SI Updated Nov 7, 2026. Figures below come from each lab’s own launch posts unless noted. Benchmarks are not interchangeable across vendors, and “Critical” under OpenAI’s Preparedness Framework is a capability-threshold label with safeguards, not an unlock for unrestricted cyber use.
Model comparisons
Terra, Carbon, and Opus 5.5: what the labs actually shipped
A dated, sourced comparison of three late-2026 frontier models. Focused on access, cybersecurity posture, and price - not a single winner score.
At a glance
1. GPT-6.1 Terra - available, cheap, labeled Critical
OpenAI positions Terra as a mid-tier workhorse: visual reasoning plus computer use, with API prices 50% below GPT-5.6 Terra and GPT-6.1 Sol. At max effort it reports 50.3% on Terminal-Bench 4.0 (vs 21.5% for GPT-5.6 Terra) and says it finishes tasks more than 3× faster than GPT-6 Astra and GPT-6.1 Sol on its own timing runs.
The policy-relevant line is cybersecurity: OpenAI says Terra meets the Critical threshold under its Preparedness Framework while keeping Astra-era protections. That is a deployment classification, not a public exploit toolkit. It also reports stronger prompt-injection resistance and fewer scope/honesty failures than GPT-5.6 Terra.
2. Gemini 4 Carbon - cyber lead, gated access
Google calls Carbon its strongest model and cites #1 on Artificial Analysis Intelligence Index v5.0 at 59. The cyber numbers it publishes are the clearest differentiator versus Terra: 88.2% SEC-bench Pro (finding), 54.8% ExploitGym (exploitation), 71.0% CWE-bench (patching), with those cyber benches evaluated without guardrails.
Access is the other half of the story. Carbon opens first to a subset of cyber defenders, then trusted Fairwind partners, with developers, businesses, and consumers later. That is both a safety choice and a market choice: the model with the loudest cyber scores is the one most people cannot call yet, while Terra ships broadly with a Critical label behind product safeguards.
3. Claude Opus 5.5 - the coding/agent baseline still in the mix
Opus 5.5 is older than the late-October Terra/Carbon pair, but it is still the publicly available Anthropic workhorse for long agentic coding. Anthropic says it matches Claude Fable 5.1 on most work, costs about 40% less than Opus 5 on typical loads, and scores best-to-date on Anthropic’s automated behavioral audit.
On cyber and biology, Anthropic treats it like its higher-risk tier: similar safeguards to Fable 5.1, with a Cyber Verification Program for vetted practitioners. OpenAI’s Terra post claims Terra is over 3× as token-efficient as Opus 5.5 at max effort and about one-fourth the price per token; treat that as an OpenAI comparison, not an independent bake-off.
How to read this without a fake ranking
- Need it today at API scale: Terra and Opus 5.5 are broadly available; Carbon is not.
- Care most about published cyber benches: Carbon’s SEC-bench / ExploitGym / CWE-bench lead is the clearest vendor claim - with gated rollout.
- Care most about cost for computer-use volume: Terra’s $1 / $5 list price is the aggressive end of this trio.
- Care most about long coding agents with a full public table from the maker: Opus 5.5 is still Anthropic’s documented default for that job.
The same late-October week also brought OpenAI’s misalignment-monitoring expansion after the Hugging Face evaluation incident. Capability headlines and oversight headlines are now the same story. See our piece on that incident for the other half.