Flash review.

GPT-6 Luna

Best if you are honest about how much intelligence your workload actually needs

$0.10 input / $0.50 output / $0.01 cached per 1M — a tenth of Terra on every leg, and the same 1,050,000-token window with the 922,000 / 128,000 split.
38 on the same index, so you give up ten points and 90% of the bill. On our support-agent model that is $90 a month against $900.
Reachable in more places than Terra: the API as gpt-6-luna, and the ChatGPT desktop app on Free and Go at Standard speed.
Also carries the same 272K long-context repricing, so the cliff is a family policy rather than a Terra quirk.
Wrong choice for computer use and multi-step visual work; right choice for everything that is really text in, structured text out.

Claude Opus 5.5

Best if you want the highest composite score and a published system card

58 on the same index against Terra's 48 — the largest gap on this page — at $4.00 / $20.00 per 1M, four times Terra's token price.
OpenAI's own comparison target, and the claim that Terra is 9× cheaper per task against it depends on Terra using a third of the tokens, which is the part you should verify on your workload.
The reference for hard reasoning, with a documented model card and no jurisdiction question.
At $3.60 per 30-step run against Terra's $1.80 for the same token counts, it is the expensive answer to a question Terra answers cheaply.
See our Claude review for plan pricing and the wider product picture.

Comparison Table

Model	Input /1M	Output /1M	Cached /1M	Context	AA index	Where you can use it
GPT-6.1 Terra	$1.00	$5.00	$0.05	1,050,000	48	API (Build+), ChatGPT Work, Codex, GitHub Copilot — not ordinary ChatGPT
GPT-6 Luna	$0.10	$0.50	$0.01	1,050,000	38	API, ChatGPT Work/Codex on paid plans, Free and Go desktop app
GPT-6.1 Sol	$2.00	$10.00	$0.10	1,050,000	52	API (Build+), ChatGPT Work, Codex
GPT-6 Astra	$10.00	$50.00	$1.00	1M	53	API, ChatGPT Work, Codex, Pro tier for chat access
Claude Opus 5.5	$4.00	$20.00	$0.20	1M	58	Anthropic API, Claude apps, major clouds
Kimi K3.1 Pro	$2.00	$10.00	$0.20	1,048,576	55	Kimi API, Kimi.ai, Kimi Work, Kimi Code, OpenRouter
MiniMax M3.1 Flash	$0.15	$0.60	$0.015	1,048,576	—	MiniMax API, MiniMax Code

Prices approximate, in US dollars, checked October 28, 2026, and they change often. Scores are Artificial Analysis Intelligence Index v4.3.2 at max reasoning effort; a dash means the model is not scored on that index at the time of writing, not that it scored zero. Terra and the two Sol/Luna rows also carry the 272K long-context repricing.

Which One Should You Choose?

Computer-use and browser automation — Terra is the buy. Screen reading, element identification and clicking at $0.03 a task with ~3× the task speed of Astra and a tenth of its price per token is the strongest case in this release. Build on the Responses API, use computer_use plus a hosted shell, and start at medium effort. Measure before you scale, because the benchmark that justifies the choice is OpenAI's.
Long documents — Terra only if you can stay under 272,000 tokens. Above that line every request doubles on input and 1.5× on output, and a 300K task goes from $0.31 to $0.62. Either chunk the document, or buy a model with a flat long-context rate: Kimi K3.1 Pro prices 1,048,576 tokens flat at $2.00 input, and Gemini 4 Argon is competitive while its promotional rate lasts.
Cheap high-frequency volume — go to Luna first and earn your way up. At $0.10/$0.50 Luna is a tenth of Terra, and most "agent" traffic is retrieval, routing and extraction where the ten-point index gap does not surface. Move the specific step that fails to Terra, and keep the rest on Luna. Mixed routing is the single biggest saving available on this family.
Maximum intelligence — Terra is the wrong rung. Claude Opus 5.5 at 58, Kimi K3.1 Pro at 55, GPT-6 Astra at 53 and Sol at 52 all beat Terra's 48. If the task is hard reasoning rather than cheap execution, paying four to ten times Terra's rate is the correct call and the cost-per-task arithmetic will usually agree once you count retries.
You want a cheaper everyday ChatGPT — this is not it. Terra is not in the standard chat surface at all. If your goal is lower spend on chat, use ChatGPT with the default model, or a free tier.
You need to self-host or fine-tune — Terra cannot help. Closed, hosted, no checkpoint, no tuning. The open-weight path is elsewhere on this page.

Cost Per Task vs Price Per Token — Why the $0.40 Matters More Than the $1.00

Every model launch is argued in price per million tokens, and that number is close to useless on its own, because tokens are not the unit anyone buys. You buy a completed task: a resolved ticket, a merged pull request, a filled form, an invoice reconciled. Two models can sit at the same rate and differ threefold on what a task costs, for two reasons that a rate card cannot show.

The first is token efficiency. A model that solves a task in 12,000 output tokens costs a third of what the identical model costs at 36,000 tokens, at the same price. This is exactly the mechanism behind OpenAI's "$0.29 a task at 57.6% success, against $0.84 for GPT-5.6 Terra" figure and behind its claim of 9× lower cost per task than Claude Opus 5.5 — the multiplier there comes from Terra using a third of the tokens at a quarter of the price. It is also why OpenAI's "parity with Astra at a tenth of the price" is not as absurd as it sounds: Astra's advantage is reasoning depth, and on a task Astra solves in one pass and a weaker model needs four attempts for, the cheaper model's cost advantage halves with every retry.

The second is the failure tax. A model that fails 43% of the time has not saved you 57% of a task's cost; it has spent that money and produced nothing, and then you pay again. This is why Artificial Analysis's $0.40 per index task at a score of 48 is the most decision-useful number in this review: it is a price for work completed on a fixed, third-party harness, measured the same way across models. On that measure Terra is 3.5× cheaper per completed task than the GPT-5.6 Terra it replaces ($1.40) and 44% cheaper than GPT-6.1 Sol ($0.72) — better than the raw rate card implies, because efficiency compounds.

The honest counterpoint is that cost per task is only comparable within a benchmark. On your workload the token counts and the pass rate are both different, and the only measurement that counts is yours: run 50 real tasks, count the tokens, count the failures, and multiply. That exercise takes an afternoon and it is worth more than every benchmark table on this page, including ours. The rule of thumb is that a cheaper model is only cheaper if it finishes — and the corollary is that the 272K cliff can quietly erase the whole advantage, since a doubled input rate applies to every token in a long request, including the ones you cached specifically to avoid paying for.

Privacy & Data

Computer use transmits your screen. This is the privacy fact that separates Terra from a text model: an agent operating a browser sends screenshots of whatever is visible — including notifications, other tabs, inbox previews and anything a colleague left open. Treat the display as an input channel, run agents on dedicated profiles or virtual machines, and never assume the model sees only the window you had in mind.
Closed and OpenAI-hosted only. There is no checkpoint, no on-premises option and no fine-tuning, so your data-processing position is whatever OpenAI's API terms say, plus whatever your region configuration adds. OpenAI offers regional processing at a 10% premium where it is available; if data residency is a requirement, price that line in.
API data handling differs from consumer surfaces. OpenAI's standard API terms do not use API content to train models by default, and enterprise agreements typically add retention limits and zero-retention options. Verify the terms your organisation is actually on rather than assuming the API defaults extend to ChatGPT Work, which is a different product with different settings.
Check the admin toggle before you blame the model. On Business, Enterprise and Edu, Terra may be disabled until an administrator enables it. If a rollout "does not work", that is the first place to look.
Plan allowances are not published multipliers. OpenAI says the same allowance covers more tasks with Terra than with GPT-5.6 Terra; no ratio was given. Capacity planning on an unquantified claim is how teams discover a limit in production.
Agent actions are your liability, not the vendor's. A model that clicks and types will eventually click the wrong thing. Put confirmations in front of destructive operations, scope credentials narrowly, and log agent actions — this is standard practice for any computer-use deployment and it matters more here because the model is cheap enough to run at volume.

Frequently Asked Questions

Can I use GPT-6.1 Terra in normal ChatGPT?

No. GPT-6.1 Terra shipped on October 26, 2026 into ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu accounts, and it is not in the standard ChatGPT chat interface. Free and Go plans cannot reach it at all. On Business, Enterprise and Edu workspaces an administrator may also have to enable the model before members see it, so a rollout that looks broken is often an admin toggle rather than a bug. Developers call it through the API as gpt-6.1-terra, and API access requires the Build tier or above. If you wanted this model as an everyday chat assistant, this release does not give you that; ChatGPT running the default model, covered in our ChatGPT review, remains the cheaper way to chat.

How much does GPT-6.1 Terra cost?

In the API the standard short-context rate is $1.00 per million input tokens, $5.00 per million output tokens and $0.05 per million cached input tokens. Cache writes are $1.25 per million for thirty minutes of retention and $2.50 for twelve hours. Batch and Flex processing run at half the standard rate, Fast mode doubles it, and regional processing adds ten percent. The trap is the long-context line: any request whose prompt exceeds 272,000 input tokens is billed at twice the input and cache rates and 1.5 times the output rate for the entire request, not just the excess. Worked example at the standard rate: 1,000 calls a day sending 20,000 input and 2,000 output tokens each is 20 million input and 2 million output tokens a day, which comes to $20 plus $10, so $30 a day and about $900 a month. In ChatGPT the price is plan-based rather than per token: Terra is included in ChatGPT Work and Codex for Plus and above.

Is GPT-6.1 Terra free?

No. There is no free path to GPT-6.1 Terra on any surface. It is absent from the free and Go ChatGPT plans, and OpenAI lists no free API tier for it: every token is billed from the first one, and API access requires a Build-tier organisation. There is no open-weights fallback either, because the model is closed and OpenAI-hosted only, and fine-tuning is not supported, so there is not even a cheaper distilled variant you could run yourself. The free options that remain are the desktop app's Free and Go tier running GPT-6 Luna, and the free chat tiers at DeepSeek, Qwen, Gemini and Kimi. Treat GPT-6.1 Terra as a paid tool with a metered API and a plan gate, and budget accordingly.

Is GPT-6.1 Terra as good as GPT-6 Astra?

Not on the composite, and the claim that it is needs a caveat. OpenAI says Terra performs on par with GPT-6 Astra on Agents' Last Exam while costing a tenth as much per token, but that is a vendor-run comparison on a benchmark OpenAI selected, and Astra is priced at $10 and $50 against Terra's $1 and $5, so the comparison is drawn in the most flattering possible frame. On Artificial Analysis Intelligence Index v4.3.2, GPT-6.1 Terra scores 48 and GPT-6 Astra scores 53, so the flagship is five points ahead. Where Terra genuinely wins is cost per finished task: $0.40 per index task against $0.72 for GPT-6.1 Sol and $1.40 for the GPT-5.6 Terra it replaces. The fair summary is that Terra is the cheapest model in the family that can still operate a browser, and that Astra is the one to reach for when five index points and the hardest reasoning matter more than the bill.

What is the 272K long-context surcharge on GPT-6.1 Terra?

OpenAI bills any request whose prompt exceeds 272,000 input tokens at double the input and cache rates and 1.5 times the output rate, and the multiplier applies to the whole request rather than only to the tokens above the threshold. That makes 272,000 a cliff rather than a slope: a 272,000-token prompt bills $0.272 of input at the standard rate, and a 272,001-token prompt bills $0.544, because the second request is priced at $2 per million instead of $1. The context window is 1,050,000 tokens with a maximum input of 922,000, so you can send more than three times the threshold in one call and pay double for all of it. On our modelled task of 300,000 input and 2,000 output tokens, the bill is about $0.62 against $0.31 at the standard rate, roughly 98 percent more for the same work. If long documents are your workload, price the 2x line into the budget before you commit, or chunk the document.

What are the best GPT-6.1 Terra alternatives?

Four, depending on what you are optimising. If you need the most capable model in the family and cost is secondary, GPT-6 Astra at $10 and $50 is the flagship, covered in our GPT-6 Astra review. If you want more intelligence per token and can live without OpenAI's computer-use tooling, Kimi K3.1 Pro at $2 and $10 scores 55 on the same index against Terra's 48, and its weights are promised for November 4, 2026. If your bottleneck is volume, GPT-6 Luna at $0.10 and $0.50 is a tenth of Terra's price and scores 38, while MiniMax M3.1 Flash at about $0.15 and $0.60 is the cheapest credible agent model on our list. If you want the highest composite score from a vendor with a published system card, Claude Opus 5.5 at $4 and $20 scores 58.

AI ModelsGPT-6.1 TerraOpenAIComputer UseAgentic AI1M ContextAPI Pricing2026

Free tools we built

Handy free tools, no sign-up

📄 PDF Merge Online

Combine multiple PDF files into one — free, no upload, files stay on your device.

📒 Bookkeeping Reviews

Honest accounting software reviews: Wave, QuickBooks, Xero & free options.

🎮 Gaming Gear Reviews

Esports hardware decoded: gaming mice, keyboards, headsets & budget picks.

📷 ID Photo Studio

Passport & visa photos with AI background changer, compressor & converter.

🧮 HK Salary Tax Calculator

Estimate Hong Kong salaries tax for 2026/27 in seconds — see what you owe before you file your tax return.

🎁 HK Benefits Quick Check

Government & benefits eligibility checker — OAA, CSSA, housing & more.

📅 HK Calendar 香港月曆

Public holidays, lunar dates & solar terms for 2020-2036 — bilingual, free, no sign-up.

★ Our own product 🏠 Home Expense Bookkeeping

Family bookkeeping: income, spending, net worth and monthly bills. Data stays on your device. HK$98 one-time.

Explore AI Tool Scout: Home About All Free Tools GPT-6 Astra Review ChatGPT Review Claude Review Gemini Review Gemini 4 Argon Review DeepSeek Review DeepSeek V4.1-Flash Review Qwen Review Kimi K3 Review Kimi K3.1 Pro Review MiniMax M3.1 Flash Review Mistral Voxtral 3 Review Inception Mercury 3 Review OpenAI Dots Review Meta Muse Review Meta Muse Spark Open Review Meta Muse Wonder Review Grok Review Perplexity Review Cursor Review GitHub Copilot Review Notion AI Review Otter.ai Review Grammarly Review Canva Review Wave Accounting Review Sora Review Runway Review Kling AI Review Hailuo AI Review HeyGen Review ElevenLabs Review Suno Review Midjourney Review Leonardo AI Review Ideogram Review Unboring AI Review LLM API Costs Gemini 4 Carbon Review Mistral Large 4 Review Claude Opus 5.5 Review Meituan LongCat-2.5 Review

AI Tool Scout — independent reviews of AI tools, updated regularly. See more free tools in our directory. Pricing, specifications and benchmark figures in this review were checked on October 28, 2026 against OpenAI's developer documentation for gpt-6.1-terra (context window, modalities, knowledge cutoff, reasoning effort, tool support and fine-tuning), OpenAI's API pricing page and the October 26 launch post, Artificial Analysis's model page, comparison pages and launch analysis for Intelligence Index v4.3.2, and same-week reporting from ghacks, MarkTechPost, Technology.org, ExplainX and Simon Willison. Benchmark results attributed to OpenAI are vendor-reported on vendor-selected harnesses and have not been independently reproduced; where OpenAI's figures and Artificial Analysis disagree, both are stated. The cost tables are our own arithmetic at published list prices using the token assumptions stated in each example, and are not vendor quotes. API prices, plan allowances, long-context thresholds and promotion windows change frequently — confirm current figures on the vendor's official pages before you buy, subscribe or deploy. Companies and products are the property of their respective owners; this review is not affiliated with OpenAI, Anthropic, Moonshot, MiniMax, Google or Meta.