Intelligence Index v5: why the new yardstick matters

The Intelligence Index v5 is good screening and a weak basis for purchase approval. The yardstick changed. On October 27, 2026, Artificial Analysis altered the test, and anyone who compares an old score with a new one risks buying the difference between two rulers rather than the difference between two models.

Facts for the decision

• The Intelligence Index v5 splits the score into four categories of 25% each.
• Evaluations with private tests or held-out responses weigh 60% of the total.

The yardstick holds for comparisons within the same version.

• A v4.3 score does not measure the same set as a v5 score.
• The aggregate score is mainly textual and in English; language and modality are evaluated separately.

Why does the Intelligence Index v5 change the buying decision?

The Intelligence Index v5 requires a fresh comparison of candidates because Artificial Analysis changed tests and weights on October 27, 2026. A v4.3 score does not measure the same set as a v5 score. The version comes before the ranking.

The method change is concrete. The Artificial Analysis announcement splits the score across knowledge work, coding, science and general, at 25% per category. FrontierCode 2.0 and Terminal-Bench-Science entered at 15% each; GDP.pdf fell from 10% to 5%. GDPval-AA v2.1, SciCode and AA-LCR v1.1 left the composition. HLE gave way to HLE-Diamond; AA-Omniscience moved to v1.1 and CritPt to the hard private subset of v2. On October 27, 2026, Claude Opus 5.5 (max) led the announcement with 58 points; an October 28 update credited 59 to Gemini 4 Carbon (high). No record proves current leadership. By November 5, the methodology already carried v5.0.1, with an improved CritPt v2 correction and a re-evaluation of models on that benchmark. A v5.0 score is not necessarily the same under v5.0.1.

The history of the AI model ranking helps separate capability from measurement.

The yardstick is not the job.

What do the 60% say about the Intelligence Index v5?

In the Intelligence Index v5, 60% is the weight of evaluations with private test sets or held-out responses, not the share of questions that are entirely private or unseen. The number does not prove the absence of contamination. Nor does it prove full reproduction by a buyer.

The intelligence methodology of the Intelligence Index v5, in version v5.0.1, separates what the index covers from what stays outside the score. Inspection has limits, and restricted access does not let a buyer reproduce every test. The methodology identifies Terminal-Bench-Science as not private even though the benchmark sits in the composition. Material that is not accessible restricts independent reproduction, and a private-test label is not an audit of grading. The aggregate score is mainly textual and in English; the source evaluates multilingual, visual and speech performance separately. Evaluating AI models for a Brazilian application therefore has to include Brazilian Portuguese when the workload requires that language. It also has to check the failure types the operation accepts, because the public result does not establish, by itself, the quality of answers inside the buyer's service.

Which benchmark measures a similar job?

In the Intelligence Index v5, benchmarks measure different jobs even when they enter the same average. FrontierCode 2.0 evaluates changes in open source repositories; Terminal-Bench-Science gathers scientific research in a terminal. Each weighs 15% of the composition, and neither guarantees an equivalent result on the buyer's workload.

FrontierCode 2.0 gathers 132 tasks, with private questions and three runs per task; one failure in blocking verification zeroes the attempt. Terminal-Bench-Science gathers 70 tasks and is not private according to the methodology. How FrontierCode grades matters. In the October 27, 2026 announcement, Claude Opus 5.5 (max) reached 46% on FrontierCode 2.0 and 59% on Terminal-Bench-Science. DeepSeek V4.1 Flash 1017 (max) scored 28% and 13% respectively. The results describe the tests, not the probability of success for an enterprise application. FrontierCode matters to whoever changes code; Terminal-Bench-Science, to researchers in a terminal. Evaluating AI models on the router also requires tying the metric to the job instead of transferring a general score straight into a production policy. For Portuguese, voice or image, the aggregate score has a clear limit.

What does the distance reveal about open-weight models?

The 3 to 12 point difference published by Artificial Analysis in the Intelligence Index v5 compares the leader and the best model classified as open weights across v4.3 and v5. On October 27, 2026, the candidate in the 12 point comparison did not have released weights yet.

On October 27, 2026, Kimi K3.1 Pro scored 46 points against 58 for Claude Opus 5.5 (max), but its weights were not released. DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) had released weights and scored 32 points each, 26 behind the leader. Those values are historical. They do not describe a verified current ranking. The 3 to 12 distance uses versions with different compositions, so it does not prove that the methodological change caused the whole widening. Nor does it turn 26 points into an error rate, lost productivity or the cost of a failure. Open weights and hosted models present distinct operating choices, but released weights are not an unrestricted license and do not prove lower cost. Before excluding a candidate, the team has to check license, availability, total cost and performance on the real task, with acceptance criteria defined in advance.

How to review AI model routing?

AI model routing should change after a comparison on the same basis and an evaluation of your own, not because an isolated score went up. The Intelligence Index v5 record has to identify version and configuration; without comparable scores, a controlled evaluation is the next step.

The policy changes when the comparison uses the same basis and the evaluation meets the criteria of the service that will receive the answers. The 25% the Intelligence Index v5 assigns to each category reflects the methodology's choice, not the company's distribution of work. Mixing versions produces a specific error. The difference between two scores can come from the yardstick, not from the model, and the switch starts answering to a method change. The Intelligence Index v5 offers public signals about the 132 FrontierCode 2.0 tasks. The policy, however, has to cover latency, failure types and total cost on your own workload, with Brazilian Portuguese when the service requires that language. When comparable results are missing, the gap stays on the record instead of being filled by a score from another version of the index. Nexforce Router is routing infrastructure; the index does not show that the product integrates the ranking or runs the tests automatically.

1. Record source, date, version, score and previous configuration.
2. Compare candidates on the same version or in the same controlled evaluation.
3. Measure quality and latency on your own workload, with explicit failure types and acceptance criteria.
4. Check total cost, availability and license before deciding.

Source: Progress Index methodology and the article sections. Art: Nexforce.

What are the frequently asked questions about the Intelligence Index v5?

The Intelligence Index v5 organizes a candidate list; it does not validate a model switch on its own. A score authorizes a decision only when version, configuration and date are recorded and the workload has been measured against defined criteria.

Anyone using the Intelligence Index v5 in a purchase decision has to record four things before comparing any score: the index version, the date of the result, the model configuration and the test that produced the number. Without that record, the comparison mixes rulers and produces a conclusion the source does not support. The next step is to measure the workload, with the failure types the operation accepts and the total cost per task. The public score serves to build the candidate list and to order the tests, not to justify the switch. When a comparable result from the same version is missing, the gap should stay on the record, and the decision then depends on the controlled evaluation, with acceptance criteria defined before the measurement and filed with the original score.

<!-- FAQPage -->

Can a v4.3 score be compared directly with v5?

No. V5 replaced evaluations and weights, so a v4.3 score and a v5 score measure different sets. Comparing the two requires recording version, date and configuration for each result. Without that record, the difference between the scores can come from the yardstick, not from the model.

Do the 60% mean that every question is private?

No. The 60% is the weight of evaluations with private test sets or held-out responses in the Intelligence Index v5. The number does not say how many questions are unseen, and it does not prove the absence of contamination. The methodology itself identifies at least one non-private benchmark among those that compose the score.

Do the 12 points represent the best open weights released so far?

No. On October 27, 2026, the model that appeared 12 points behind the leader did not have released weights yet. The models with released weights cited in the announcement were 26 points behind. The comparison is historical and does not describe the current ranking of the Intelligence Index v5.

Which index deals with progress over time?

The Progress Index, from the same organization, estimates a common scale for comparing models over time. The Intelligence Index v5 measures models on current uses and does not form a direct series across versions. For routing decisions, neither of them replaces a controlled evaluation of your own workload.

References and Further Reading

The four primary Artificial Analysis pages support the composition of the Intelligence Index v5, the dated results of October 27, 2026 and the difference between comparing models and tracking progress over time. None of them carries independent measurement.

The Artificial Analysis announcement of October 27, 2026 presents v5.0 and the historical results cited here. The October 28 update records another leader, not a current ranking. The intelligence methodology, consulted up to v5.0.1 on November 5, lists the tests, the weight per category and the language and modality limits, and records the rescoring on CritPt v2. The Progress Index describes why scores from different versions do not form a comparable series and proposes a common scale for temporal analysis. To take the analysis into the operation, the reading on AI cost per task covers the cost per attempt, which the aggregate score does not measure. Every result has a date, and the date changes the reading of any score, including another score from the same version.

Which evidence should authorize the next switch?

The next routing change needs comparable results and a workload that meets the company's criteria. The Intelligence Index v5 guides the initial selection; it does not decide the policy on its own. Without scores from the same version, a controlled evaluation is the appropriate basis.

Recording the version and configuration of the Intelligence Index v5 prevents comparing v4.3 and v5 scores as results of the same test. The distinction also holds inside the v5 family: the announcement of October 27, 2026 describes v5.0, while the methodology consulted on November 5 already carried v5.0.1 and the improved CritPt v2 correction. The score is not the contract. Quality, latency, total cost, availability and license have to be examined on your own workload, with Brazilian Portuguese when needed. If the index does not offer comparable scores, the controlled evaluation of your own provides another basis for comparing candidates, provided the team defines acceptance criteria and failure types. Without enough evidence from that evaluation, there is no case for the switch. The first decision fits in a spreadsheet cell; the second has to fit the operation.