![GPT-6.1 Terra in ChatGPT Work and Codex: $1/$5 [2026]](https://shattered.io/wp-content/uploads/2026/10/gpt-6-1-terra-chatgpt-work-codex-api-pricing-2026-gen.webp)
OpenAI’s developer documentation describes GPT-6.1 Terra as a “fast, efficient model for visual reasoning, precise computer use, and clear, natural writing.” The model carries the API ID gpt-6.1-terra, runs inside ChatGPT Work and Codex, and costs $1 per million input tokens and $5 per million output tokens at standard rates. That is the confirmed core of a story many headlines frame as OpenAI upgrading ChatGPT and Codex.
The word upgrade needs a caution flag. OpenAI’s materials say Terra is available in Work and Codex. They do not call it an upgrade, so this analysis uses availability language and treats the upgrade framing as headline shorthand. If you want a side-by-side against Anthropic and Google flagships, our best AI model comparison of Terra, Opus 5.5 and Argon covers that ground. This piece sticks to what OpenAI has put on the record, what the numbers mean in dollars, and what nobody has confirmed yet.
Don't miss new tech stories on Google
Add Shattered once in the Google app and our stories show up in your news suggestions.
What OpenAI Confirmed About GPT-6.1 Terra #
Start with the facts that survive scrutiny. OpenAI announced GPT-6.1 Terra, and its materials list a short, specific set of properties. Everything below comes from OpenAI’s own announcement and documentation, not from third-party testing.
- OpenAI announced GPT-6.1 Terra on April 9, 2026, according to the verified record for this story.
- The official API model ID is
gpt-6.1-terra. - The model is available through the OpenAI API, ChatGPT Work and Codex.
- ChatGPT availability covers Plus, Pro, Business, Enterprise and Edu plans, depending on plan, region and workspace settings.
- Standard API pricing is $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens.
- The context window is 1.05M tokens and maximum output is 128K tokens.
- The listed knowledge cutoff is June 1, 2026.
- Using the model in Codex requires Codex CLI 0.159.0 or later.
Equally important is what the record leaves out. No named OpenAI executive or engineer appears on the record in the cited official materials, so this article carries no expert quotes. Benchmark scores and launch-pricing details beyond the figures above also sit outside the confirmed set. Later sections flag each gap instead of papering over it.
GPT-6.1 Terra Availability: ChatGPT Work, Codex and Plan Tiers #
The ChatGPT Learn models page describes the models that power ChatGPT Work and Codex, and Terra is one of them. The rollout covers five paid ChatGPT tiers. Plus and Pro serve individuals, while Business, Enterprise and Edu serve organizations. OpenAI qualifies that list with three variables: your plan, your region and your workspace settings. A Business admin can therefore see a different picture than a Plus subscriber in another country.
Which surfaces get Terra
Terra lives in two ChatGPT surfaces and one developer surface. In ChatGPT it appears in Work and Codex. In the API it appears under the gpt-6.1-terra ID. OpenAI’s help center, as retrieved during our research, describes Work as the place for longer multi-step jobs and finished deliverables, and Codex as the home for software development and technical work. That split matters because it tells you what OpenAI expects Terra to do all day: act on screens, read images and produce clean drafts.
What the Chat exclusion means
Terra is not part of the general Chat experience. A subscriber who opens a normal conversation will not find it in the picker, which is the single most common misreading of the headlines. If you want Terra, you open Work or Codex. Research notes drawn from OpenAI’s help center also say the rollout does not change overall plan usage limits, though we could not verify that line against a second source, so treat it as reported rather than confirmed.
Why the Dates Do Not Line Up #
Readers should know about a date problem before they quote this story. The verified record dates OpenAI’s announcement to April 9, 2026. Yet the same record lists a knowledge cutoff of June 1, 2026, which falls nearly two months after that announcement date. A model cannot know about events that happen after its announcement unless its training data was extended or the announcement date refers to something other than this specific checkpoint.
Other reporting complicates things further. Our earlier coverage on this site, which cited GitHub’s changelog for Copilot, put Terra’s arrival in that product on October 26, 2026. A research pass for this article also returned October 26 as the date for the Work and Codex rollout note. Those two dates and the April 9 date do not reconcile, and we cannot settle the question from the materials in hand.
The practical takeaway is simple. Use April 9, 2026 only when you cite OpenAI’s announcement date, and use late October when you describe when the model reached Work and Codex or Copilot, labeling the source each time. If OpenAI publishes a correction or a clarifying note, that note should override anything printed here.
GPT-6.1 Terra Specs at a Glance #
The table below collects every spec OpenAI has confirmed. Treat it as a checklist for procurement and engineering reviews. Anything absent from the table is unconfirmed.
| Spec | Confirmed value | Source type |
|---|---|---|
| Model name | GPT-6.1 Terra | OpenAI announcement |
| API model ID | gpt-6.1-terra | OpenAI model documentation |
| Announcement date | April 9, 2026 | OpenAI announcement |
| ChatGPT surfaces | Work and Codex (not Chat) | OpenAI help materials |
| ChatGPT plans | Plus, Pro, Business, Enterprise, Edu | OpenAI help materials |
| Context window | 1.05M tokens | OpenAI model documentation |
| Maximum output | 128K tokens | OpenAI model documentation |
| Knowledge cutoff | June 1, 2026 | OpenAI model documentation |
| Reasoning effort | none, low, medium (default), high, xhigh, max | OpenAI model documentation |
| Codex requirement | Codex CLI 0.159.0 or later | OpenAI Codex documentation |
Two details stand out. The 1.05M-token window is large enough to hold a mid-sized codebase or a long stack of screenshots in one request. The 128K output ceiling means Terra can write long documents in a single pass. Neither number says anything about how well the model uses that space, and OpenAI has not published a long-context accuracy result we could verify.
GPT-6.1 Terra API Pricing: $1, $0.05 and $5 Per Million Tokens #
Standard pricing has three lines. Input costs $1 per million tokens, cached input costs $0.05 per million and output costs $5 per million. Two ratios drive every budget conversation. Cached input costs 95 percent less than fresh input, a 20x discount. Output costs five times as much as input, so a chatty model is more expensive than a terse one even at identical prompt sizes.
The table below turns those rates into dollars for five workloads. The arithmetic uses only the published rates. The workload sizes are our illustrations, not OpenAI benchmarks, and they ignore any surcharge, tool fee or Fast mode premium that may apply.
| Workload (illustrative) | Input tokens | Output tokens | Cost at standard rates |
|---|---|---|---|
| One screen-reading agent step | 20,000 fresh | 1,000 | $0.025 |
| Long Codex-style session | 50,000 fresh plus 450,000 cached | 30,000 | $0.2225 |
| Near-full context, no caching | 900,000 fresh | 128,000 | $1.54 |
| Near-full context, fully cached | 900,000 cached | 128,000 | $0.685 |
| 10,000 drafting requests | 20M fresh (2,000 each) | 8M (800 each) | $60.00 |
Compare the two near-full-context rows. Caching cuts the input bill from $0.90 to $0.045, yet the 128K output still costs $0.64, so the total only drops by about 55 percent. Output dominates once the input is cached. For high-volume drafting, the 10,000-request example shows that output tokens account for $40 of the $60 total. Teams that cap response length will save more than teams that merely shave prompts.
For plan-side costs, our ChatGPT vs Claude pricing breakdown walks through the subscription ladder, which sits on top of these per-token rates.
Reasoning Effort: Six Settings and a Medium Default #
Terra supports six reasoning-effort settings: none, low, medium, high, xhigh and max. Medium is the default. OpenAI’s guidance tells developers to start at medium and raise the setting only if the quality gain justifies the added time and tokens. That advice reads like a cost warning in friendly clothes. Higher effort buys more thinking, and thinking takes time and tokens.
The none setting deserves attention. According to OpenAI’s documentation, GPT-6 Astra and GPT-6.1 Sol do not support none, while GPT-6.1 Terra, GPT-6 Sol and GPT-6 Luna do. For latency-sensitive jobs like form filling or catalog matching, that gives Terra a path to near-instant responses that the two higher tiers lack. The documentation also lists Terra among the models whose Fast mode supports EU data residency, a detail that compliance teams will want to check against their own rules.
A minimal request looks like the sketch below. OpenAI’s guidance says to call models through the Responses API, and to migrate Chat Completions tool calls there. Confirm the exact parameter shape on the model page before shipping, because SDK details can shift between releases.
from openai import OpenAI
client = OpenAI()
response = client.responses.create(
model="gpt-6.1-terra",
reasoning={"effort": "medium"},
input="Match this product photo description to the closest catalog entry.",
)
print(response.output_text)
A sensible rollout runs your own task set at none, medium and high, records cost and accuracy for each, and picks the cheapest setting that clears your quality bar. That test costs a few dollars at Terra’s rates and saves far more at production volume.
Where Terra Sits in the GPT-6 Lineup #
OpenAI’s developer documentation lays out four current models, each with a one-line pitch. The pitches tell you how OpenAI wants customers to choose.
- GPT-6 Astra: the highest-intelligence option, aimed at the most demanding reasoning, coding and professional work.
- GPT-6.1 Sol: balanced speed, cost and intelligence, with near-Astra performance for complex work at lower cost.
- GPT-6.1 Terra: fast and efficient, for visual reasoning, precise computer use and clear, natural writing.
- GPT-6 Luna: the fastest and most cost-effective, for focused, high-volume tasks.
Terra therefore occupies the middle-to-lower price band with a specialty in screens and images. The documentation suggests uses such as matching product photos to catalog entries, comparing technical drawings, acting on application screens and drafting or revising text. It also names a predecessor. The migration notes tell developers moving from GPT-5.6 Terra to compare both models on representative images and workflows at the same reasoning effort. That wording confirms a Terra line existed before this release and that OpenAI expects regressions to be possible.
A short timeline of the 2026 lineup
Our AI coding assistants pilot guide cites GitHub’s changelog for how fast these models reached Copilot. Per that reporting, Copilot added GPT-6 Astra on September 4, GPT-6 Sol and GPT-6 Luna on September 22, GPT-6.1 Sol on September 29 and GPT-6.1 Terra on October 26, 2026. Seven models arrived inside seven weeks across vendors, and the pace explains why teams struggle to keep model choices stable for even a quarter.
Codex and Terra: What Developers Need to Know #
Codex users need one version check. Terra requires Codex CLI 0.159.0 or later, so run codex --version before you hunt for the model in a picker. Teams that pin the CLI in CI images should bump the pin first, or the model will not appear for automated runs.
codex --version
OpenAI’s help center, per our research, also says Codex Cloud can start and continue coding tasks from desktop, web or mobile, using reusable environments that hold repositories, tools and dependencies. Terra’s pitch suits the front end of that workflow: reading screenshots of failing UIs, acting on application screens and writing clear pull request descriptions. Heavier refactors may still belong to Astra or Sol, based on OpenAI’s own positioning.
Adoption data frames the stakes. The Stack Overflow 2026 Developer Survey reported 65.5 percent of respondents using Claude Code against 29.5 percent using OpenAI Codex, a gap we examined in our Codex vs Claude Code analysis. A cheaper, faster Codex model is one way OpenAI can narrow that gap without rebuilding the product. Whether it works depends on quality, which no verified benchmark has yet measured for Terra.
Competitive Comparison: Terra Against Claude, Gemini and DeepSeek #
Price is the one axis where Terra’s numbers are firm, so the table below compares list prices drawn from OpenAI’s materials and our earlier published coverage of rival models. Price says nothing about quality. A cheaper model that fails more often can cost more per finished task.
| Model | Input per 1M tokens | Output per 1M tokens | Terra’s price versus this model |
|---|---|---|---|
| GPT-6.1 Terra | $1.00 | $5.00 | Baseline |
| GPT-6 Astra | $10.00 | $50.00 | Terra costs 10 percent as much |
| Claude Opus 5.5 | $4.00 | $20.00 | Terra costs 25 percent as much |
| Gemini 4 Argon (through Nov. 30) | $2.00 | $10.00 | Terra costs 50 percent as much |
| GPT-6 Luna | $0.10 | $0.50 | Terra costs 10 times as much |
| Claude Haiku 5.5 (prompts to 100K) | $0.10 | $0.50 | Terra costs 10 times as much |
| DeepSeek V4.1-Flash (peak) | $0.30 | Not listed here | Terra costs about 3.3 times as much on input |
The pattern is clear. Terra undercuts every flagship in the table by at least half, yet it sits well above the small-model tier. Our DeepSeek vs Claude pricing comparison shows how aggressively DeepSeek and Anthropic’s Haiku tier have pushed the floor, so Terra has to win on capability for screens and images rather than on raw price.
Google’s side carries a clock. Per our Gemini 4 Argon pricing analysis, Argon’s current rates double on December 1 and its context limit is 1,048,576 tokens, close to Terra’s 1.05M. After the step-up, Terra’s price advantage over Argon should widen if nothing else changes. Anthropic’s Opus 5.5 competes on a different promise, which is depth on long agent runs rather than low cost.
Market Impact: Who Gains and Who Loses #
Start with the winners. Teams that run screen-driven automation gain the most, because Terra’s pitch targets exactly that job at a price that makes 20,000-token screen reads cost about two cents. Support operations, back-office data entry and catalog maintenance all involve repetitive visual work where per-step cost decides whether an agent pays for itself.
Prompt-heavy applications gain as well. The 20x cached-input discount rewards products that resend the same long instructions or documents, which describes most agent frameworks and many retrieval systems. An engineering lead who redesigns prompts to keep a stable cacheable prefix can cut input spend by up to 95 percent on those tokens. That optimization costs nothing in model quality.
Losers are harder to name. Rival mid-tier models face pricing pressure, and subscribers who assumed Terra would appear in everyday Chat face disappointment. OpenAI’s own higher tiers could also feel some cannibalization if customers find Terra good enough for work they used to send to Sol or Astra. That outcome is plausible, yet no sales data supports it.
For broader context, our AI and machine learning coverage tracks how these model releases ripple through pricing and tooling across vendors. The consistent finding is that cost per finished task, not headline intelligence, now decides most enterprise model choices.
What Remains Unconfirmed About GPT-6.1 Terra #
Several claims circulating around the launch lack support in OpenAI’s cited materials. Treat each as an open question until OpenAI or an independent tester provides evidence.
- The upgrade framing. OpenAI’s materials establish availability in Work and Codex. They do not establish that Terra upgrades ChatGPT as a whole.
- Benchmark results. No verified score appears in the official materials cited here. Any figure you see without a named, linked test setup deserves skepticism.
- Extra pricing. Only the three standard rates are confirmed. Launch promotions, Fast mode premiums and tool surcharges are not.
- Named statements. No on-record quote from an OpenAI employee appears in the cited materials.
- The announcement date. April 9, 2026 conflicts with a June 1, 2026 knowledge cutoff and with late-October rollout dates reported elsewhere.
- Plan usage limits. Reports say the rollout leaves limits unchanged, but we did not find a second source.
Good reporting here means saying unconfirmed out loud. Readers who build budgets or contracts on this model should pin every assumption to a document OpenAI controls, and re-check that document before each renewal.
Five Predictions for GPT-6.1 Terra #
The following forecasts are our own analysis. They rest on the confirmed facts above and on the pace of releases through 2026, and none comes from OpenAI.
- Screen-driven agents move to fast tiers first. With a 20,000-token screen read costing roughly two cents, expect pilots in support and back-office automation to default to Terra-class models before they touch flagship tiers.
- Cached-input pricing becomes a headline spec. A 20x discount changes budgets more than a modest drop in base rates, so rivals will likely answer with their own caching discounts rather than cutting list prices.
- Reasoning effort turns into a routing decision. Teams will build routers that send easy tasks to none or low and reserve high or max for hard ones, because the setting directly controls time and tokens.
- Terra’s footprint widens beyond Work and Codex. OpenAI has shipped the model to the API and two ChatGPT surfaces, and wider placement would follow the pattern of earlier models. This remains a forecast, not an announcement.
- Lineup churn continues. Seven models in seven weeks across vendors suggests another GPT-6.x release before spring 2027, so avoid long contracts that lock you to a single model ID.
If any forecast fails, the likeliest culprit is quality. Cheap, fast models only win when they pass the same acceptance tests as expensive ones.
FAQ: GPT-6.1 Terra in ChatGPT Work and Codex #
What is GPT-6.1 Terra? It is an OpenAI model that the company describes as fast and efficient, built for visual reasoning, precise computer use and clear, natural writing. The API ID is gpt-6.1-terra.
Where can I use it? Terra is available through the OpenAI API, in ChatGPT Work and in Codex. It is not part of the general Chat experience.
Which ChatGPT plans include it? OpenAI lists Plus, Pro, Business, Enterprise and Edu. Access depends on your plan, region and workspace settings.
How much does the API cost? Standard rates are $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. OpenAI has not confirmed other launch pricing in the materials we reviewed.
How big is the context window? The context window is 1.05M tokens, and maximum output is 128K tokens. The listed knowledge cutoff is June 1, 2026.
What do I need to use Terra in Codex? You need Codex CLI 0.159.0 or later. Run codex --version to check your install.
Which reasoning effort should I pick? OpenAI’s documentation recommends starting at medium, the default, and raising it only when the quality gain justifies the added time and tokens. Terra also supports none, low, high, xhigh and max.
Is this really an upgrade to ChatGPT and Codex? OpenAI’s materials confirm that Terra is available in Work and Codex. They do not establish the word upgrade, so we treat that wording as unconfirmed.