Grok 4.7 vs Gemini 4 Carbon: Pricing, Specs and Access

Grok 4.7 and Gemini 4 Carbon are both "flagship" releases from the last two months, but they are not really competing for the same job yet. Grok 4.7, from xAI, is a generally available, cheap, fast model built for coding and long office-style work. Gemini 4 Carbon, from Google DeepMind, posts the higher benchmark scores and leads on cybersecurity tasks, but as of this writing it is gated to a small group of cyber defenders and Fairwind Program partners — most developers cannot call it at all. If you need a model you can actually use today, that difference matters more than any leaderboard score.

What's new in each model

Grok 4.7 launched on September 21, 2026. In xAI's own announcement, the company describes it as its most capable model for coding and knowledge work, trained with a longer reinforcement-learning run weighted toward tasks that take many hours rather than a single prompt-and-response exchange. It keeps the same base price as its predecessor, Grok 4.6, while posting higher scores on xAI's own coding and long-task benchmarks, and it was built to work natively with the Grok Bot harness that powers agentic workflows.

Gemini 4 Carbon arrived five weeks later, on October 28, 2026, announced by Koray Kavukcuoglu, SVP of Google DeepMind, in a post titled "Gemini 4 Carbon: our next frontier model" on Google's own blog. Google positions it as its strongest model yet, built specifically for cybersecurity defense and "long-horizon" autonomous tasks — the kind of multi-step agentic jobs that run for hours without a human checking in. It follows Gemini 4 Argon, which had gone generally available only two weeks earlier. For background on what was confirmed at launch, see Pandromeda's earlier look at what's confirmed about Google's new AI.

The two launches make for an odd comparison on paper: Carbon wins more benchmark categories, but Grok 4.7 is the only one of the two you can actually sign up and use right now.

Pricing head-to-head

Both companies publish token pricing, though Carbon's is explicitly introductory and due to rise. Grok 4.7 splits its pricing by context length: requests under 200,000 tokens get the base rate, and once a request crosses that line, every token in it — not just the overflow — bills at the higher tier. Gemini 4 Carbon's pricing is flat regardless of context length, at least for now, with a steep discount on cached input tokens.

Google has said plainly that Carbon's current $2/$10 rate is introductory — it matches the launch pricing Gemini 4 Argon carried before its own price roughly doubled, and Google has told reporters the same step-up applies to Carbon. xAI has made no equivalent announcement for Grok 4.7; its $2/$6 base rate is simply described as "the same price and speed" as Grok 4.6.

Context window and knowledge cutoff

xAI's own model documentation lists Grok 4.7's context window at a flat 500,000 tokens, with a knowledge cutoff of May 2026. That is the number that determines which of the two pricing tiers above applies to a given request.

Google has not published an equivalent context-window figure for Gemini 4 Carbon in its launch materials. The closest public data point is a benchmark called GraphWalks, which tests long-context reasoning across a "256K to 1M" token range and where Carbon scored 96.2% — but that is a benchmark description, not a confirmed maximum context length for the model. Readers who want a hard context-window number for Carbon should treat it as unconfirmed until Google publishes model documentation to match. This is one of several places where a reasoning model's headline spec sheet lags behind its benchmark claims at launch.

Coding and agentic benchmarks

On xAI's own benchmark table, Grok 4.7 (run at its highest "xhigh" reasoning effort) scored 46.3% on CursorBench 4.0, up from 40.4% for Grok 4.6, and 71.0% on DeepSWE v1.1. On AA Briefcase v1.1, a multi-hour office-work benchmark, it reached an Elo of 1,657, and on the Harvey Legal Agent Benchmark it scored 19.6% — well ahead of the 2.5% and 6.7% xAI attributes to two rival models in the same table. xAI's table is self-reported, and at least one independent tracker has listed different numbers for the same tests at different reasoning-effort settings, so the comparisons are not fully controlled.

Google's benchmark table for Gemini 4 Carbon, published on its own model page, shows an Artificial Analysis Intelligence Index v5.0 score of 59 — enough, Google says, to rank first on that index. Carbon also posted 61.2% on FrontierSWE 2 (behind a rival model's 65.5% on the same row of Google's own table), 65.2% on Terminal-Bench 4.0, 81.5% on OSWorld-2.1's offline subset (behind Claude Opus 5.5's 87.2%, again by Google's own table), and 1,768 Elo on WebDev Arena. For a look at how one of those rivals stacks up elsewhere, see Pandromeda's comparison of Claude Opus 5.5 and Sonnet 5.5.

Net effect: Carbon leads on the aggregate intelligence index Google cites, but it does not sweep every coding or agentic category even on Google's own numbers — and Grok 4.7's numbers come from a model anyone can actually call.

Longer-running agentic work

Both companies are chasing the same trend: models that run autonomously for hours rather than answering one prompt at a time. Grok 4.7's own table reports 37.6% on Terminal-Bench 4.0, up sharply from Grok 4.6's 20.3%, and an Elo of 1,695 on GDPval at its "xhigh" reasoning setting — a tier below the 1,735 xAI attributes to one rival model in the same chart, but well ahead of Grok 4.6's 1,605. xAI frames the jump as evidence that Grok 4.7 verifies its own work better over long sessions rather than drifting off task, which is also the rationale it gives for pairing the model natively with the Grok Bot harness used in its agent products.

Gemini 4 Carbon's own benchmark table leans even harder into this category, since Google built it explicitly for "long-horizon" tasks. On Google's figures, Carbon scored 77.4% on Agents' Last Exam v1.1, 53.9% on AutomationBench (a Zapier-built workflow-automation test), 58.6% on Blueprint-Bench 2, and 80.0% on RiemannBench. On Vending-Bench 2 — a benchmark that has an agent run a simulated vending-machine business over a long simulated timeline — Carbon ended with a simulated bank balance of $16,484, which Google cites as a sign of sustained, coherent decision-making rather than a one-off lucky run. None of these figures have independent, reproduced confirmation outside Google's own reporting, and the same caveat applies to xAI's numbers above.

Grok 4.7 also exposes four separate reasoning-effort settings — low, medium, high and xhigh — letting a caller trade latency and cost for more deliberate, multi-step reasoning on a per-request basis. That knob is one of the more practical differences between the two models for anyone actually building an agent: Grok 4.7 lets you dial reasoning effort up or down inside a single API, while Carbon's access gate means most teams cannot run that kind of cost/quality experiment on it at all yet.

Security and safety benchmarks

This is where the two models' different design goals show up most clearly. Gemini 4 Carbon was built specifically for cybersecurity defense, and its scores reflect that: 88.2% on SEC-bench Pro (vulnerability detection), 71.0% on CWE-bench (patching known weaknesses), and 54.8% on ExploitGym, which Google calls a new state of the art for exploitation benchmarks. Google also reports Carbon more than doubled Gemini 4 Argon's success rate at generating secure code internally, cut prompt-injection attack success from 0.7% to 0.0%, and scored a 0.0% attack success rate on the Gray Swan Indirect Prompt Injection benchmark.

Grok 4.7's safety numbers are aimed at a different problem: resisting jailbreaks and misuse rather than defending code. xAI reports it let through only 3.3% of risky dual-use prompts on HackerBench v0.3 and scored 62.4% on LatchBio's biosafety benchmark, and says select cybersecurity partners get invite-only access to its red-team capabilities. Neither company's safety numbers have been independently reproduced in results we could verify; both are the companies' own reporting.

Availability: who can actually use each one

This is the practical dividing line. Grok 4.7 is live now through the Grok API, inside Cursor, and through Grok Build (x.ai/build), with third-party coding harnesses, model routers and cloud platforms also listing it. Anyone with an API key can call it today.

Gemini 4 Carbon is not generally available. Google's own announcement says it is starting with "a subset of cyber defenders," expanding gradually to other trusted partners through its Fairwind Program, with wider access for developers, businesses and consumers planned "as soon as possible" — no date given. Google says it is participating in a voluntary U.S. government pre-release testing process and is still refining guardrails before a broader launch. In practice, unless you already hold Fairwind or cyber-defender access, you cannot get an API key for Carbon right now.

Which one should you actually use?

For nearly everyone reading this, the decision is simpler than the benchmark tables suggest, because only one of these models is available to sign up for:

• General coding, agents and budget-sensitive API work: Grok 4.7 — it is live, priced per the table above, and plugs into Cursor and other coding tools today.
• Multi-hour office and document work: Grok 4.7's AA Briefcase and Harvey Legal Agent scores suggest it is a reasonable fit, though always validate on your own workload before committing a budget.
• Cybersecurity defense (vulnerability detection, patching, exploit research): Gemini 4 Carbon is purpose-built for this and posts the strongest published numbers — but only if your organization already has Fairwind or cyber-defender access.
• Everyone else waiting on Carbon: there is no workaround. Google has not published a signup queue or a GA date, so budgeting around Carbon's eventual $4/$20 standard pricing is premature until it actually ships broadly.

What's next

Watch two things. First, whether Google sets a concrete general-availability date for Gemini 4 Carbon, and whether its promised price step-up to $4/$20 per million tokens lands on the same kind of calendar deadline Argon's introductory pricing had. Second, whether xAI ships a documented context-window increase or a cheaper "fast" tier beyond the 2x-speed/2x-price variant it has already mentioned — context length and per-token cost are usually the first specs vendors revise once a flagship model is a few months old. Until Carbon opens up, the realistic head-to-head for most teams is Grok 4.7 against whichever generally available Gemini or Claude model fits the budget, not Carbon itself.

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