Terra replaces the previous flagship for most production workloads, trading Sol's ultra multi-agent mode and pro reasoning ceiling for a large context window and explicit prompt caching. It's built for engineering teams running coding agents at scale, not researchers chasing the hardest 5% of reasoning tasks.
Terra sits in the middle of OpenAI's model lineup, scoring 72.3% on SWE-bench Verified and 68.7% on GPQA Diamond as of its July 2026 release. It matches the prior flagship's performance without Sol's ultra mode or programmatic tool calling, aimed at everyday coding and knowledge work.
Provider: OpenAI · Family: GPT-5.6
Context window: 200,000 tokens · Max output: 64,000
Input modalities: text, image · Output: text, tool-calls, code
About GPT-5.6 Terra
GPT-5.6 Terra is the balanced tier of OpenAI's GPT-5.6 family, launched July 9, 2026 for general availability. It delivers performance competitive with the prior flagship, at roughly half that model's cost. Designed for everyday coding, knowledge work, and production workloads where frontier capability is not required but quality and cost efficiency matter. Terra shares the core GPT-5.6 architecture: strong reasoning, token efficiency improvements over the previous generation, explicit prompt caching with cache breakpoints (30-min TTL), persisted reasoning across turns, and support for reasoning effort levels none through xhigh (pro mode and max effort reserved for Sol). It does not include ultra multi-agent mode or programmatic tool calling; those stay exclusive to Sol. The model supports a 200K context window with 64K max output tokens. Native modalities: text and image input; text, tool-calls, and code output. Vision capabilities match Sol. There's no native audio or video I/O. Terra is reachable through the OpenAI API (Responses API and Chat Completions), ChatGPT, Codex, and Microsoft Copilot, plus other model gateways. Knowledge cutoff is June 2026. See the pricing FAQ for exact per-token rates. Enterprise buyers get Zero Data Retention eligibility, US and EU data residency, plus the standard compliance certifications listed under governance.
Pricing
Input $2.50/M, Output $15.00/M, Cached input $0.3125/M (1.25x base). 50% of Sol pricing. No batch discount at launch. No ultra mode, no PTC, no pro mode. Available on OpenAI API, ChatGPT, Codex, Microsoft 365 Copilot, OpenRouter, Vercel, Cloudflare, Snowflake, Databricks Mosaic.
Key Features
- Flagship-Class Performance at Half Cost: Matches the prior generation's flagship on SWE-bench, GPQA, and MMLU-Pro within 2-3 points while costing exactly half per token.
- Explicit Prompt Caching: Cache breakpoints with a 30-min TTL and writes at 1.25x the base rate, giving predictable costs for repeated-context workloads.
- Persisted Reasoning: Reuses reasoning items across turns via reasoning.context, improving multi-turn quality and cache efficiency.
- Full Reasoning Effort Range: Supports none, low, medium, high, and xhigh so compute matches task complexity without a pro-mode ceiling.
- Broad Distribution: Reachable through the OpenAI API, ChatGPT, Codex, Vercel, Cloudflare, Snowflake, and Databricks, plus Microsoft 365 Copilot.
Pros
- Best price-to-performance in the family: matches the prior flagship at exactly half its per-token cost.
- Explicit caching and persisted reasoning optimize multi-turn agent economics, which matters for production fleets.
- Faster than Sol (85 vs 75 tok/sec) with lower latency (1000ms vs 1200ms p50), better for latency-sensitive paths.
- Broad availability across the API, ChatGPT, Codex, Copilot, and gateway partners means no capacity waitlists.
- Strong safety parity with Sol, including high CBRN and cyber marks with below-high self-improvement, keeps it enterprise compliant.
Cons
- No ultra multi-agent mode, so it can't parallelize complex decomposable tasks the way Sol reduces wall-clock time on those.
- No programmatic tool calling, so tool-heavy loops need standard turn-based calling, costing more tokens and turns.
- No pro reasoning mode, so the xhigh ceiling limits quality on the hardest reasoning, where Sol's pro and max modes win on novel algorithms.
- Context tops out well below Sol's largest window; long-document workloads still need Sol or explicit caching strategy.
- Vision only for multimodal input, with no audio or video; Sol and Gemini 3.1 lead on multimodal breadth.
Benchmarks
- math: 82.3
- mmlu: 91.4
- mmlu pro: 81.2
- aime 2025: 84.1
- arc agi 2: 25.8
- humaneval: 94.5
- live bench: 64.3
- lmarena elo: 1382
- gpqa diamond: 68.7
- lmarena rank: 5
- aider polyglot: 71.2
- swe bench verified: 72.3
- humanitys last exam: 18.9
- artificial analysis intelligence index: 58
- artificial analysis price blended per m: 8.75
- artificial analysis speed tokens per sec: 85
Frequently Asked Questions
How much does GPT-5.6 Terra cost per 1M tokens?
Terra costs $2.50 per 1M input tokens and $15.00 per 1M output tokens, with cached input at $0.3125 per 1M, a 90% discount over uncached. That's exactly half of Sol's per-token price, with no batch discount at launch.
How does GPT-5.6 Terra compare on benchmarks vs GPT-5.6 Sol?
Terra scores 72.3% on SWE-bench Verified against Sol's higher mark, and 68.7% on GPQA Diamond, trailing Sol on the hardest reasoning tasks. Sol adds ultra multi-agent mode and a pro reasoning ceiling that Terra doesn't have, but Terra matches GPT-5.5-class performance at half Sol's per-token cost.
Is GPT-5.6 Terra open source or proprietary?
Terra is proprietary. OpenAI serves it only through its own API, ChatGPT, Codex, gateway partners, and select cloud platforms. There are no downloadable weights and no open license.
Does GPT-5.6 Terra train on user data?
By default OpenAI doesn't use API inputs or outputs to train its models unless you opt in, and Terra supports Zero Data Retention for enterprise customers. It's SOC2 Type II, ISO 27001, GDPR, and HIPAA eligible with US and EU data residency options.
Who is GPT-5.6 Terra best for and who should avoid it?
Terra fits engineering teams running production coding agents and cost-sensitive API applications that don't need frontier-level reasoning. Skip it for the hardest 5% of reasoning or research tasks, long-document work past its context limit, or anything needing Sol's ultra multi-agent parallelization; route those to Sol instead.