Benchmarks · checked 2026-11-18
Gemini 4 Neon benchmarks: index and vendor claim, side by side
One composite index from a third party, one win count from the model maker. Same model, different test sets, different owners — so they stay in separate columns with separate sources.
Two columns, because they measure different things
The left column is a third party's composite index. The right column is the model maker's own claim. They are printed side by side only so you can see who said what — never added together, averaged, or blended into a sentence.
Third party
Artificial Analysis Intelligence Index: Gemini 4 Neon (xhigh) = 61, third overall, behind GPT-6.2 Astra and Claude Fable 5.2.
What it is: one composite number built by Artificial Analysis from its own fixed evaluation set, weighted by its own method.
Source: artificialanalysis.ai/articles/gemini-4-neon-intelligence-gain, checked 2026-11-18.
Vendor
Google's claim: Gemini 4 Neon wins 16 of 19 tests in the comparison Google published.
What it is: a win count over a test selection and scoring method chosen by the model maker. It is not an index, and 16 of 19 is not a score out of 100.
Source: Google's launch material as summarised on llm-stats.com/models/gemini-4-neon, checked 2026-11-18.
Why the two numbers cannot be merged
They differ in three ways that no arithmetic can reconcile.
- Who chose the tests. The index uses a fixed set that its publisher maintains across models so numbers stay comparable over time. A vendor comparison uses tests the vendor selected for its own announcement.
- How results turn into one number. An index applies a published weighting to many sub-scores. A win count treats every test as equal and discards the margins: winning by one point and winning by forty both count as one win.
- What happens next time. The index is re-run as new models ship, so the 61 and the rank can move without any change to Gemini 4 Neon itself. A win count belongs to the day of the announcement and is never revisited.
A sentence like “61 and 16 of 19” is therefore not a summary. It is two different claims from two different owners, and any sentence that fuses them is making a third claim that neither source supports.
What the index figure does not tell you
The index is a capability summary. It says nothing about these, all of which change a bill:
- Price per token. A model that scores higher can cost several times more per million output tokens. The price matrix carries both rates.
- Latency and throughput. Not part of the index, and not published as a single number for these models in the sources checked.
- Context window in practice. Gemini 4 Neon publishes 1.0M tokens. The index does not measure how well a model performs at the far end of its window.
- Cost per task. This is a different metric from price per token: the same task can need very different token counts from different models. This site does not compute it, because doing so honestly would need per-task token counts from a controlled run, and we have not run one.
Where the benchmark claim sits next to the price
At list prices verified 2026-11-18, and at the site's standard volume of 10,000,000 input plus 2,000,000 output tokens a month with no cache hits:
| Model | Input $/1M | Output $/1M | Monthly cost | AA index | Vendor claim |
|---|---|---|---|---|---|
| Gemini 4 Neon | $3.00 | $15.00 | $60.00 | 61 | 16 of 19 tests won |
| Claude Opus 5.5 | $4.00 | $20.00 | $80.00 | not published here | not published here |
| Claude Opus 5.5 (fast) | $8.00 | $40.00 | $160.00 | not published here | not published here |
Those two “not published here” cells are a limit of this page, not a claim that no figure exists: we have not verified an index value or a vendor claim for Claude Opus 5.5, so the cells stay empty rather than carrying a number we did not check. The cost column is arithmetic on the two published rates, nothing else.
Reading a 61 from a composite index
A composite score is a weighted average, and a weighted average moves when the weights move. Three questions worth asking of any index before quoting it:
- Which version. Artificial Analysis publishes versions of its index; a score belongs to the version it was calculated against.
- Which effort setting. “xhigh” in the figure above is an effort setting, so 61 describes that setting, not every way the model can be run.
- Which comparison set. “Third overall” describes a position within a list that grows every time a model ships. It is a position on a date, not a fixed property.
The same discipline applies to the 16 of 19: it is a statement about 19 particular tests, chosen by Google, on one day. Fewer tests, a different selection, or a different tie rule would all produce a different count.
Questions about the two figures
Does this site compute its own benchmark score?
No. We publish other people's numbers with the owner's name attached. A score of our own would need a weighting we could not justify from published data, and it would hide exactly the disagreement this page tries to show.
Should the index or the vendor claim be trusted?
Treat each as a claim from an interested party with a stated method. The index publisher sells a comparison service; the vendor sells the model. Both are checkable, which is why the sources are linked rather than summarised.
Why put prices on a benchmark page at all?
Because a capability number with no price next to it cannot answer the question most readers arrive with: what does using this cost me per year. The arithmetic is on the home page calculator; the rates are on the price matrix.
Will the numbers on this page change?
Yes. Index values are recalculated as models ship, and vendors restate claims. Every figure here carries the date it was checked; a figure without a date would be unusable.
Prices and benchmarks verified 2026-11-18. ModelCost is not affiliated with Google or Anthropic.