Picking the right AI model is no longer about raw power. It is about price per task. On October 26, 2026, OpenAI launched GPT-6.1 Terra, a mid-tier model that costs half as much as its predecessor and comes close to OpenAI's most advanced model on several tasks. For an SME paying for AI tools by usage, this release illustrates a very concrete trade-off: the most expensive model is not always the one to use.
At a glance
- OpenAI launched GPT-6.1 Terra on October 26, 2026, available in ChatGPT (Plus, Pro, Enterprise, Edu, inside ChatGPT Work and Codex) and through the API as
gpt-6.1-terra. - Its API price is $1 per million input tokens and $5 per million output tokens ($0.05 for cached input), which is 50% cheaper than GPT-5.6 Terra, the model it replaces.
- On Agents' Last Exam (professional tasks), Terra scores 60.0% at $1.01 per task, versus 59.3% for GPT-6 Astra, OpenAI's most advanced model, which costs $7.24 per task.
- Analysts quoted by InfoWorld (Pareekh Jain, Sanchit Vir Gogia) warn that cheaper tokens do not guarantee net savings: extra human review time can cancel out the gain.
- For an SME, the right approach is not picking "the best model", but the cheapest model that reaches the quality bar needed for each specific task.
GPT-6.1 Terra: a model built for the mid-tier
GPT-6.1 Terra sits between two extremes in OpenAI's lineup: GPT-6 Luna, the cheapest and simplest model, and GPT-6.1 Sol followed by GPT-6 Astra, the most powerful and most expensive models. OpenAI describes Terra as "a fast, efficient model that combines detailed visual reasoning and precise computer use" (source: OpenAI announcement, cited by InfoWorld, October 27, 2026).
In practice, Terra targets high-volume tasks: document processing, customer support, routine code, information extraction. These are use cases where paying for a frontier model doesn't make economic sense, but where an entry-level model sometimes lacks nuance.
Before (GPT-5.6 Terra)
With GPT-6.1 Terra
The pricing table: four tiers, one API
OpenAI now structures its lineup in clear tiers. Understanding this table is the first step to managing AI costs.
| Model | Input price (per M tokens) | Output price (per M tokens) | Positioning |
|---|---|---|---|
| GPT-6 Luna | $0.10 | $0.50 | Simple tasks, very high volume |
| GPT-6.1 Terra | $1 | $5 | Everyday tasks with moderate reasoning |
| GPT-6.1 Sol | $2 | $10 | Complex reasoning, advanced code |
| GPT-6 Astra | $10 | $50 | The most demanding tasks, research |
Source: OpenAI's October 26, 2026 announcement, figures reported by InfoWorld. For comparison, Anthropic's Claude Sonnet 5.5 charges $2 input and $10 output per million tokens.
The price ratio between Astra and Terra reaches 10x. That is exactly the gap OpenAI is trying to close: on the internal Agents' Last Exam test, Terra reaches 60.0% success versus 59.3% for Astra, at roughly a seventh of the cost per task.
On visual reasoning (MMMU-Pro test), Terra also improves, from 81.7% to 85.5% success, at a cost per task ten times lower than its predecessor. These figures remain benchmarks reported by OpenAI itself; no independent body has publicly re-verified them yet, a detail worth keeping in mind before treating them as absolute fact.
Why a lower price doesn't mean a lower bill
This is the most useful point for an SME, and the least intuitive one. Two analysts interviewed by InfoWorld add an important nuance to this release.
The analysts' warning
Pareekh Jain (EIIRTrend & Pareekh Consulting) recommends testing each model on your own workloads, measuring completion rate, accuracy, cost per completed task, and human intervention time, rather than relying solely on vendor-published benchmarks. Sanchit Vir Gogia (Greyhound Research) goes further: on one example coding task, 5.4 extra seconds of human review time were enough to cancel out the savings made on token price.
The price of a token is only part of the real cost of an AI workflow. You also need to count the time spent checking, correcting, or re-running a poorly executed task. That is why the quality margin sometimes matters more than the listed price.
How an SME picks the right model, task by task
With four to five pricing tiers at each provider (OpenAI, Anthropic, Google), the temptation is either to keep the most expensive model "to be safe", or to shift everything to the cheapest one to cut costs. Both are costly shortcuts in the long run.
Map your AI tasks
Test the mid-tier model first
Measure the full cost, not just the token price
Reserve the top-tier model for critical tasks
This approach echoes what LUWAI has recommended since the first multi-tier pricing models launched in 2026: the challenge is no longer picking a single provider, but allocating the right model to the right task, and revisiting that allocation with every new release.
FAQ
Does GPT-6.1 Terra replace ChatGPT Plus for an SME?
No, Terra is a model available inside ChatGPT (Plus, Pro, Enterprise, Edu) and through the API, not a separate subscription tier. An SME already using ChatGPT can access it without changing plans, depending on their account configuration.
Is Terra as good as GPT-6 Astra?
On some tests like Agents' Last Exam, Terra comes close to Astra's results (60.0% versus 59.3%) at a much lower cost. On computer use (OSWorld 2.1), however, Claude Sonnet 5.5 still leads both, with 83.9% success. The right choice depends on the specific task.
How do I know if a cheaper model is good enough for my business?
By testing it on a real sample of your own tasks, not just on published benchmarks. Measure success rate, review time needed, and total cost, not just the price per token.
Should I switch AI providers every time a new model ships?
No. Switching providers carries integration and training costs. It is generally more cost-effective to regularly re-evaluate model allocation with your current provider, as OpenAI's, Anthropic's, and Google's multi-tier lineups now allow.
Weighing several AI models for your business? Check our AI model selection resources or see how other SMEs structured their AI costs in our success stories.
