Gemini 4 Carbon's Two Scores of 59: Intelligence and Coding Agents Differ

Gemini 4 Carbon led the overall intelligence evaluation with 59 points. It also scored 59 in a coding-agent evaluation, where it placed second. These numbers belong to different indices; neither is a 59% success rate.

Artificial Analysis's October 28 report makes that distinction explicit. Google's announcement that day describes a staged rollout beginning with selected cyber defenders and trusted partners. This is an October 30 source review, with no Carbon calls or independent benchmark reproduction.

Write the unit of evaluation beside the score

This table restates the evaluator's October 28 announcement, not a live leaderboard or a CodeBridge measurement.

The first asks about capabilities across tasks. The second evaluates a coding system in a particular execution environment. Such a system includes tools, observations, retries, context management and stopping rules alongside the model. Equal-looking scores do not make the measured objects identical.

The gap between 62 and 59 is not a three-percentage-point success-rate gap. The announcement alone also does not establish statistical significance or superiority on your repository. That requires task-level results and uncertainty estimates.

Holding the harness fixed answers only one kind of decision

Two comparisons serve different purposes.

To choose a product available today, testing each product with its own harness can be appropriate. Users receive the whole system, especially when they cannot swap its components. The conclusion must remain about those products, configurations and tasks.

To isolate a model substitution, keep the harness, tool schemas, budget, environment and workload fixed. If both models are not supported in the same harness, do not manufacture a comparison. Settings called high and max do not establish equal compute budgets either.

The published ranking can narrow a shortlist. It cannot isolate the causal effect of changing only the model. Our model versus harness article develops that distinction.

What the 23% cost difference actually measures

Artificial Analysis reports standard-price Intelligence Index task costs of $7.76 for Carbon and $10.04 for Opus 5.5 max. Save this arithmetic as cost.mjs and run node cost.mjs with Node.js 18 or later. No packages or API calls are needed.

const carbon = 7.76;
const opus = 10.04;
const reduction = 100 * (opus - carbon) / opus;
console.log(`${reduction.toFixed(1)}%`);

Actual local output:

22.7%

This reproduces only the relative difference between two published costs. It does not reproduce the evaluation or measure a 22.7% coding-cost saving. Mixing discounted and standard prices, or dividing this cost by a Coding Agent Index score, combines incompatible measurements.

Failed development attempts also cost money. On a fixed workload, divide total spend across all attempts by the number of tasks passing external verification. With zero successes, report the metric as undefined. State separately whether review time and test infrastructure are included.

Prepare the comparison record before access arrives

This is a proposed CodeBridge record, not an observed result or an API configuration.

comparison_goal: choose an available coding product
task_id: regression-017
repo_commit: <fixed commit>
environment_digest: <pinned test image>
system: <model + harness version + reasoning setting>
budget: <same time limit, cost ceiling, retry limit>
acceptance: <failing regression plus preserved behavior>
result: <pass / fail / blocked; independent test evidence>
spend: <all attempts, cache and discount rules included>

For a regression repair, prepare both the failing test and valid behavior that must survive. Keep final verification separate from the agent's execution so test edits or misreported tool errors cannot define success. A model name and final diff alone will not reconstruct failed attempts or their costs.

This is a future evaluation procedure. No dataset or model evaluation was run here. We have not invented a restricted-access command, model ID, hardware configuration or unpublished execution budget.

Carbon's announcement can change the shortlist. Changing a team's tool requires access, a usable harness, a verifiable workload and a budget together. The different denominators in its security evaluations are covered in our existing Carbon security article.

The published paid course below covers execution, verification and loops around models. It does not provide Carbon access or promise reproduction of this evaluation. The link is promotional.