Gemini 4 Carbon Broke the Robot Arm: What a 7-Hour 33-Minute Lab Run Says Before the Score Exists

There is no score yet. That is the first thing to understand about the Gemini 4 Carbon robotics result, and almost everything else in this piece follows from it. What exists, as of 13 November 2026, is a single post from Artificial Analysis on 12 November stating that Gemini 4 Carbon spent 7 hours and 33 minutes inside the AA-Robotics v0.1 lab, "bashing the arm around so much it broke a gripper," with the added note that the lab gave it a retry on that attempt. The five-model leaderboard published on 14 October still lists the same five rows it listed then — GPT-6 Astra 69.4, GPT-6.1 Sol 64.6, Gemini 4 Argon 55.3, Claude Opus 5.5 48.8, Claude Fable 5.1 46.3 — and Carbon is not among them. No aggregate, no per-task breakdown, no grasp rate, no bill.

So treat this as an early signal rather than a result: a duration, a hardware failure, and one retry, from the lab that owns the benchmark, published before its own table has been updated. Everything below that goes beyond those three facts is inference, labelled as such. The reason it is still worth reading now is that Carbon's predecessor, Gemini 4 Argon, is the slowest model in that lab by a wide margin, and 7 hours 33 minutes lands almost exactly on top of Argon's number — which makes the duration readable even without a score.

The only public number is a clock

Artificial Analysis describes AA-Robotics as measuring "how well language models understand, plan and act in the physical world by controlling a robot arm." The rig is a 6-axis Hiwonder NexArm with five joints and a gripper, watched by four cameras — three on the table, one on the wrist — across eight tabletop tasks: press a red button, put a crocodile toy in a rectangle, pour a drink into a cup, tidy blocks into a box including its lid, finish a shape-matching board, write "hello world," build the tallest foam tower, and measure a box to within 2 mm per dimension. Attempts are zero-shot, each scored 0 to 100 with partial credit. Models are dropped in through their own coding harnesses at high effort — Codex for the GPT models, Claude Code for Claude, Antigravity for Gemini — and have to write the Python that plans and executes the motion themselves.

Against that, the Carbon datapoint is thin on purpose. Seven hours thirty-three minutes is 453 minutes. It is not stated whether that covers all eight tasks, whether it includes the retried attempt, or what the run cost. Anyone quoting it as a performance figure is reading a stopwatch as a scoreboard.

What it does say, unambiguously, is that Carbon was put through the lab and the equipment did not survive intact. The benchmark's own publication documents dozens of failure modes — empty grasps, dropped bottles, blocks nudged off the table, a pen tested on the page itself — and not one of them is a broken gripper. Hardware destruction is a category the October write-up never had to describe, because none of the five models in it produced one.

Seven hours thirty-three, measured against Argon

The baseline matters more than usual here, because the comparison is unusually clean. Artificial Analysis published lifetime run times for each of the five October models: GPT-6 Astra 242 minutes, GPT-6.1 Sol 224, Claude Opus 5.5 253, Claude Fable 5.1 249, and Gemini 4 Argon 419. Carbon's 453 minutes sits 34 minutes above Argon — about 8% longer — and roughly 200 minutes above the four non-Gemini models, which cluster in a 224-to-253-minute band.

• Per-task pace, if the session covered all eight tasks — Carbon about 56.6 minutes per task versus Argon's 52.4, Astra's 30.3 and Sol's 28.

• Against a human — AA puts people at roughly 20 seconds per task, so 453 minutes across eight tasks is about 170 times slower than a person doing the same tabletop work.

• Share of the clock spent planning — Argon's split was 42% understanding, 45% planning, 6% cameras, 6% arm. Astra and Sol inverted it, spending 13% of the run on actual arm motion and the majority on planning.

That last line is where the broken gripper becomes interesting rather than funny. In the October data, the two highest-scoring models (Astra at 69.4 and Sol at 64.6) spent the least time understanding the scene and the most time planning; Argon, the slowest and third-best, spent 42% of its run just looking. A model that deliberates longer and then damages the hardware is not exhibiting the same failure the October mistake wall described. The wall is full of models that misjudged state — "reports the lid level and aligned while it rests on a wedge," "counts a tilted yellow hexagon as placed." Those are perception and self-assessment errors. Bashing an arm until a gripper breaks is a command-authority error: the model asked the hardware for something the hardware could not take. Whether Carbon did that by driving into a collision limit, over-torquing a grasp, or simply repeating a motion it had already been told failed, AA has not said, and the 11-second clip attached to the post does not settle it.

The retry is the methodological question, not a footnote

"We gave it a retry on that one" is the most consequential half-sentence in the post, and it deserves more scrutiny than the broken gripper. AA-Robotics is built on zero-shot attempts — the October publication says so explicitly, and the whole point of a zero-shot physical benchmark is that the model gets one go, as a person would. Granting a redo after equipment failure is defensible: a snapped gripper is the lab's problem, not a measure of the model's planning, and scoring the destroyed attempt as a zero would punish the rig's fragility.

But it creates a comparability problem the eventual score will have to answer. If exactly one attempt was voided and re-run, the aggregate is comparable to the other five and the retry is a footnote. If the retry was granted more than once, or only on tasks where Carbon had already failed for model-side reasons, the number that lands on the leaderboard will not mean what the other five numbers mean. The honest thing to look for when the table updates is a line stating how many attempts were voided, on which task, and on what criterion — the same way the October post documented that Argon's grasp sample was n=75 against Astra's n=95.

There is a second, quieter measurement problem. Argon's timings in the October publication are explicitly estimates: Antigravity, the Gemini harness, does not timestamp steps, so AA interpolated them by tokens between timed arm commands. Carbon runs in the same harness. If the 7-hour figure was measured the same way, it is not directly comparable to the minute counts generated by Codex and Claude Code — and the 34-minute gap over Argon is well inside the error a token-interpolation method could introduce over a 453-minute run. That does not make the duration meaningless. It makes "Carbon is 8% slower than Argon" a claim the data does not support, even though "Carbon took a very long time" is safe.

What Carbon's text profile would predict

None of this is surprising against Carbon's published behaviour, which is the part we can check independently. On Artificial Analysis's v5 index the model scores 60, fourth of 78, and its profile is that of a deliberate, verbose reasoner rather than a fast one: 28.4 minutes per task against a 13.0-minute median, 72 output tokens per second, 64.71 seconds to the first answer token, and 205 million tokens of output across the index — described by AA as "very verbose" against a 72.8-million median. It also costs $5.48 per index task, ranking 59th of 77 on cost efficiency.

Read those together and the lab result has a shape before the score arrives. A model that spends a minute before its first token on a text prompt, emits nearly three times the median volume, and already took 28 minutes per task on a static benchmark is not going to be quick in a loop where every step requires reading four camera feeds and then writing motion code. The 7-hour figure is what Carbon's text profile looks like when it is attached to a robot arm. Whether that deliberation produced better manipulation is exactly what the missing score would tell us, and the broken gripper is at least weak evidence that more thinking did not translate into gentler acting.

What a run like this costs, and why the calendar matters

Cost per task is the number AA has not published for Carbon, and guessing it would be worthless because the existing spread is enormous: GPT-6 Astra's 242-minute run cost $16.36 per task, GPT-6.1 Sol's cost $1.93 — an 8.5x gap between two models from the same vendor doing the same eight tasks. Across eight tasks that is roughly $131 against $15. About 70% of the spend went to input and context rather than output, which is what four camera feeds and a growing transcript do to a token bill. Carbon is more verbose than either on text, so if anything its per-task number should land high — but that is a prediction, not a figure, and the only honest use of it is as a budgeting range rather than a quote.

The calendar is the part you can act on. Carbon is at $2 per million input and $10 per million output tokens until 30 November 2026; on 1 December those become $4 and $20 with the model unchanged. Cached input moves from $0.10 to $0.20, batch and flex run at half rate, priority at 1.8 times standard. Any long agentic loop of the kind AA just ran is mostly cached context and repeated camera frames, which means a robotics-style workload is precisely the shape that benefits most from the current rate and hurts most after it. Re-running this experiment in December costs twice what it costs this week.

How to read the number when it lands

When AA updates the table, three things decide whether Carbon is a step forward or another Argon.

• The aggregate against 55.3 — Argon's score. A Gemini model at or below its own predecessor, one release later, would make the family's robotics story a story about harness latency rather than capability.

• Grasp success against the 28-to-45% band: Astra 45% (n=95), Fable 44% (n=89), Opus 40% (n=100), Argon 36% (n=75), Sol 28% (n=111). This is the metric that would register a model that breaks grippers, because a command that damages hardware and a grasp that misses the object come from the same absence of force feedback.

• The per-task profile, task by task. Argon's was 100, 100, 100, 40, 0, 45, 57, 0 on button, crocodile, pour, tidy, shapes, write, tower, measure. The measure-the-box task is the one nobody solves: Astra 0, Argon 0, Fable 0, with only Sol and Opus at 17. A 2 mm tolerance across three dimensions is a precision task, and precision is where a model that bashes an arm should show up worst.

There is also a fair question about what the benchmark is really scoring. Carbon ran inside Antigravity at high effort, as Argon did. A 7.5-hour session is a joint test of the model's planning and the harness's ability to keep a long loop coherent — and the October data already showed harness-shaped artefacts, from Argon's 25-minute first action to Fable's ten minutes of study before touching anything. Carbon's result will be quoted as a model number regardless. It is at least partly an agent-loop number.

If you want to reproduce the experiment instead of waiting

The useful thing about a public benchmark with a public rig is that the score is not the only way to learn from it. Carbon is callable now — Google made gemini-4-carbon generally available on the paid Gemini API on 10 November 2026, after announcing it to a gated cyber-defender cohort on 28 October — and it is on OrcaRouter as google/gemini-4-carbon at Google's own $2/$10 introductory rate, on the same OpenAI-compatible endpoint as the other 200-plus models on the key. Because we pass provider list price through with 0% markup, the 1 December repricing lands on our side the day Google applies it, rather than after a resale cycle.

Two practical notes before anyone points a fresh model at expensive hardware. First, a first-week endpoint is not a production endpoint: our own aggregated figures for google/gemini-4-carbon currently read a 30.2% error rate and a 6.89-second p50 time to first token over the last seven days, measured across the provider path we route. That is normal for a model three days into general availability and it is also exactly why you do not hard-code one endpoint into a loop that runs for seven hours — automatic failover across providers is what keeps a long agentic session from dying at hour four. Second, the routing DSL lets you keep a cheap model in the loop for the steps that do not need Carbon, which matters at 453 minutes: most of that session was reading frames and re-planning, not deciding anything hard.

What would change this read is not a better clip of the arm. It is AA publishing a score with the voided attempts documented, and a grasp rate next to the other five. Until then the accurate summary is narrow: Gemini 4 Carbon is the first model in that lab to cost the lab a piece of hardware, it took longer than anything the lab has measured except its own predecessor, and the retry that followed is a detail the eventual number will have to carry. The 12 November post is a real datapoint about how this model behaves when it is given hands. It is not yet a result about how well it uses them.