ModelsArchitectures & capability
OpenAI launches GPT-6.1 Terra, a $1/$5-per-million-token mid-tier model for computer use and vision
OpenAI's GPT-6.1 Terra sits between Luna and Sol, targets computer use, vision and technical writing, and costs $1/$5 per million input/output tokens. Testing shows sharply lower hallucination and per-task cost, while one report assigns it a Critical cybersecurity rating.
From the Simon Willison article of October 26, 2026 and cited sources: Terra is priced at $1 per million input tokens and $5 per million output tokens, down from $2/$12 for GPT-5.6 Terra. [7] [1] [8] [9]
SOURCES[7] Artificial Analysis: GPT-6.1 Terra pushes the cost… artificialanalysis.ai[1] Simon Willison: GPT-6.1 Terra simonwillison.net[8] Hacker News discussion news.ycombinator.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Terra is priced at $1 per million input tokens and $5 per million output tokens, down from $2/$12 for GPT-5.6 Terra. [7] [1] [8] [9]
- Original source
- Simon Willison article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 26, 2026 article published
- Covered
- Oct 27 edition
- Evidence
- 10 sources across 10 sites
Reach: not measured yet
Terra is priced at $1 per million input tokens and $5 per million output tokens, down from $2/$12 for GPT-5.6 Terra. It sits between Luna and the more capable Sol and Astra. An OpenAI researcher described it as better than Sol and Astra at technical communication, similar on vision and computer use but faster, and somewhat weaker on most other tasks. It is available in the API and in ChatGPT Work and Codex. [7] [1] [8] [9]
In Artificial Analysis's independent testing, Terra scored 48 on the Coding Agent Index, compared with 54 for Sol and 31 for Luna. Its hallucination rate on AA-Omniscience fell from 88% to 38%, and cost per task dropped to about $0.40 from about $1.40. The same testing found that maximum reasoning effort falls below the cost-efficiency frontier set by the model's own lower settings. [7]
One outlet reports that Terra carries a Critical cybersecurity rating under OpenAI's Preparedness Framework, the level GPT-6 Astra reached first. The outlet cites 97.1% on ExploitBench and a 29.6% prompt-injection bypass rate when no system-level controls are in place. Critical means the model can autonomously exploit hardened systems or carry out new end-to-end attacks. [2] [5] [4]
cost per task fell about 70% while coding-agent scores stayed close to Sol's. Implication: high-volume computer-use agents become cheaper to run. If the cyber rating holds, sandboxing and prompt-injection defenses matter more at this price point.
Executive brief
OpenAI's new mid-tier model, GPT-6.1 Terra, is reported to carry a "Critical" cybersecurity rating under OpenAI's Preparedness Framework. That is the same top rating GPT-6 Astra received when it became the first model to reach that level. Terra also costs $1/$5 per million input/output tokens, and OpenAI says it matches Astra on one agent benchmark at about a tenth of the per-token price (Cybersecurity News; Interesting Engineering). OpenAI is pitching it at computer use, visual reasoning and workplace writing. Independent testing by Artificial Analysis found a large drop in hallucination rate and strong cost efficiency.
What changed and event timeline
GPT-6 Astra ships as a "Critical" cyber model
OpenAI rated Astra Critical for cybersecurity, a first. It paused frontier training runs to build new safeguards before release ().
Astra's system card details exploit chains
OpenAI's system card reports that Astra built a browser exploit chain in 29 hours and showed more control over its own chain-of-thought than earlier models ().
OpenAI launches GPT-6.1 Terra
OpenAI's launch post frames it as "fast and capable computer use at a lower price" (). Interesting Engineering gives 27 October as the date ().
Artificial Analysis publishes independent scores
Terra comes in 4 points below GPT-6.1 Sol on the Intelligence Index, and its hallucination rate falls from 88% to 38% ().
An OpenAI researcher explains where Terra fits
On Hacker News, OpenAI's Ted Sanders called Terra better than Astra and Sol at technical communication, similar on vision and computer use but faster, and "a bit worse on most other things" (;).
Capabilities and access
- Model: GPT-6.1 Terra. API ID
gpt-6.1-terra. Available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users (BleepingComputer). - Pricing: $1 per million input tokens, $5 per million output, $0.05 per million cached input, $1.25 per million for cache writes.
- Target tasks: filling out forms, video editing, reading technical drawings, desktop automation (Basic Tutorials).
Read the full section
- Model: GPT-6.1 Terra. API ID
gpt-6.1-terra. Available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users (BleepingComputer). - Pricing: $1 per million input tokens, $5 per million output, $0.05 per million cached input, $1.25 per million for cache writes. GPT-5.6 Terra cost $2/$12 (Artificial Analysis).
- Target tasks: filling out forms, video editing, reading technical drawings, desktop automation (Basic Tutorials).
- Lineup position: Luna (cheapest), then Terra, then Sol and Astra (most capable) (Simon Willison).
Technical analysis for researchers and developers
- No architecture details: none are documented. Sanders said only that later-trained models may use new data, techniques or architectures (Hacker News).
- Reasoning effort matters a lot: results depend heavily on the effort setting.
- Max effort is a weak deal: Artificial Analysis found that at maximum effort the model falls below the cost-efficiency frontier set by its own lower settings (Artificial Analysis).
Read the full section
- No architecture details: none are documented. Sanders said only that later-trained models may use new data, techniques or architectures (Hacker News).
- Reasoning effort matters a lot: results depend heavily on the effort setting. Terminal-Bench 4.0 scores of 50.3% are at maximum effort, and error rates of 16.7% (low) and 4.3% (max) are reported per setting (BleepingComputer).
- Max effort is a weak deal: Artificial Analysis found that at maximum effort the model falls below the cost-efficiency frontier set by its own lower settings (Artificial Analysis).
- Cost of a heavy prompt: Willison's "pelican riding a bicycle" drawing test at max effort cost 11.3 cents and took 4 minutes 11 seconds (Simon Willison).
Claims and evidence
- Speed: more than 3× faster than Astra and Sol. Vendor-reported (BleepingComputer).
- Matches Astra at a tenth of the cost: matches Astra on Agents' Last Exam at about 1/10 the token cost, and runs tasks at about 9× lower cost than Claude Opus 5.5.
- Coding Agent Index: 48, against 54 for Sol and 31 for Luna.
Read the full section
- Speed: more than 3× faster than Astra and Sol. Vendor-reported (BleepingComputer).
- Matches Astra at a tenth of the cost: matches Astra on Agents' Last Exam at about 1/10 the token cost, and runs tasks at about 9× lower cost than Claude Opus 5.5. Vendor-reported (Interesting Engineering).
- Coding Agent Index: 48, against 54 for Sol and 31 for Luna. Independent (Artificial Analysis).
- Cost per task: about $0.40, against about $1.40 for GPT-5.6 Terra. Independent (Artificial Analysis).
- Size of the price cut is disputed: OpenAI says 50% (BleepingComputer). The prior $12 output price means output fell about 58%, and a Hacker News commenter claimed 75% (Hacker News).
Context and prior work
- Previous generation: GPT-5.6 came in three tiers (Sol, Terra, Luna), all rated "High" for cybersecurity.
- What "Critical" means: autonomously finding and exploiting zero-day vulnerabilities across hardened systems, or carrying out new end-to-end attack strategies (InfoQ).
- Competitor comparison: Willison compares Terra's spot in the lineup to Anthropic's Sonnet (Simon Willison).
Read the full section
- Previous generation: GPT-5.6 came in three tiers (Sol, Terra, Luna), all rated "High" for cybersecurity. GPT-5.6 Terra could not build end-to-end exploit chains against hardened targets on its own (Codersera).
- What "Critical" means: autonomously finding and exploiting zero-day vulnerabilities across hardened systems, or carrying out new end-to-end attack strategies (InfoQ).
- Competitor comparison: Willison compares Terra's spot in the lineup to Anthropic's Sonnet (Simon Willison).
Limitations, safety and contested findings
- Cyber rating: among the reviewed sources, only one outlet reports the Critical rating.
- OpenAI's own caveats: OpenAI reportedly says exposure to historical vulnerabilities in training may inflate these scores, and that reliable exploitation remains difficult (Cybersecurity News).
- Safeguards: Terra reportedly uses the same safeguards as Sol and Astra, with detection systems updated for the new model (Cybersecurity News).
Read the full section
- Cyber rating: among the reviewed sources, only one outlet reports the Critical rating. It cites 97.1% on ExploitBench, 24.5% arbitrary code execution on recent vulnerabilities, and a 29.6% prompt-injection bypass rate without system-level controls (Cybersecurity News).
- OpenAI's own caveats: OpenAI reportedly says exposure to historical vulnerabilities in training may inflate these scores, and that reliable exploitation remains difficult (Cybersecurity News).
- Safeguards: Terra reportedly uses the same safeguards as Sol and Astra, with detection systems updated for the new model (Cybersecurity News).
- Hallucinations: a 38% hallucination rate on AA-Omniscience is a large improvement but still high (Artificial Analysis).
- Naming and choice confusion: Hacker News users criticized the naming scheme and the staggered, one-model-at-a-time releases (Hacker News).
Business and practitioner implications
- Cheaper agents: Terra is a natural default for high-volume computer-use and vision agents where Sol's price is hard to justify (Simon Willison).
- Pick effort deliberately: benchmark at low and medium effort first, since max effort gives the worst value (Artificial Analysis).
- Higher-risk tooling, lower price: a reportedly Critical-rated cyber model at mid-tier prices raises the stakes for agent sandboxing and prompt-injection defenses, especially given the reported 29.6% bypass rate without controls (Cybersecurity News).
Read the full section
- Cheaper agents: Terra is a natural default for high-volume computer-use and vision agents where Sol's price is hard to justify (Simon Willison).
- Pick effort deliberately: benchmark at low and medium effort first, since max effort gives the worst value (Artificial Analysis).
- Higher-risk tooling, lower price: a reportedly Critical-rated cyber model at mid-tier prices raises the stakes for agent sandboxing and prompt-injection defenses, especially given the reported 29.6% bypass rate without controls (Cybersecurity News).
- ChatGPT subscribers: Sanders calls it "not a bombshell release." The main gains are going further before hitting usage limits and faster computer use (Hacker News).