GPT-5.6 Terra

OpenAI discloses no parameter count, layer count, attention type or training scale for Terra, so those fields are null. The public facts are operational: a 1,050,000-token window with a 922,000-token input cap and 128,000-token output cap, text and image input, text output, and reasoning tokens that bill as output. Reasoning items can be reused across Sol, Terra and Luna but not with GPT-5.5, which suggests the three models share a reasoning format. The GPT-5.6 system card calls Terra "a capable lower-cost option" and says Sol and Terra are served with newly added activation classifiers that monitor internal activations during generation. Images are processed at original dimensions up to 65,535 pixels per side, within a 30,000-patch limit.

All Artificial Analysis (AA) figures below are at max effort unless stated; AA tests each effort level separately, and the default medium setting scores lower. Coding (9.0): 88.0% on Terminal-Bench 2.1 (72.3% at medium), ahead of GPT-5.5 at xhigh (84.3%) and Claude Sonnet 5 at max (80.5%) and level with Gemini 3.8 Flash at high (87.6%); SciCode 55.0%; LMArena coding 1518 at xhigh, level with Sonnet 5 at high (1520). Reasoning (9.0): GPQA Diamond 92.5% and Humanity's Last Exam 42.9% (87.2% and 33.3% at medium), in line with GPT-5.4 at xhigh (92.0% and 43.7%). Math (8.6) has no verified competition-math score and is set from the reasoning results. Agentic (9.0): τ²-Bench Telecom 86.3% (72.8% at medium), below GPT-5.5 at xhigh (93.9%), plus the widest hosted tool list in its price band: computer use, hosted shell, apply_patch, code interpreter, MCP, tool search, skills and programmatic tool calling. OpenAI's system card adds a caution: on its destructive-actions test Terra avoids overwriting user data at 0.81, against 0.83 for Sol and 0.88 for GPT-5.5, and scores 0.37 on avoidance plus correctness, against 0.44 for both. Long context (8.6): 83% on AA-LCR (74% at medium) with 922K usable input. Vision (8.5): MMMU-Pro 80.7%, above GPT-5.5 at xhigh (79.9%) and Sonnet 5 at max (77.3%) but below Gemini 3.8 Flash at high (85.6%); 1267 on LMArena's vision board. Document/OCR (8.2) benefits from original-resolution image input. Instruction following (8.5): IFBench 71.2% (62.2% at medium), below GPT-5.5 at xhigh (75.9%) and GPT-5.4 mini at xhigh (73.3%). Function calling (9.0): OpenAI's GPT-Red testing gives a 0.061% direct prompt-injection success rate and 0.946 on search and function-calling injection robustness, the best in the GPT-5.6 family. Safety calibration (8.0): refusals are well calibrated in OpenAI's testing, but AA-Omniscience records an 88% hallucination rate (a wrong answer rather than an abstention on questions it cannot answer correctly), against 39% for Sonnet 5 at max and 55% for Gemini 3.8 Flash at high. Real-time data (6.8) comes from the web search tool. On OpenAI's length-adjusted HealthBench Professional, Terra scores 57.7, against 51.8 for GPT-5.5 and 60.5 for Sol (vendor-reported).

Artificial Analysis figures were read on 2026-09-21 and are all independent measurements; the front-matter stores the max-effort results from the page cited. Each competitor is shown at the effort level named. The arena entries test Terra at xhigh, and Gemini 3.8 Flash's arena ratings rest on far fewer votes (5,076 text and 1,434 coding, against 28,119 and 7,783 for Terra), so its lead carries wider error bars. MMMU-Pro, IFBench and AA-LCR are different tests from the template's MMMU, IFEval and MRCR keys, which stay null, as do MMLU-Pro, AIME 2025, SWE-bench Verified, LiveCodeBench and Aider Polyglot.

Artificial Analysis measures Terra on OpenAI's own API at 91.8 output tokens/sec with 2.0 s to the first token at the default medium effort (the front-matter values). At high it measures 89.4 tokens/sec and 3.2 s, at xhigh 98.4 tokens/sec and 18.4 s, and at max 106.5 tokens/sec with about 224 s before the first token, because hidden reasoning runs first. That is slower per token than Gemini 3.8 Flash at high (329 tokens/sec) and DeepSeek V4.1-Flash at max (226), faster than Claude Sonnet 5 at high (71), and about half the 176 tokens/sec of GPT-5.4 mini at medium. OpenAI states that Fast mode runs up to 2.5x faster than Standard at twice the price, including for prompts over 272K tokens. The GPT-5.6 safety classifiers can also pause a stream for several seconds on sensitive cyber or biology content. latency_tier is medium at the default effort; treat xhigh and max as batch settings.

What it costs in practice: 10M input and 2M output tokens a day, uncached, is about $44 on Terra, against $16.50 on GPT-5.4 mini, $55 on GPT-5.4, $40 on Claude Sonnet 5, $15 on Gemini 3.8 Flash at its introductory rate through 2026-12-31 ($30 from 2027-01-01) and $5.40 on DeepSeek V4.1-Flash at peak rates. Effort matters as much as list price, because reasoning tokens bill as output. Artificial Analysis spent $447 running its Intelligence Index on Terra at medium (score 30), $1,187 at xhigh (38) and $2,501 at max (42). The same run cost $1,438 on GPT-5.4 mini at xhigh (24), $5,294 on GPT-5.5 at xhigh (38), $6,998 on Claude Sonnet 5 at max (38), $1,623 on Gemini 3.8 Flash at high (41, introductory price) and $477 on DeepSeek V4.1-Flash at max (39.5). On Amazon Bedrock, global cross-Region inference matches OpenAI's rates, and in-Region or geographic inference adds 10% ($2.20/$13.20).

Terra is API-only and proprietary: no open weights, no self-hosting, no fine-tuning. On OpenAI it is served through the Responses API, Chat Completions and Batch; Realtime, Assistants and fine-tuning endpoints are not supported. Only one snapshot exists, gpt-5.6-terra, so there is no dated version to pin. Microsoft made the GPT-5.6 family GA in Microsoft Foundry on 2026-07-09 in Global Standard and Data Zones Standard deployments, including a new Asia-Pacific Data Zone. Amazon Bedrock made Terra GA on 2026-07-13 in US East (N. Virginia), US East (Ohio) and US West (Oregon) through cross-Region inference profiles (us.openai.gpt-5.6-terra, global.openai.gpt-5.6-terra), and in AWS GovCloud (US-West and US-East) on 2026-08-24. Bedrock lists the model as EOL no sooner than 2027-07-13 and does not offer Priority or Flex tiers for it. OpenAI's own regional processing covers the United States and Europe (EEA + Switzerland) for Terra; EU processing requires Zero Data Retention, Modified Abuse Monitoring, Eyes Off or Safety Retention. OpenRouter routes Terra to OpenAI, Azure and Bedrock.

Under OpenAI's Preparedness Framework, OpenAI treats Terra as High capability in both Cybersecurity and Biological and Chemical risk, below Critical, and deploys safeguards matched to that level. Sol and Terra carry newly added activation classifiers that can pause streaming while a second check decides whether to block the output; a two-tier topical classifier and safety reasoner covers all three GPT-5.6 models. The system card reports that GPT-5.6 models show a greater tendency than GPT-5.5 to go beyond the user's intent in agentic coding, although absolute rates stay low. It also reports strong refusal on harmful cyber requests from production data (0.987 not-unsafe) and a low over-refusal rate on benign advanced biology (0.978 not-overrefuse). API data is not used for training unless you opt in; abuse-monitoring logs are kept for up to 30 days, and Chat Completions and Responses are eligible for Zero Data Retention. OpenAI offers BAA coverage for eligible configurations, though web search with live internet access is not HIPAA-eligible. compliance is left null because no certification list is recorded for this page; check OpenAI's trust portal for SOC 2 and ISO status. Send a stable safety_identifier for end-user traffic, as OpenAI recommends.

Official OpenAI SDKs cover Python, TypeScript, Java, Go and .NET, and the OpenAI Agents SDK supports the Responses tool surface directly. Microsoft positions GPT-5.6 in Foundry Agent Service with the Microsoft Agent Framework, and standard OpenAI-compatible integrations such as LangChain, LlamaIndex and Vercel AI SDK work with the model ID. On OpenRouter, the apps sending the most Terra traffic as of 2026-09-21 are Framer (868B tokens), Hermes Agent (114B) and Codex (108B). OpenAI's models page recommends Terra for workloads that "balance intelligence and cost", and its deprecation notices route a long list of retiring models to it, so adoption is set to grow through the 2026 shutdowns. Popularity tier: mainstream.