On October 26, 2026, OpenAI shipped GPT-6.1 Terra. The most interesting part isn't the model itself, it's the pricing: the new mid-tier generation launched at exactly half the standard API price of its predecessor, $1 in / $5 out per million tokens, versus $2/$12 for GPT-5.6 Terra.
Read that carefully: this is a new generation launching at half price, not an older model getting discounted months after release. Those two things sound similar but mean very different things. One is clearance pricing. The other permanently moves the anchor for what the mid tier costs, and every future model in this slot will be measured against $1/$5.
Here's where the savings actually come from, the cost lever most launch coverage missed, and what you'll actually pay now that it's live on third-party platforms.
What "half price" actually means
Side by side, per million tokens (as of October 2026):
| Item | GPT-5.6 Terra (previous gen) | GPT-6.1 Terra | Change |
|---|---|---|---|
| Input | $2.00 | $1.00 | 50% |
| Output | $12.00 | $5.00 | ~58% |
| Cached read | $0.20 | $0.05 | ~75% |
| Cached write | — | $1.25 | — |
OpenAI's official line is "50% lower standard API prices," but the real cuts run deeper: output is down roughly 58%, and cached reads are a quarter of what they were. If you run high-volume calls and reuse system prompts or repeated context, the cached-read line saves you more than the headline number suggests.
There's a cheaper tier for asynchronous work: Batch/Flex mode takes another 50% off standard pricing, at $0.50 input / $2.50 output. The tradeoff is no real-time guarantee, which is fine for offline batch jobs, backfills, and overnight runs. In the other direction, Fast mode runs about 2x standard pricing, trading money for lower latency.
Mid-tier positioning: the volume workhorse
Terra sits in the middle of the four-tier GPT-6.1 ladder (Luna < Terra < Sol < Astra), built around two things: visual reasoning and computer use, plus technical writing and automated evaluation. Its job is to replace GPT-5.6 Terra as the volume workhorse. The implicit message: you no longer need the flagship to run computer-use tasks.
The official pitch: it matches flagship GPT-6 Astra on Agents' Last Exam, at one-tenth the price, with average task completion more than 3x faster than both Astra and GPT-6.1 Sol. Those are OpenAI-reported figures, not independently reproduced, so take them as vendor claims.
Third-party numbers give a firmer anchor. Artificial Analysis measured an Intelligence Index of 48, versus 52 for Sol and 38 for Luna; hallucination rate dropped from 88% to 38% versus the previous generation, and the low-reasoning-tier error rate fell from 53.6% to 16.7%. In plain terms: near-flagship intelligence at mid-tier pricing, with the weak spots noticeably tightened.
The real cost lever: reasoning.effort=none
The detail most launch coverage skipped: GPT-6.1 Terra is the only model in the family that exposes a reasoning.effort: "none" tier, and even Sol doesn't support it.
What does that mean? Models in this family run a reasoning chain by default, on a six-tier scale from low to high: deeper reasoning means smarter output, but also higher cost and latency. Plenty of batch tasks (format conversion, first-pass classification, templated generation) never need deep reasoning, and you used to pay for thinking you didn't use. Terra lets you switch reasoning off entirely, and once it's off you get full control of temperature/top_p. For anyone running high-frequency agents or batch pipelines, that's a direct cost switch: run the same task with deep reasoning off and Batch mode on, and the bill shrinks by another large factor.
If you're already running batch jobs on Sol, these two walk through the setup and tuning, and the same approach applies with the reasoning dial turned down on Terra: How to use GPT-6.1 Sol with the HiAPI API and GPT-6.1 Sol prompt recipes.
Platform availability: live on HiAPI
Within days of launch, multiple relay and aggregation platforms have added Terra. Markups vary a lot between platforms, so check each one's ratio rather than just confirming it's listed.
On HiAPI, GPT-6.1 Terra is live on its model page, speaking both Responses and Chat Completions, with a 1.05M-token context window. Platform pricing is set at a quarter of official:
| Model | Input ($/M tokens) | Output ($/M tokens) |
|---|---|---|
| GPT-6.1 Terra | $0.25 | $1.25 |
| GPT-6.1 Sol | $0.50 | $2.50 |
| GPT-6 Astra | $2.50 | $12.50 |
| Claude Opus 5.5 | $2.02 | $10.08 |
(Same-platform comparison, base tier, as of October 2026. Base tier covers contexts up to 272K tokens.)
Cached reads run $0.0125 and cached writes $0.3125. For documents over 272K tokens, the long-document tier applies: $0.50 input / $1.875 output, pricier than base but it means a 1.05M-token window fits a whole document in one call with no chunking pipeline.
The math is simple: on HiAPI, Terra is half the price of Sol and a full tier cheaper than Astra and Opus 5.5.
What this means for developers
Three workload types need a cost recalculation:
- Computer-use agents: Terra brings computer use to mid-tier pricing, so tasks that used to require a flagship now run at $1/$5 (with OpenAI claiming it's 3x faster than the flagship). For a hands-on look at switching models inside coding tools, see Switching models in Claude Code and Codex via HiAPI.
- Visual agents: visual reasoning is a headline capability, so screenshot understanding and UI parsing no longer need to borrow the flagship.
- Batch text workloads: stack reasoning.effort=none, Batch/Flex, and prompt caching, and per-token cost drops to a fraction of what it was.
For a cross-model comparison of intelligence versus price, this piece tests Terra's tier against Sol and Claude Opus 5.5: Claude Opus 5.5 vs GPT-6.1 Sol. And if world models are on your radar, our Genie 4 breakdown covers that direction.
FAQ
Is GPT-6.1 Terra just a renamed, discounted GPT-5.6 Terra?
No. It's an entirely new generation released on October 26, 2026, and its standard price launched at half the predecessor's. The older model wasn't discounted; it was replaced on pricing by the new one.
Is $1/$5 the official price or a platform price?
That's OpenAI's official standard API price per million tokens ($1 input, $5 output). HiAPI's platform price is $0.25/$1.25, a quarter of official, with cached reads at $0.0125. Prices are as of October 2026; check each platform's live page for current numbers.
When should I use the Batch/Flex tier?
At $0.50 input / $2.50 output, it's another 50% off standard pricing, with no real-time guarantee. Use it for offline batch processing, data cleaning, and scheduled overnight jobs. Don't use it for latency-sensitive work.
Does reasoning.effort=none make it noticeably dumber?
Depends on the task. Shallow work like format conversion, classification, and templated generation barely degrades with deep reasoning off. For multi-step reasoning, start low and step up until you find the quality/cost balance. Terra's six tiers make that tuning possible in a way Sol doesn't allow.
How does the $0.05 cached read save money?
Keep repeated system prompts and context prefixes stable so the platform cache hits, and hit tokens bill at $0.05, just 5% of the standard input price. For high-frequency calls, this line item is often the biggest lever.
How does HiAPI's long-document tier work?
Base tier (contexts up to 272K tokens) is $0.25/$1.25. Above 272K, the long-document tier applies at $0.50/$1.875. The context ceiling is 1.05M tokens, so a full long document fits in a single call.
Key takeaways
- "Half price at launch" is a reset of the mid-tier anchor, not a promotion: GPT-6.1 Terra standard pricing is $1/$5, with output actually down ~58% and cached reads down ~75%. Every future mid-tier model competes against this number.
- Official claims (matching the flagship at one-tenth the price, 3x faster) are vendor-reported; third-party testing (Intelligence Index 48, hallucination 88% to 38%) backs the "near-flagship intelligence at mid-tier pricing" story.
- reasoning.effort=none is unique to Terra in this family (Sol doesn't support it), and stacking it with Batch/Flex is the single biggest cost lever on the invoice.
- Terra is live on HiAPI at $0.25/$1.25 (a quarter of official), cached reads at $0.0125, 1.05M context, and a long-document tier at $0.50/$1.875. On the same platform it's half the price of Sol and a full tier below Astra and Opus 5.5. Prices as of October 2026; the model page shows live numbers.
