Inter-office memo

To
Everyone using Gemini 4 Carbon in Antigravity, the Gemini app or the API
From
Mo, Chief Executive
Date
November 18, 2026, 6:16 AM CST
Re
Why Gemini 4 Carbon is slow

Why is Gemini 4 Carbon so slow? What the tests show, and what to use instead

Mostly because it does more work, not just because it runs slower. Kernel’s traces show about seven times GPT-6 Astra’s steps, and Artificial Analysis times Carbon at 34.5 minutes a task at high, against 5.3 for GPT-6.2 Astra at the same score. Update Antigravity, which stopped cutting off long thinking on November 17, and use Orchestrator or Hydrogen for routine work.

How slow Carbon is

Gemini 4 Carbon is Google’s strongest model, in Antigravity on Google AI Ultra since November 10, and in the Gemini app on Ultra and on the Gemini API’s paid tier. Since it reached Ultra, the complaint has been the same. On November 11 the developer Theo posted: “I don’t really know how to properly emphasize this. Gemini 4 Carbon is unacceptably slow in real world usage.” He timed one job at 22 minutes 21 seconds on Carbon at low, against 9 minutes 33 seconds on Claude Opus 5.5 at high. His first take, three minutes earlier and seen more than 500,000 times, began “It’s slow, awkward, stupid, and hallucinates all the time,” and went on to say Google had “made a ton of progress.”

On November 18, Kernel’s analysis of why reached the front page of Hacker News. Running Carbon on one of its private evaluations, Kernel found it “spent close to an hour completing tasks that routinely took Opus and Astra only a few minutes.”

Artificial Analysis, which measures each model through its maker’s API, says the same in numbers. Its “minutes per task” is the output tokens a model writes per task of its Intelligence Index, divided by its output speed:

Model and thinking level Intelligence Index Minutes per task Time to first answer Output speed
Gemini 4 Carbon, high 60 34.5 74.86 s 59 tokens/s
Gemini 4 Carbon, medium 58 29.2 27.81 s 59 tokens/s
Gemini 4 Carbon, low 54 29.8 8.74 s 48 tokens/s
Gemini 4 Argon, high 50 29.3 47.96 s 66 tokens/s
Gemini 4 Hydrogen, high 51 8.5 17.44 s 167 tokens/s
Claude Opus 5.5, high 52 9.3 20.43 s 108 tokens/s
Claude Opus 5.5, max 59 24.6 353.86 s 117 tokens/s
GPT-6.2 Astra, high 60 5.3 22.07 s 56 tokens/s
GPT-6.2 Astra, xhigh 62 6.8 52.89 s 54 tokens/s

Figures as Artificial Analysis showed them on November 18, 2026. At the same score of 60, Carbon at high takes about six and a half times as long per task as GPT-6.2 Astra at high. Artificial Analysis calls it “among the slowest” in its price tier, where the median is 12.4 minutes. Argon, which paid Google AI Pro gets instead of Carbon, is nearly as slow.

It does more work, not just slower work

The table holds the first clue: Carbon writes 59 tokens a second, a little faster than GPT-6.2 Astra’s 56, and still takes far longer. Kernel’s traces explain why. Slower inference accounts for “16–30% of Carbon’s extra time compared with Astra, and 45–49% compared with Opus.” The rest is work:

  • More steps. Carbon took about seven times as many steps as GPT-6 Astra and eleven times as many actions, such as reading a file or running a command. It wrote seven times as much, and because each step resends a longer conversation, it processed more than sixteen times as much input.
  • Work nobody asked for. “Half of Carbon’s steps went into work the ticket never asked for”: auditing code the fix didn’t touch, and guarding against cases the bug report had already ruled out. In one sample run, Carbon found the cause in about six minutes and then wrote no code for another 25.
  • Less done per step. Astra made more than two tool calls per step, often reading several files at once; Carbon made about one. And “39% of the lines it read were ones it had already seen, against 7% for Astra.” Kernel estimates these two habits alone account for about a quarter of Carbon’s extra time.

It isn’t the coding agent around it: Kernel also ran Carbon in OpenCode and mini-SWE-agent and “saw no meaningful difference.” Nor is it simply the thinking level: on low, Carbon “often failed to produce a fix at all,” going in circles instead. When Kernel’s harness asked “do you already have enough to make the fix?” every five turns, one run finished in 28 minutes instead of 37. Kernel’s conclusion: “the slowness seems to come from how Carbon behaves, not what it knows.” That’s one company’s private test, with the check-in tried on a single run, but it fits the public numbers above.

What Google has said

Google hasn’t called it a problem, but its own pages describe it. The model card lists, under known limitations: “There may also be occasional slowness or timeout issues, and at times the model may use more tokens to maximize performance, especially at higher thinking levels.” The developer guide says Carbon “spends more thinking tokens per turn to explore thoroughly and verify its work, which increases per-turn token use and latency compared to Argon,” and tells developers: “Hard prompts can take several minutes. Use background execution, or set HTTP client timeouts to at least 15 minutes.”

Google also says what Carbon is for. Its Antigravity launch post calls it “ideal for high-leverage coordinator, architecture, cybersecurity auditing, and adversarial verification roles while lighter, faster models handle the parallel file edits and routine implementation tasks.” Two days later Google added Orchestrator, where Carbon plans and Gemini 4 Hydrogen does the work. On Terminal-Bench 4.1, Google says, it “reached near-Carbon task success rates while, on average, consuming 24% less quota and completing 19% faster.” Those are Google’s own numbers.

Google has shipped fixes for long turns that time out. Antigravity App 2.33.0, dated November 17, fixed “an issue where long reasoning turns timed out after five minutes of thinking.” Antigravity CLI 1.5.10, posted to GitHub late on November 16, US time, stopped canceling and restarting “long, high-effort model responses” that went silent for five minutes, and “now waits up to 10 minutes before it treats a silent response as stalled.” It also waits and retries when the API asks it to hold off for up to 4 minutes, so “per-minute rate and capacity limits no longer end your turn.” Google notes that new versions “roll out gradually and may take a few days to reach all users.”

What Google hasn’t said: whether it’s working to make Carbon faster, or whether the reports below are an incident.

Slow, or broken?

Some of what people report this week goes beyond slow. On November 16, a Google AI Ultra subscriber wrote on Google’s developer forum that Carbon, Argon and Hydrogen were “Frequently unusable or extremely slow,” with “an immediate ‘Model unavailable’ error with a retry loop.” In a thread viewed more than 1,700 times, Ultra users report the opposite: Carbon and Argon skipping their thinking and answering at once, in Antigravity and, some say, on gemini.google.com. One wrote on November 17 that it “Seems fixed”; others said a day later that it wasn’t. Another says Carbon is “constantly stuck on ‘Working’” at every thinking level, and asks “Did I subscribe to Ultra just to use Hydrogen?” A fourth, on November 18, blames dropped prompt caches for turns that take “several minutes” and use up quota.

No one from Google has replied in these threads. A forum regular, not marked as Google staff, answered one that slow first responses come from “real-time serving load and deeper reasoning budgets rather than a rate limit on your account.” Google’s AI Studio status page lists no incident since October 30, and Antigravity has no status page of its own.

The difference matters. A slow Carbon is still thinking: wait, or move the job. A Carbon that answers at once without thinking, or loops on “Model unavailable” while your quota is untouched, points to a fault on Google’s side, and our guide to when Antigravity is down covers what to do.

Who it affects

Among Google’s consumer plans, Carbon is only on Google AI Ultra, at $99.99 a month for “5x higher usage limits vs. AI Pro” or $199.99 for 20x, by Google’s plans page. It’s also on Enterprise, and on the Gemini API’s paid tier, which charges $2 per million input tokens and $10 output through November 30, then $4 and $20 from December 1, with thinking tokens billed as output.

Slow also means costly. Antigravity’s plans page says its “rate limits correlate with the amount of work done by the agent,” so “more complex tasks consume more quota.” A model that takes many more steps per task spends them against your five-hour and weekly limits too.

What to do

  1. Update Antigravity. Check for App 2.33.0 or CLI 1.5.10 or later, so long thinking no longer fails at five minutes.
  2. Use Orchestrator for everyday work. Antigravity’s models page calls Orchestrator the “recommended default” for “feature work, bug fixes, refactoring, and multi-file debugging.” Ultra has Orchestrator (Carbon) and Orchestrator (Argon); paid Pro has Orchestrator (Argon). It runs “with a fixed configuration,” so you can’t set its thinking level.
  3. Give routine tasks to Hydrogen. At high, it takes 8.5 minutes per task on Artificial Analysis’s measure, against Carbon’s 34.5. Keep Carbon for what Google says it’s for: hard debugging, design and review.
  4. Write tighter tasks. Kernel’s runs went long on checks nobody asked for. Say what’s out of scope, and tell Carbon to make the fix once it has found the cause.
  5. Lower the thinking level for quick questions only. Low brings Carbon’s first answer in about 9 seconds instead of 75, but whole tasks barely get shorter, and Kernel found low often never fixed the bug.
  6. Switch to Claude inside Antigravity. Ultra and paid Pro also have Claude Opus 5.5 and Sonnet 5.5, and Antigravity’s models page shows usage for “Claude and GPT models” apart from Gemini’s, each with a weekly and a five-hour limit.
  7. On the API, plan for long calls. Follow Google’s advice: background execution or timeouts of at least 15 minutes, and a max_output_tokens of at least 32,768, since Carbon “can spend more of the limit on thinking.”

Carbon in Enjoy

Enjoy runs Antigravity on your Google account, through the server Google offers apps such as Zed and JetBrains IDEs, beside Claude Code, Codex and the other labs’ agents, each on its own plan. Its model list is the Gemini models Google offers your account, and its effort levels for Antigravity are Model default, Low, Medium and High. Google documents Orchestrator in its own app and CLI, not in other apps. When Carbon is still auditing an hour in, you can switch agents in the middle of the conversation: select Stop and move to beside Working, and Claude Code or Codex picks the job up with the conversation so far, on the plans you already have. Enjoy is a monthly subscription, with the first seven days free (pricing). It doesn’t make Carbon faster, or add it to an account Google hasn’t given it to.

We’ll update this memo if Google says more about Carbon’s speed, or about this week’s reports.

8.Questions.

Why is Gemini 4 Carbon so slow?

Mostly because it does far more work per task. Kernel traced Carbon’s runs and found about seven times as many steps as GPT-6 Astra, half of them on work the bug report never asked for, with slower inference explaining 16% to 30% of the gap. Google says Carbon “spends more thinking tokens per turn,” which “increases per-turn token use and latency.”

Is Gemini 4 Carbon down, or just slow?

Usually just slow, but some Google AI Ultra subscribers report more than that. Since November 16, threads on Google’s forum describe Carbon and Argon stuck on “Working,” failing with “Model unavailable,” or answering at once without thinking. Google hasn’t replied in them, and its AI Studio status page lists no incident since October 30.

Does a lower thinking level make Gemini 4 Carbon faster?

It brings the first answer sooner but barely shortens whole tasks. Artificial Analysis measures Carbon’s first answer token at 8.74 seconds on low against 74.86 on high, yet 29.8 minutes per task on low against 34.5 on high. Kernel found that on low, Carbon “often failed to produce a fix at all.”

What should I use instead of Gemini 4 Carbon in Antigravity?

For day-to-day work, Orchestrator, which Google’s docs call the “recommended default.” Carbon plans and Gemini 4 Hydrogen makes the edits, and Google says it finished 19% faster than Carbon alone on Terminal-Bench 4.1. Hydrogen on its own suits routine tasks, and Ultra and paid Pro also offer Claude Opus 5.5 and Sonnet 5.5, on a separate quota.

Is Gemini 4 Carbon slower than Claude Opus 5.5 and GPT-6.2 Astra?

Yes, on every public measure so far. Theo timed one job at 22 minutes 21 seconds on Carbon at low and 9 minutes 33 seconds on Opus 5.5 at high. Artificial Analysis puts Carbon at 34.5 minutes per task at high, Opus 5.5 at 9.3 at high and 24.6 at max, and GPT-6.2 Astra at 5.3 at high.

Will Google make Gemini 4 Carbon faster?

Google hasn’t said. Its model card lists “occasional slowness or timeout issues” as a known limitation, and its Antigravity updates of November 16 and 17 stop long thinking from timing out at five minutes. That keeps slow turns from failing, but doesn’t make Carbon do less work.

Enjoy Agentic Industries · Checked against its sources on