{"slug": "gpt-6-1-terra-pricing-explained-what-a-cache-heavy-agent-session-costs-the-272k", "title": "GPT-6.1 Terra pricing explained: what a cache-heavy agent session costs, the 272K-token surcharge, and how it compares with Sol, Luna and Claude Sonnet 5.5", "summary": "OpenAI released GPT-6.1 Terra on October 26, 2026, priced at $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens, according to OpenAI's API documentation read on October 27. Against the GPT-5.6 Terra model it replaces ($2 input, $0.20 cached, $12 output), the cut is 50% on fresh input, 58% on output and 75% on cache reads, and on a worked example session of 10 million input tokens (97% cached) plus 200,000 output tokens, GPT-6.1 Terra costs $1.79 versus $4.94 for GPT-5.6 Terra and $3.57 each for GPT-6.1 Sol and Claude Sonnet 5.5. Any request over 272K input tokens is billed at 2x input and cache rates and 1.5x output for the whole request, even though GPT-6.1 Terra accepts up to 922K input tokens within its 1,050,000-token context window.", "body_md": "# GPT-6.1 Terra pricing explained: what a cache-heavy agent session costs, the 272K-token surcharge, and how it compares with Sol, Luna and Claude Sonnet 5.5\n\nOpenAI released GPT-6.1 Terra on October 26, 2026, at $1 per million input tokens and $5 per million output. This page takes the price sheets of OpenAI and Anthropic and works out what one realistic agent coding session costs on each model, where the cut is bigger than half, and where one long prompt doubles the bill.\n\n**This explains reporting by**\n\n[OpenAI, \"Introducing GPT-6.1 Terra\" (October 26, 2026)](https://openai.com/index/introducing-gpt-6-1-terra/).\nRead the original first:\n\n[https://openai.com/index/introducing-gpt-6-1-terra/](https://openai.com/index/introducing-gpt-6-1-terra/)\n\n## In one minute\n\n- GPT-6.1 Terra costs $1 per million input tokens, 5 cents cached and $5 output. GPT-5.6 Terra cost $2, 20 cents and $12.\n- So the cut is 50% on fresh input, 58% on output and 75% on cache reads. Agents mostly read cache, so their bills drop by more than half.\n- On our example session (10M input tokens, 97% cached, 200K output), Terra costs $1.79, GPT-5.6 Terra $4.94, GPT-6.1 Sol and Claude Sonnet 5.5 $3.57 each, and GPT-6 Luna 23 cents.\n- Any request over 272K input tokens is billed at 2x input and cache rates and 1.5x output for the whole request, even though Terra accepts up to 922K input tokens.\n- Artificial Analysis measures 40 cents per task for Terra against $1.40 for GPT-5.6 Terra, and scores it 4 points under Sol and 10 over Luna.\n\n## What OpenAI shipped\n\nOpenAI released GPT-6.1 Terra on October 26, 2026. It is the middle model in OpenAI's lineup: above GPT-6 Luna, below GPT-6.1 Sol and GPT-6 Astra. OpenAI pitches it for everyday work such as drafting, reading technical drawings, filling forms and computer use.\n\nThe facts below come from the model's page in OpenAI's API docs, read on October 27:\n\n- Model name gpt-6.1-terra, the only snapshot listed.\n- Price per million tokens: $1 input, $0.05 cached input, $5 output. Cache writes cost $1.25 with 30-minute retention, or $2.50 with 12-hour retention.\n- Context window 1,050,000 tokens, with at most 922,000 input tokens and 128,000 output tokens. Knowledge cutoff June 1, 2026.\n- Reasoning effort levels none, low, medium (the default), high, xhigh and max.\n- Batch and Flex prices are 50% lower than Standard. Fast mode is 2x Standard. Regional processing adds 10%.\n\nThe Decoder reports it is also in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users.\n\n## The price sheet, side by side\n\nPer million tokens, standard tier, from each vendor's own price page on October 27, 2026 (input / cached input / output):\n\n- GPT-6.1 Terra: $1.00 / $0.05 / $5.00\n- GPT-5.6 Terra (the model it replaces): $2.00 / $0.20 / $12.00\n- GPT-6.1 Sol: $2.00 / $0.10 / $10.00\n- GPT-6 Astra: $10.00 / $1.00 / $50.00\n- GPT-6 Luna: $0.10 / $0.01 / $0.50\n- Claude Sonnet 5.5: $2.00 / $0.10 (cache hit) / $10.00\n- Claude Opus 5.5: $4.00 / $0.20 (cache hit) / $20.00\n\nTwo things stand out. First, Terra is exactly half of GPT-6.1 Sol on every line, and Sol and Claude Sonnet 5.5 have the same list prices. Second, against the old Terra the cut is not even: 50% on fresh input, 58% on output (from $12 to $5) and 75% on cached input (from 20 cents to 5).\n\n## What one agent session costs\n\nList prices hide how agents spend tokens. A coding agent sends the same long context again on every turn, so most of its input tokens are cache reads. Max Woolf measured cache reads at 96% to 99% of agent coding tokens in his own sessions, so the cache price matters more than the headline input price.\n\nHere is a worked example. It assumes one session of 10 million input tokens, 97% of them cache reads, and 200,000 output tokens. Those numbers are our assumption, so change them to match your own logs:\n\n```\n# One agent coding session: 10M input tokens, 97% of them cache reads, 200K output tokens.\n# Prices are per million tokens, from each vendor's own price page on October 27, 2026.\nprices = {  # (input, cached input, output)\n    'GPT-6.1 Terra':     (1.00, 0.05,  5.00),\n    'GPT-5.6 Terra':     (2.00, 0.20, 12.00),\n    'GPT-6.1 Sol':       (2.00, 0.10, 10.00),\n    'GPT-6 Luna':        (0.10, 0.01,  0.50),\n    'Claude Sonnet 5.5': (2.00, 0.10, 10.00),\n    'Claude Opus 5.5':   (4.00, 0.20, 20.00),\n}\ntokens_in, cached_share, tokens_out = 10_000_000, 0.97, 200_000\nfor model, (inp, cached, out) in prices.items():\n    cost = (tokens_in * cached_share * cached\n            + tokens_in * (1 - cached_share) * inp\n            + tokens_out * out) / 1_000_000\n    print(f'{model:18} ${cost:.2f}')\n```\n\nOutput:\n\n```\nGPT-6.1 Terra      $1.79\nGPT-5.6 Terra      $4.94\nGPT-6.1 Sol        $3.57\nGPT-6 Luna         $0.23\nClaude Sonnet 5.5  $3.57\nClaude Opus 5.5    $7.14\n```\n\nOn this session Terra costs 64% less than GPT-5.6 Terra and half of Sol or Sonnet 5.5. Luna is still about 8 times cheaper again.\n\nThe example leaves out cache writes and assumes every model uses the same number of tokens for the same job. They do not. A model that thinks longer or needs more turns spends more tokens, so the measured cost per task below is the better guide.\n\n## The 272K-token rule\n\nTerra's page says it accepts up to 922,000 input tokens. But the same page says prompts with more than 272,000 input tokens are priced at 2x the input and cache rates and 1.5x output, for the full request. GPT-6.1 Sol and GPT-6 Astra carry the same rule.\n\nThe jump is a cliff, not a slope. With no caching and 5,000 output tokens:\n\n- A 250,000-token prompt costs 25 cents of input plus 2.5 cents of output, about 28 cents.\n- A 300,000-token prompt costs 60 cents of input plus 3.75 cents of output, about 64 cents.\n\nSo 20% more prompt costs about 2.3 times as much. If your agent stuffs whole repositories or long documents into one call, measure where your prompts land before you count on the savings.\n\n## What the benchmarks say\n\nArtificial Analysis tested Terra on launch day:\n\n- Cost per task at max effort: 40 cents for GPT-6.1 Terra, against $1.40 for GPT-5.6 Terra.\n- Intelligence Index: 6 points above GPT-5.6 Terra at max effort, 4 points below GPT-6.1 Sol and 10 above GPT-6 Luna. At lower effort levels the gain over the old Terra is 7 to 13 points.\n- Coding Agent Index v2.0: Terra 48, Sol 54, Luna 31.\n- AA-Omniscience hallucination rate: 38%, down from 88% for GPT-5.6 Terra.\n\nOpenAI's Ted Sanders gave the simplest guide: \"If you care about quality only, pick Astra. If you're a bit price sensitive, pick Sol. If you're very price sensitive, pick Terra or Luna.\" He also called Terra \"not a bombshell release\", according to Simon Willison's notes.\n\n## How to decide\n\n- If you run GPT-5.6 Terra today, switching is close to free money: the new model is cheaper on every line and scores higher. Rerun your own evals first.\n- If you run GPT-6.1 Sol or Claude Sonnet 5.5 on routine agent work, Terra is half the list price. On Artificial Analysis's coding index it trails Sol by 6 points, so test it on your real tasks, not on one demo.\n- If your job is simple extraction or classification, GPT-6 Luna is still about a tenth of Terra's price.\n- Log input tokens per request. Anything near 272K needs splitting or trimming, or the price cut disappears for that call.\n- Check cache hit rates. Most of the saving for agents comes from the 5-cent cache read, so a prompt layout that breaks caching throws much of it away.\n\n## Who is affected\n\n| Case | Status | \n|---|---|\n| Teams running GPT-5.6 Terra through the API | Terra 6.1 is cheaper on every price line and scores higher on Artificial Analysis. Rerun your evals and switch. | \n| Agent and coding workloads on GPT-6.1 Sol or Claude Sonnet 5.5 | Terra is half the list price. Worth a test on real tasks, since it trails Sol on most benchmarks. | \n| Workloads with prompts over 272K tokens | Billed at 2x input and cache rates and 1.5x output for the whole request. The cut can vanish. | \n| ChatGPT Work and Codex users on Plus, Pro, Enterprise and Edu | Can pick Terra, per The Decoder. No API bill involved. | \n\n## What to do\n\n- Pull a week of API logs and compute your own cache-read share and output share. Plug them into the script above.\n- Run your eval set on gpt-6.1-terra at the effort level you use today, then at one level higher, and compare cost per passing task.\n- Add an alert for any request over 250K input tokens, so you see the 272K line coming.\n- Keep the stable part of every prompt (system prompt, tools, repository context) at the front, so it stays cached.\n\n## What is still unknown\n\n- Our session numbers (10M input, 97% cached, 200K output) are an example, not a measurement. Your mix will differ.\n- The example assumes each model spends the same tokens on the same job. Real token use differs by model and effort level, which is why Artificial Analysis's cost per task is the better comparison.\n- OpenAI's speed and efficiency claims against Astra, Sol and Claude Opus 5.5 come from OpenAI's own tests.\n- Artificial Analysis says Terra gained about 20 points on Terminal-Bench 4.0, while The Decoder reports 50.3% against 21.5%. We did not resolve the difference.\n- Prices are as published on October 27, 2026. Vendors change them often.\n\n## Sources\n\n[AI News Report](https://theainewsreport.com/)· every headline, every morning.", "url": "https://wpnews.pro/news/gpt-6-1-terra-pricing-explained-what-a-cache-heavy-agent-session-costs-the-272k", "canonical_source": "https://theainewsreport.com/2026-10-27-gpt-6-1-terra-pricing-agent-cost-explained.html", "published_at": "2026-10-27 15:31:24+00:00", "updated_at": "2026-10-27 15:43:36.383566+00:00", "lang": "en", "topics": ["large-language-models", "ai-products", "ai-agents", "artificial-intelligence"], "entities": ["OpenAI", "GPT-6.1 Terra", "GPT-5.6 Terra", "GPT-6.1 Sol", "GPT-6 Luna", "Claude Sonnet 5.5", "Anthropic", "Artificial Analysis"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/gpt-6-1-terra-pricing-explained-what-a-cache-heavy-agent-session-costs-the-272k", "markdown": "https://wpnews.pro/news/gpt-6-1-terra-pricing-explained-what-a-cache-heavy-agent-session-costs-the-272k.md", "text": "https://wpnews.pro/news/gpt-6-1-terra-pricing-explained-what-a-cache-heavy-agent-session-costs-the-272k.txt", "jsonld": "https://wpnews.pro/news/gpt-6-1-terra-pricing-explained-what-a-cache-heavy-agent-session-costs-the-272k.jsonld"}}