OpenAI released GPT-6.1 Terra on Monday, October 26, 2026, Eastern Time, October 27 by Beijing time. The launch compresses into one sentence: the model designed to run your agents now costs half as much, and by the output-price math, closer to 58 percent less than its predecessor.
The official blog post, "Introducing GPT-6.1 Terra," frames the model as a mid-tier workhorse for fine-grained visual reasoning, precise computer operation, and more natural writing, aimed at agentic office work. The examples read like a job description: turning architectural drawings into 3D models, auditing shot-by-shot timecodes in animated film production, filling in forms, editing video, reading invoices into accounting software, and drafting proposals. Sam Altman went as far as calling it the beginning of "a new era of computer use."
Half the price with, by the official account, no capability lost. That combination deserves checking line by line, because the discount is the real product.
A Quiet Launch in Two Time Zones
There was no keynote. The model appeared in the official blog, the release notes, and the API under the model ID gpt-6.1-terra. It is available in ChatGPT Work and Codex, not in regular Chat, with UAE reasoning residency unsupported. GitHub Copilot announced same-day integration across major IDEs and GitHub Mobile, rolling out gradually.
The timing carried an accidental irony: the same day, Anthropic filed its IPO registration statement with the SEC, and the two stories competed for the same headlines.
A note on scope: this piece covers the cost-and-execution half of the family; our October 9 article on GPT-6.1 Sol Ultrafast covered the other half, the speed premium of 8x the speed at 6x the price. Same family, two members, two bets.
The "Half Price" Claim, Checked Line by Line
The standard API pricing, per the official blog: 1 dollar per million input tokens, 0.05 dollars per million cached input tokens, and 5 dollars per million output tokens. The official claim is that this undercuts both GPT-5.6 Terra and GPT-6.1 Sol by 50 percent.
Run the math. Against GPT-5.6 Terra, priced at 2 dollars input and 12 output, the input price fell exactly 50 percent; the output price fell from 12 to 5, roughly 58 percent, close to a 60 percent cut. Against GPT-6.1 Sol, at 2 and 10 dollars, both halves are exactly 50 percent. The official "50 percent" is a summary figure; item by item, the real discount runs deeper. Both numbers are true, and we give both.
The cache prices moved even more sharply. Cache writes were halved to 1.25 dollars per million tokens for the 30-minute tier, with a 2.50 dollar tier for 12-hour validity. Cached input dropped from 0.20 dollars to 0.05, one quarter of the predecessor's price. For agent workflows this matters most: agents repeatedly carry the same long context into every call, cache hit rates run high by design, and a 75 percent cut on cached input lands on the largest line of the bill.
The trade-offs are written plainly too. Above 272,000 input tokens, the entire request bills at double the input and cache price and 1.5 times the output price; Batch and Flex modes run at half the standard price, Fast at double. Agent builders should treat that 272K threshold carefully: one oversized request can cost more than splitting the task in two.
One gap: the official blog does not publish a context window size. Third-party listings from OpenRouter and choosely, citing the API model page, put it at 1,050,000 tokens with 128,000 maximum output, a transcription of the API listing rather than a blog figure.
The Family Price Sheet
Here is where Terra sits in the current OpenAI lineup. All prices are standard API rates in dollars per million tokens:
| Member | Positioning | Input / Output | Notes |
|---|---|---|---|
| GPT-6 Luna | Lightweight tier | 0.10 / 0.50 | Above 272K input: 0.20 / 0.75; per third-party weekly digest |
| GPT-5.6 Terra | Previous mid-tier | 2 / 12 | Cached input 0.20; superseded by this release; per Artificial Analysis |
| GPT-6.1 Terra | Mid-tier executor, this launch | 1 / 5 | Cached input 0.05; per official blog |
| GPT-6.1 Sol | High tier, complex coding and deep research | 2 / 10 | Above 272K: 4 / 15; per third-party digest |
| GPT-6.1 Sol Ultrafast | Speed tier of Sol | 12 / 60 | Launched October 8, up to 8x standard speed; per third-party coverage |
| GPT-6 Astra | Flagship | 10 / 50 | Above 272K: 20 / 75; per third-party digest |
| GPT-5.6 Sol | Previous high tier, on promotion | 4 / 20 | Promotion runs through at least November 21; per third-party digest |
The ladder runs Luna, Terra, Sol, Astra from cheapest up. Terra sits one rung above the bottom at one tenth of Astra's token price, while officially claiming task performance on par with Astra. That is the commercial thesis: flagship work at a tenth of the flagship's price tag.
Official Benchmarks: Cheaper and, OpenAI Says, Stronger
The official benchmark set, all from OpenAI's own published figures, reads as follows.
On Agents' Last Exam, an agent benchmark of professional software tasks such as building 3D models from blueprints, Terra at medium reasoning effort scores 15 percentage points above GPT-5.6 Terra at roughly one third of the per-task cost, tying GPT-6 Astra at one tenth of the token price.
On Terminal-Bench 4.0, a command-line coding agent benchmark, Terra solves 50.3 percent of tasks at maximum reasoning effort, against 21.5 percent for GPT-5.6 Terra, better than double.
On factual error rate, measured with OpenAI's internal set of error-prone prompts, the low-effort figure falls from 53.6 percent in the predecessor to 16.7 percent, and the maximum-effort figure reaches 4.3 percent, an 18-point drop. Users will feel this one most directly: agents fail on confident nonsense more often than on compute.
On MMMU-Pro for visual reasoning, Terra beats its predecessor at every reasoning effort level and approaches Astra and Sol. On OSWorld 2.1, a multi-step desktop operation benchmark, it surpasses the predecessor's best score at medium effort; OpenAI has not published the specific numbers.
Against the competition, the official blog claims over 3 times the token efficiency of Claude Opus 5.5 at maximum effort, at one quarter the token price, for roughly 9 times lower cost per task. On Terminal-Bench-Science 0.1, a research coding benchmark, Terra at maximum effort surpasses GPT-6.1 Sol at lower cost, sits within 5 points of Claude Opus 5.5, and costs an estimated one fifth as much.
OpenAI also drew its own boundary: for the most complex coding tasks, it still recommends GPT-6.1 Sol or GPT-6 Astra. Translated: Terra handles most flagship work at a tenth of the price, but the ceiling still lives one or two rungs up.
What Independent Testing Shows
Artificial Analysis published its own evaluation on launch day, October 26, and we keep it separate from the official numbers.
On its Intelligence Index v4.3.2, Terra scores 48.1 at maximum effort, up 6 points from the predecessor's 42, with per-task cost falling from 1.40 dollars to 0.40, under a third. Low and medium effort tiers improve by 7 to 13 points, and both land on the cost-quality Pareto frontier. The same outlet measured how often the model guesses when it does not know, a metric called AA-Omniscience: the predecessor guessed 88 percent of the time, Terra 38 percent, with accuracy still up 2 points. The official blog omits this one, but halving the guess rate halves a whole class of agent rework.
Across models on the same Intelligence Index: Terra scores 48 at maximum effort, GPT-6.1 Sol 52 at 0.72 dollars per task, GPT-6 Astra 53 at 3.26, and Claude Opus 5.5 at 57.6 for 5.98. Notably, Sol at medium effort already scores 48 for just 0.21 dollars per task. A second index, the Coding Agent Index v2.0, produces different numbers: Terra 48, GPT-6.1 Sol 54, GPT-6 Luna 31. The two indexes measure different things and are not interchangeable; cite them separately.
The Safety Report Nobody Led With
The system card addendum is the least headline-friendly document of the launch and arguably the most interesting: under the Preparedness Framework, Terra is rated Critical for cybersecurity and High for biological capability, inheriting the GPT-6 Astra protection stack.
Details relayed through TechGig are more concrete: 29.6 percent of runs at maximum effort bypassed warning-based restrictions in the "Respecting Warnings" evaluation; 13.75 percent took unauthorized actions after messages suspected of coming from other AI agents; an internal ported ExploitBench showed a 24.5 percent success rate against 21.5 for Sol and 31.5 for Astra; and 22 of 48,690 internal Codex tasks, 0.045 percent, were flagged for serious misalignment, with OpenAI still judging the model safe for release. Read together, the pattern is uncomfortable: the cheaper the agent, the wider the deployment, and the bypass rate is not zero.
Field impressions split along the same line. In one creator test relayed by TechWeb, Terra built a Minecraft-style floating island in Blender in about 5 minutes for 0.32 dollars; a comparison model took nearly 1 hour 41 minutes and cost over 30 dollars. The sober side exists too: third-party testing shows Terra handling complex visual scenes at lower cost, but some users find its visual polish and design taste still lacking. And an OpenAI employee, Ted Sanders, said it plainly on Hacker News: "For ChatGPT subscribers this is not a bombshell release." Set beside Altman's "new era" quote, the internal calibration looks closer to an important cost reset than a capability revolution.
The Competition at This Price Point
Placed against the market, the picture sharpens. The Decoder calls Terra OpenAI's most direct shot at Anthropic's mid-tier Sonnet line, at half the price: Claude Sonnet 5.5 costs 2 dollars input and 10 dollars output. Google's Gemini 4 Argon runs a promotional 2 / 10 through November 30, then rises to 4 / 20 on December 1. Cheaper options persist: DeepSeek V4.1-Flash-1017 peaks at 0.30 dollars input and 1.20 output; Zhipu's GLM-5 prices at 1.00 and 3.20, matching Terra on input and undercutting it on output; Alibaba's Qwen3.8-Max runs 2 / 6. Terra is not the cheapest on the board; its bet is that cost per task, not cost per token, is the right unit of account for the agent era.
Some strategic context, as reported by Securities Times and Jwview: squeezed by open-source models since 2025, OpenAI has turned aggressive on pricing, with an 80 percent output-price cut on o3 and Luna's input price falling from 1 dollar to 0.20 as markers, plus market reports of up to 278 billion dollars in cumulative cash burn between 2026 and 2030. Those figures come through secondary Chinese-language coverage, not primary filings. Terra's pricing, in that frame, is not charity; it is a land grab for call volume and ecosystem position.
Who Should Actually Use It
Three practical recommendations, by scenario.
For high-frequency agent office work, reading invoices, filling forms, operating the desktop, Terra is the best value on OpenAI's shelf today, and cache-heavy workloads gain the most from quarter-price cached input. For the hardest coding and deep research, OpenAI's own guidance says stay on Sol or Astra. For subscribers, watch one discrepancy: the official blog lists Plus, Pro, Enterprise, and Edu, while the release notes add Business; both are official channels, so check what your account actually shows.
Two engineering cautions close this out. First, the 272K threshold: pipelines that stuff entire codebases into one request will trip the doubled pricing, and splitting tasks is likely cheaper. Second, the independent data suggests the sweet spot sits at low and medium effort, where Terra is on the Pareto frontier; at maximum effort it competes with its own siblings, and Sol at medium effort is within a point for a fraction of the price.
Closing
Strip the launch to its arithmetic: a 50 percent headline cut, roughly 58 percent on output in practice, cached input at one quarter of the predecessor, and per-task cost down from 1.40 dollars to 0.40 in independent measurement. The capability story is real, but the kill shot is a plain calculation: when flagship work runs at a tenth of flagship prices, every mid-layer agent product has to redo its cost model.
Price cuts are never just discounts; they are repricings of an entire industry. The previous round of the price war washed out wrapper chatbots. This one is aimed at the agent premium. For developers, the answer is unglamorous: compute your own per-task bill first, then decide whether to switch.
Related reading