Gemini 4 Carbon took first place on the Artificial Analysis Intelligence Index v5.0 with 59 points1, according to Google's launch table. Claude Opus 5.5 scored 58.2 GPT-6 Astra came in at 57.3 When Google announced Carbon on October 28, 2026,4 only selected cyber defenders could use it, with additional trusted partners phased in through its Fairwind Program.
We set out to write about the top-ranked model that most builders cannot touch. That story expired on November 10, 2026.4 Google opened Carbon to paying developers and enterprises that day. The gate lasted 13 days.
The top spot did not last much longer. Requesty's November 4 snapshot of the index had Carbon first. FourWeekMBA read it in second place on November 15. Local AI Zone put it fourth on November 19.
So the useful question is smaller than the headline. When the best model reaches the public in stages and the lead moves every week, how does a team pick one?
Measured in October, sold in November
Carbon followed a staged release. Its capability was measured in public before most buyers could reach it, and access and price opened in steps after that. Each step answers a different buying question. A leaderboard only reads the first one.
| Gemini 4 Carbon5Requesty · high effort, Nov 4 table | $7.76 |
|---|---|
| Claude Opus 5.55Requesty · Nov 4 table | $10.04 |
| GPT-6.1 Sol3Requesty · scored 55 | $1.27 |
| Carbon price doublesGoogle · $4 input, $20 output per million tokens | DEC 1 |
Step one was defenders only, starting October 28. Google reports Carbon hit 88.2% on SEC-bench Pro6, which tests finding vulnerabilities. It scored 54.8% on ExploitGym4, which tests exploiting them, and 71.0% on CWE-bench7, which tests patching them. Google says those security runs happened without guardrails.
Step two was the paying API, starting November 10. Google's release post put Carbon on the paid tier of the Gemini API under the model ID gemini-4-carbon. Google AI Ultra subscribers also began getting it in the Gemini app. There is no free tier.
Step three is the list price. Google charges $2 per million input tokens and $10 per million output tokens through November 30,4 2026. On December 1 those become $4 and $20.8 A token is a chunk of text about the size of a short word. You pay for the tokens you send and the tokens you get back.
The reason for the gate is plain. A model that scores 54.8% on exploitation can help an attacker as easily as a defender4. Whether safety was the whole motive is unclear. A restricted pilot also lets a provider ration compute and collect feedback from expert users before the crowd arrives.
Why a one-point lead buys almost nothing
A one-point lead is smaller than the swing from a single setting. Artificial Analysis lists Carbon at 60 in high thinking mode4, 58 in medium and 54 in low3. Thinking mode sets how long the model reasons before it answers, and longer reasoning costs more money and time. A team that ships at medium effort gets 58. That is the same figure Google's table gives Opus 5.5, though on a slightly different index version.
The lead also decays fast. Carbon went from first to fourth across 15 days of snapshots. A model chosen on rank alone gets re-chosen within weeks, and every re-choice is engineering time that ships nothing new.
Judge a model by what each finished job costs you. Rank is a reported number. The bill is cash. Requesty's November 4 table priced each index task, meaning the average cost to run one item of the benchmark suite.
On that table, Carbon at high effort cost $7.76 per task5. Opus 5.5 cost $10.04 and GPT-6 Astra cost $6.25. GPT-6.1 Sol scored 55 at $1.27.3 Carbon cost 6.1 times as much as Sol for four more points.
The December price change moves the math again. FourWeekMBA's arithmetic from a leaderboard reading on November 15 put Carbon at $5.48 per index task9. Doubled on December 1, that becomes $10.96, about 9% above Opus 5.5 at $10.04.5 Two snapshots eleven days apart gave Carbon two different task costs. Treat that gap as a warning against trusting any one snapshot.
Access status is just as slippery. Local AI Zone published an analysis on November 11 built on the claim that Carbon had no purchase path. It ran a correction on November 19, because Google had shipped the model on November 10. A dedicated analysis site kept a wrong premise up for eight days. A busy engineering team skimming headlines will make the same mistake.
The scores themselves deserve caution. The Batch, from DeepLearning.AI, noted that most launch figures came from Google's own testing. During the gated phase, developers had no way to check them against their own code. The data is mixed on how well security benchmarks predict ordinary product work.
We think the deeper lesson is about sequence. The leaderboard sees a model on day one of a staged release. A buyer lives with it on day ninety, at list price, under real load.
Today the evidence supports a narrow decision. Carbon earns a slot as a challenger for hard tasks, especially security work, where Google's numbers are strongest. Rank alone does not yet earn it the production default. That call changes if Carbon wins on your own tasks at the effort level you ship and at the December 1 price.
Staged launches may be standard by 2031
Every major event in Carbon's commercial life fits inside 34 days, from October 28 to December 1. Google shipped Carbon about two weeks after Gemini 4 Argon. Carbon held first place for about two weeks of snapshots. The gate itself ran 13 days.
Put those together and a pattern appears. If gates run about as long as leads, the public meets each new leader around the time it loses first place. We doubt Google keeps a two-week release rhythm for five years. What may persist is the ratio between gate length and lead length.
Run the numbers forward to 2031 with one assumption you can check against reality. Suppose the leaderboard gets one new leader a month across all labs. That is 12 changes a year and 60 over five years. A team that re-picks on rank would face 60 decisions4.
The risk is lopsided. A thin routing layer, one config value that names the model, takes a few days to build once. Skip it and you migrate every time the leader changes. Teams that keep their own task set get an edge that compounds, because each new release takes an afternoon to test.
Providers face their own fork. Some will sell peak scores behind gates. Others can counterposition on stable access and predictable pricing, which enterprise buyers already weigh heavily. Whether gating spreads beyond models with exploit skills is unclear, because Carbon is one model from one lab.
Rebuild your model shortlist before December 1
Google's price change on December 1 gives you a natural deadline. Run this process once before that date and once after. Then rerun it every Monday.
Step one: pull 25 real tasks from your last month of logs. Pick ones where a wrong answer costs you something. Next to each, write the correct output or a pass rule a script can check.
Step two: fill three tiers. The production default is the broadly available model you trust today. The challenger is the strongest model for your hardest tasks, and Carbon is the obvious candidate. The fallback is a cheaper model that keeps the service running when the others fail or change price.
Step three: run all 25 tasks on each model at the thinking level you would actually ship. That comes to 75 runs. For each model, log cost per completed task and time to finish. A completed task is one that passed your rule, so divide total spend by passes only, since failed runs still cost money.
For a budget, use napkin math. If each run cost as much as one of Requesty's index tasks for the leaders, at $6.25 to $10.04,5 the 75 runs would land between about $469 and $753. Your real tasks are probably lighter. Measure your first five runs and scale from those.
Step four: add an access column. Check whether your account can call the model today, in your region, on its paid tier. Write the post-December price next to the current one. A model you cannot call scores zero, whatever its rank.
Step five: route every call through one config value that names the model. Swapping tiers then takes a one-line change. Test the swap once on purpose so you know it works before an outage forces it.
Expect the first pass to break. Schema errors and timeouts will show up early. A schema error is output in the wrong shape for your code, such as JSON with a missing field. Count each one as a failed task, because in production it is one.
On December 2, rerun the challenger at $4 and $20.8 If its cost per completed task still beats your default on the hard tasks, it keeps the slot. If it doesn't, move it to watch and check again next Monday.
Practise it step by step: Build the Verification Habit
The bottom line
Pick the model that wins your tasks at the price you will pay.
Carbon's gate lasted 13 days and its lead lasted about as long, so the leaderboard captured it at its peak moment. Buyers live with a model at list price, under real load, months later. A small private task set and a one-line routing switch turn each new release into an afternoon of testing. Carbon has earned a challenger slot for hard and security work, and the December 1 rerun will tell you whether it earns more.
