Gemini 4 Carbon vs GPT-6.1 Terra: Two New Models in One Week, Double the Price Gap and an 8.4-Point Index Gap

The two models are rolling out in opposite ways: Carbon opens this week to a subset of cyber defenders, then expands to more trusted partners through the Fairwind Program, with no date set for general availability to developers, businesses and consumers; Terra has been available since Oct 26 to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and to developers through the API as gpt-6.1-terra.

Their positioning differs: Carbon is Google's strongest model to date, aimed at cybersecurity and long-horizon engineering work, which Google says can run for days on vulnerability discovery and patching across large codebases and low-level systems; Terra is aimed at visual reasoning, computer use and writing, handling everyday work such as drafting proposals, interpreting technical drawings, filling in forms and editing videos.

Pricing moves in different directions: Carbon's launch pricing matches Argon, with input and output prices each doubling after the launch period; Terra's standard pricing is half that of its predecessor GPT-5.6 Terra and of GPT-6.1 Sol, positioning it as the execution layer for high-volume everyday tasks.

A note on benchmark methodology: Artificial Analysis' index measures both models on the same nine evaluations, while each vendor's own benchmarks use its own methodology and chosen tasks; every figure below names its source, configuration and measurer, and scores across different benchmarks are not directly comparable.

1. Maker and release date: Carbon was released by Google DeepMind on Oct 28, 2026; Terra was released by OpenAI on Oct 26, 2026.
2. Context window: Carbon: 1M tokens (Artificial Analysis model details); Terra: 1.05M tokens, maximum output 128,000 tokens, knowledge cutoff June 1, 2026 (OpenAI model documentation).
3. API input and output price: Carbon: $2 input and $10 output per million tokens at launch, $4 and $20 after the launch period; Terra: $1 input and $5 output per million tokens at standard pricing.
4. Cached input price: Carbon: cached input at a 95% discount, $0.10 per million at launch and $0.20 at standard pricing; Terra: cached input at $0.05 per million.
5. Long-prompt and batch pricing: Terra prompts over 272,000 input tokens are billed at twice the input rate and 1.5 times the output rate for the whole request, with Batch and Flex at 50% off; no long-prompt premium has been published for Carbon.
6. Open weights: Neither model's weights have been published; Terra's parameter count has also not been published.
7. Availability and subscription: Carbon opens this week to a subset of cyber defenders and expands through the Fairwind Program, with no general-availability date and no standalone subscription price published; Terra is available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, with GitHub Copilot and AWS Bedrock launching the same day, Oct 26, and no standalone subscription price published.
8. Input modalities: Carbon accepts text, image, video and speech input and outputs text; Terra accepts text and image input and outputs text, with audio and video input not listed as supported on its model page.

1. Independent index: Artificial Analysis Intelligence Index v5.0 (Oct 28, 2026): Carbon at high effort scores 59.4, at $7.76 per task (standard pricing); Terra at max effort scores 51.0, at $0.96 per task; a gap of 8.4 points.
2. Carbon cybersecurity trio: Published by Google, measured without guardrails: SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0%.
3. Carbon other published scores: Artificial Analysis AA-Omniscience v1.1: 54, with 58% accuracy and an 11% hallucination rate; Google's published WebDev Arena ranking: 4th at 1768 Elo, up 89 Elo from its predecessor.
4. Terra computer use: Published by OpenAI: Agents' Last Exam v1, 57.6% at medium effort and 60.0% at max effort; OSWorld 2.1 offline set, 66.7% at medium effort and 74.2% at max effort.
5. Terra vision and coding: Published by OpenAI: MMMU-Pro, 85.5% at max effort and 84.6% at medium effort; Terminal-Bench 4.0, 50.3% at max effort, against 21.5% for its predecessor GPT-5.6 Terra.
6. Terra factual-error rate: OpenAI's internal error-inducing prompt set: 16.7% at low effort, against 53.6% for its predecessor; 4.3% at max effort.
7. Not published: Carbon has not published Agents' Last Exam, OSWorld or MMMU-Pro scores; Terra has not published SEC-bench Pro, ExploitGym or CWE-bench scores.

1. For vulnerability discovery and patching: If the task is finding, validating and patching vulnerabilities and access is already within the Fairwind defender cohort, Carbon fits, with published results covering discovery, validation and patching.
2. For computer use available now: If the task needs computer operation, screen reading, form filling and everyday document handling in Work, Codex or the API right away, Terra fits, with all three channels open since Oct 26.
3. For lower per-token list pricing: On list pricing, Terra fits, at a standard $1 input and $5 output per million tokens, against Carbon's launch $2 and $10.
4. For four input modalities: If the task spans text, image, video and speech input, Carbon fits; if the task needs a stated 128,000-token maximum output and a June 1, 2026 knowledge cutoff, Terra fits.
5. For a 1.05M context: If context needs exceed 1M tokens, Terra's 1.05M fits; for the higher index score at a standard $7.76 per task, Carbon's 59.4 fits (measured at high effort, against Terra at max effort, so configurations differ).