What happened
Artificial Analysis announced Intelligence Index v5 on October 27, 2026. The company describes the Intelligence Index as a composite measure of model capabilities across tasks intended to reflect real-world usage. The central methodological change is that 60% of the v5 score now comes from evaluations using private test sets or held-out answers. Artificial Analysis also launched Progress Index v1, a separate measure intended to track model capability over time on one continuous scale.
The v5 composite assigns equal 25% weights to four categories. Knowledge Work consists of AA-Briefcase at 15%, AutomationBench-AA at 5%, and GDP.pdf at 5%. Coding consists of FrontierCode 2.0 at 15% and Terminal-Bench 4.0 at 10%. Science consists of Terminal-Bench-Science at 15% and a hard CritPt v2 subset at 10%. General capability consists of AA-Omniscience v1.1 at 15%—10% accuracy and 5% non-hallucination—and HLE-Diamond at 10%. GDPval-AA v2.1, SciCode, and AA-LCR v1.1 have been removed.
Two benchmarks are new to the Index. Cognition's FrontierCode 2.0 contains 132 tasks in real open-source repositories and grades an agent on whether a repository maintainer would merge its change. A failed blocking check receives a zero. The tasks remain private, and Artificial Analysis says it runs each task three times in its Stirrup agentic harness. Terminal-Bench-Science contains 70 scientific-research tasks performed in a terminal environment; experts curate and review the tasks, and a task passes only when every test passes.
Three elements have been upgraded. HLE-Diamond replaces Humanity's Last Exam with a refined 1,000-question version split evenly between reasoning and knowledge. AA-Omniscience v1.1 removes saturated and public-set questions, adds questions, and revises ambiguous or erroneous ones; Artificial Analysis has also released a recalibrated public subset on Hugging Face. CritPt v2 is represented by a private 40-question hard subset, derived from 70 research-level physics questions after excluding 30 solved by seven frontier models in all five launch attempts. Models use an offline Python sandbox in Stirrup for that evaluation.
On the overall v5 index, Artificial Analysis reports Claude Opus 5.5 (max) at 58, GPT-6 Astra (max) at 57, and GPT-6.1 Sol (max) at 55. It says Anthropic and OpenAI occupy the top five positions, with Gemini 4 Argon (high) next at 50. No single model leads all nine evaluations: Claude Opus 5.5 leads FrontierCode 2.0, HLE-Diamond, and AA-Omniscience v1.1, while GPT-6 Astra leads Terminal-Bench-Science and the CritPt v2 hard subset. The reported cost per Intelligence Index task among the ten highest-scoring displayed models ranges from $0.96 for GPT-6.1 Terra (max) to $13.04 for Claude Fable 5.1 (max).
What the archive adds
This is a new version of a known evaluation story, rather than a wholly new benchmark category. Artificial Analysis launched the Decision Index v1.0 treated bounded decisions as a distinct task class. Our note on FrontierCode 2.0 and Terminal-Bench-Science recorded the two evaluations that now enter this composite. A related note argued that component tasks and scoring rules make an aggregate more inspectable, without proving that its mix matches a deployment: Artificial Analysis publishes per-benchmark results.
Trend signal
The release supports verification bandwidth is the scarce engineering resource. FrontierCode's blocking checks and maintainer-merge criterion, plus Terminal-Bench-Science's all-tests-pass rule, make acceptance evidence part of the measurement rather than treating plausible generated output as success. That does not establish that the benchmarks predict production outcomes, but it reinforces the value of explicit verification contracts.
Why now
Artificial Analysis frames v5 as its biggest upgrade since v4 launched in January. Its stated rationale for increasing private and held-out material is implicit in the design: private tasks and answers can remain outside public benchmark exposure, while revised datasets remove saturated questions. The release also arrives with newly launched FrontierCode 2.0, allowing the composite to incorporate that coding evaluation immediately.
Connections to other events
The new composite directly incorporates the two task suites covered in New benchmarks: Cognition's FrontierCode 2.0 and Terminal-Bench-Science. It also sits beside Artificial Analysis's Decision Index, where the publisher separated classification, tool calling, knowledge, and reasoning rather than relying only on a broad aggregate. Together, these releases point to a growing set of narrower task measurements feeding higher-level comparisons.
Why it matters
Teams frequently need a short list of models before they can run their own evaluations. v5 offers a more structured input to that process because it reports task areas, component weights, private-set use, benchmark-specific leaders, and a per-task cost comparison. It also shows why an overall rank should not be a deployment decision. Claude Opus 5.5 leads the composite, while GPT-6 Astra leads the reported science and physics subsets; other models lead individual evaluations. A team building coding automation, terminal-based research workflows, or knowledge-heavy assistants should therefore look at the relevant component and then test its own acceptance criteria. The reported 13-fold-plus cost range among leading models adds an operational constraint: a marginal score difference can come with a materially different task budget.
What happens next
- Benchmark publishers will continue replacing public or saturated items with private or held-out material. (6 months, 65%). Watch for: revised test sets and explicit contamination controls.
- Teams using the composite will compare component scores and cost per completed task, not only the overall rank. (3 months, 60%). Watch for: vendor evaluation reports broken out by workload.
Limits and uncertainty
The source is Artificial Analysis's own announcement and methodology description; it does not independently validate the task selection, scoring, harness configuration, or the relation between these scores and a particular production system. Private datasets may reduce public exposure, but they also limit outside inspection of individual items. The announced figures apply to named model configurations and effort settings, not to every API configuration or agent harness. Cost per task is reported for the displayed leading models, but the source does not establish a team's total operating cost, including retries, tool calls, engineering time, or human review. The open-weight comparison also distinguishes a model whose weights are not yet released from models with released weights.
Skill to build now
Create a small evaluation pack from work your team already must accept: a coding change with blocking checks, a research task with required artifacts, and a knowledge task that penalizes unsupported claims. Run candidate models more than once, record pass criteria before prompting, and measure both pass rate and end-to-end cost. Keep the tasks private where possible, but document their category, success rule, and failure modes so the comparison remains useful as models and prompts change.
Reader takeaway
Intelligence Index v5 is most useful as evidence about evaluation design, not as a universal answer to which model to buy. Its private and held-out task share, all-pass checks, and workload-specific results make the ranking more informative than a single public quiz. Still, the relevant benchmark is the one that reproduces your own definition of a completed task.
Open question
Will v5's private-task mix predict outcomes on independently run production workloads better than the public-heavy versions it replaces?