Gemini 4 Carbon vs Opus 5.5 vs GPT-6.1 Terra: Price and Benchmarks

Short version: GPT-6.1 Terra is the cheapest of the three at $1 input / $5 output per million tokens and the fastest for everyday computer-use tasks. Claude Opus 5.5 is the best-documented all-rounder for coding and long agentic work at $4 / $20 per million tokens, with a full published benchmark table from Anthropic. Gemini 4 Carbon posts the highest score on the independent Artificial Analysis Intelligence Index and leads on cybersecurity benchmarks, but Google has only opened it to a small group of vetted cyber-defense partners, not to the general public, so most readers cannot use it yet.

Which model wins, in one paragraph

There isn't a single winner, because the three models were not built for the same job. Gemini 4 Carbon is Google's new top-end model, pitched first at cybersecurity defenders rather than ordinary developers. Claude Opus 5.5 is Anthropic's workhorse for coding and multi-step agentic work, with the most detailed public benchmark table of the three. GPT-6.1 Terra is OpenAI's mid-tier, high-throughput model for computer use — reading screens, clicking through apps, and running long task chains cheaply and quickly. If your decision comes down to price and raw availability today, Terra is the easy pick. If it comes down to coding and agent benchmarks you can actually check against a published table, Opus 5.5 has the edge. If it comes down to who leads an independent intelligence ranking and who is strongest on security work, that's Carbon — assuming you can get access to it at all right now.

What each company actually shipped

Google's blog post announcing the model, published October 28, 2026, describes Gemini 4 Carbon as its "next frontier model," built on top of the publicly available Gemini 4 Argon. Google is not opening Carbon to everyone at once. Access starts this week with "a subset of cyber defenders" and expands gradually to trusted partners through what Google calls the Fairwind Program. Developer, business, and consumer access is described only as "coming soon," with no date attached. Google also says it is participating in the U.S. government's voluntary pre-release testing process for the model.

Anthropic's launch announcement for Claude Opus 5.5, dated September 22, 2026, positions it as the model most people should start with for "long-running agentic coding and knowledge work," while reserving Claude Fable 5.1 for the heaviest reasoning jobs. Opus 5.5 is already live across the Claude apps and the Claude API under the model ID claude-opus-5-5.

OpenAI's announcement says it shipped GPT-6.1 Terra on October 26, 2026, as a mid-tier update aimed squarely at computer-use workloads — the kind of task where a model is reading a screen, operating software, and acting across multiple apps rather than just answering questions in a chat window. OpenAI's own system-card addendum for the model is dated the same day. Terra sits below GPT-6.1 Sol and GPT-6 Astra in OpenAI's lineup, at a noticeably lower price.

Price and context window, side by side

Here is what each company's own pricing and model documentation states. Figures are the standard list price per million tokens; none of the three companies publishes the other two's numbers, so this table pulls each row from that company's own page.

Two caveats are worth flagging. Google's post does not state a context window or a max output figure for Carbon anywhere in the announcement, which is unusual for a flagship launch and likely reflects the staged, partner-first rollout. And OpenAI's pricing for Terra steps up for very long prompts: once an input passes 272,000 tokens, the entire request is billed at 2x the input and cache rate and 1.5x the output rate, according to OpenAI's model documentation. Anthropic's cache pricing for Opus 5.5 is $0.20 per million tokens for cache reads and $5 per million for cache writes, against $0.50 and $6.25 on the previous Opus 5; OpenAI prices Terra's cached input at $0.05 per million and cache writes at $1.25 per million with 30-minute retention. Anthropic also offers a Fast mode for Opus 5.5 at $8 input / $40 output per million tokens for up to 2.5x the speed.

What the benchmark tables actually say

All three companies grade their own models, so every number below is self-reported unless marked otherwise. Treat them as each company's claim about its own model, not as an independently audited result.

Anthropic's table for Opus 5.5 compares it against the larger Claude Fable 5.1 and the previous Claude Opus 5, all at max effort unless stated:

Anthropic also reports a GDPval-AA v2.1 Elo score of 1846 for Opus 5.5, against 1735 for Fable 5.1 and 1708 for Opus 5 — GDPval-AA is Anthropic's own Elo-style economic-tasks ranking, not an outside benchmark.

Google's claims for Gemini 4 Carbon lean heavily toward security work. Google cites an Artificial Analysis Intelligence Index v5.0 score of 59, which it says puts Carbon in first place; Artificial Analysis is a third-party benchmark aggregator, so that ranking isn't Google grading its own homework, though Google is the one publicizing it. On Google's own security suites, Carbon scores 88.2% on SEC-bench Pro (vulnerability finding), 54.8% on ExploitGym (exploitation), and 71.0% on CWE-bench (patching). On WebDev Arena, Google reports an Elo of 1768, placing Carbon fourth, up 89 points from its prior model. Google also highlights two safety-flavored metrics rather than raw capability: a 0.0% "cheating" rate on Terminal-Bench 4.0 (versus 0.4% attempted), and a 0.0% attack-success rate on the Gray Swan indirect-prompt-injection benchmark, down from 0.7% for Gemini 4 Argon.

OpenAI's claims for GPT-6.1 Terra, per its announcement and developer documentation, center on speed and cost rather than peak capability. OpenAI says Terra solves 50.3% of tasks on Terminal-Bench 4.0 at maximum reasoning effort, more than double the 21.5% it attributes to the previous GPT-5.6 Terra. On an internal error-inducing prompt set, OpenAI reports the error rate at low reasoning effort falling from 53.6% (GPT-5.6 Terra) to 16.7% (GPT-6.1 Terra). OpenAI also says Terra finishes tasks more than three times faster on average than both GPT-6 Astra and GPT-6.1 Sol, based on its own timing runs — a speed and cost claim more than a quality one.

Coding, agents, and computer use

If the job is writing and debugging code inside a terminal or editor, Opus 5.5 is the one of the three with a full, apples-to-apples published table against its own prior generation: Terminal-Bench 4.0, FrontierCode, and CursorBench scores all moved up double digits over Opus 5. Readers who want more detail can check our write-up of Claude Opus 5.5's launch for the full pricing and access breakdown.

If the job is operating software directly — clicking through a browser, filling forms, chaining steps across apps — GPT-6.1 Terra is the one OpenAI built and priced specifically for that at scale, and it is the cheapest of the three to run continuously. See how Terra stacks up against GPT-6.1 Sol for the full picture against OpenAI's own higher-tier model.

Gemini 4 Carbon's own coding-adjacent number is its WebDev Arena placement (4th, Elo 1768), alongside the security-specific SEC-bench Pro and CWE-bench scores. Google is framing Carbon first as a security tool for defenders, not as a general coding assistant competing directly with Opus 5.5 or Terra on day one. For a wider view of where Carbon's predecessor landed against Claude and GPT models, see our rundown of Gemini 4 Carbon's rollout and pricing.

Safety classifications and health benchmarks

OpenAI's system-card addendum for GPT-6.1 Terra, published the same day as the model, classifies the model as Critical for cybersecurity and High for biological and chemical risk under OpenAI's Preparedness Framework, with AI self-improvement rated below the High threshold. On health-related evaluations, OpenAI reports a length-adjusted HealthBench Professional score of 62.7 and a MentalHealthBench overall score of 55.5%.

Google's safety framing for Gemini 4 Carbon is built around the Gray Swan indirect-prompt-injection result (0.0% attack-success rate) and the Terminal-Bench cheating-rate metric described above, plus the stated participation in U.S. government pre-release testing. Anthropic has not published a comparable standalone safety table for Opus 5.5 alongside the capability benchmarks on its announcement page; its safety work for the model is documented separately in its usual model and system-card materials.

Who can actually use each model today

This is the sharpest practical difference between the three. Claude Opus 5.5 and GPT-6.1 Terra are both live now for ordinary developers and subscribers: Opus 5.5 through the Claude apps and the Claude API, Terra through the OpenAI API and ChatGPT. Gemini 4 Carbon is not in that position yet — Google's own post is explicit that initial access goes to a limited set of cyber-defense partners, with broader developer, business, and consumer availability left as "coming soon" and undated. If search demand for "Gemini 4 Carbon vs Opus 5.5" is already climbing, as it appears to be, most of that traffic is for a model most people cannot sign up for yet.

Which one should you actually use

For general coding work, long agent runs, and tasks where you want a public, detailed benchmark table to check Anthropic's claims against, Claude Opus 5.5 is the most practical of the three right now. For high-volume, latency-sensitive computer-use work — browser automation, large batches of screen-reading tasks, anything billed by the token at scale — GPT-6.1 Terra's $1/$5 pricing and OpenAI's speed claims make it the cost-effective default. Gemini 4 Carbon is the one to watch rather than the one to build on today: its security benchmarks and Artificial Analysis ranking are the strongest of the three on paper, but until Google widens access past cyber-defense partners, it isn't a realistic option for most developers or businesses evaluating this comparison.

What's next

The most consequential open variable is Google's access timeline — Carbon's "coming soon" language leaves developer and consumer availability undated, so the real three-way comparison on equal footing (same sign-up process, same API self-serve access) doesn't fully exist yet. Anthropic has already signaled that Claude Sonnet 5.5 and Claude Haiku 5.5 are coming in the weeks after Opus 5.5, which will reshape the price-to-performance picture at the lower end. OpenAI, for its part, now has three recent GPT-6-family releases in play — Astra, Sol, and Terra — at three different price points, and how it continues to segment that lineup will matter as much as any single Terra benchmark. Expect pricing on all three sides to keep moving as introductory rates (like Carbon's) expire and as each company responds to the others' numbers.

Frequently asked questions

Is Gemini 4 Carbon available to the public yet? No. Google's October 28, 2026 announcement says initial access goes to a subset of cyber-defense partners through its Fairwind Program, with developer, business, and consumer access described only as coming soon.

What does each model cost per million tokens? Gemini 4 Carbon is $2 input / $10 output introductory, rising to $4 / $20. Claude Opus 5.5 is $4 input / $20 output. GPT-6.1 Terra is $1 input / $5 output. All figures are each company's own standard API list price.

Which model has the biggest context window? Of the two with a disclosed figure, GPT-6.1 Terra's 1.05-million-token window is slightly larger than Claude Opus 5.5's 1-million-token window. Google has not published a context window figure for Gemini 4 Carbon.

Are the benchmark scores independently verified? No. Every score in this comparison comes from the model's own maker. The one partial exception is Google's citation of the Artificial Analysis Intelligence Index, a third-party aggregator, though Google is the one highlighting that result.

Is GPT-6.1 Terra a replacement for GPT-6 Astra? No. OpenAI positions Terra as a cheaper, faster, mid-tier model for computer-use workloads, sitting below GPT-6.1 Sol and GPT-6 Astra in its lineup rather than replacing them.

Which model should a developer pick today? It depends on the workload: Opus 5.5 for coding and agentic tasks with a published benchmark table, Terra for cheap, fast, high-volume computer-use work, and Carbon only if you already have access through Google's partner program.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage