The arms race to construct smarter AI fashions has a measurement downside: the checks used to rank them have gotten out of date virtually as shortly because the fashions enhance. On Monday, Synthetic Evaluation, an unbiased AI benchmarking group whose rankings are carefully watched by builders and enterprise patrons, launched a significant overhaul to its Intelligence Index that essentially modifications how the trade measures AI progress.
The brand new Intelligence Index v4.0 incorporates 10 evaluations spanning brokers, coding, scientific reasoning, and common data. However the modifications go far deeper than shuffling check names. The group eliminated three staple benchmarks — MMLU-Professional, AIME 2025, and LiveCodeBench — which have lengthy been cited by AI corporations of their advertising and marketing supplies. Of their place, the brand new index introduces evaluations designed to measure whether or not AI methods can full the form of work that folks really receives a commission to do.
sort: embedded-entry-inline id: 1bCmRrroGCdUb07IuaHysL
“This index shift displays a broader transition: intelligence is being measured much less by recall and extra by economically helpful motion,” noticed Aravind Sundar, a researcher who responded to the announcement on X (previously Twitter).
Why AI benchmarks are breaking: The issue with checks that prime fashions have already mastered
The benchmark overhaul addresses a rising disaster in AI analysis: the main fashions have turn into so succesful that conventional checks can now not meaningfully differentiate between them. The brand new index intentionally makes the curve tougher to climb. Based on Synthetic Evaluation, prime fashions now rating 50 or under on the brand new v4.0 scale, in comparison with 73 on the earlier model — a recalibration designed to revive headroom for future enchancment.
This saturation downside has plagued the trade for months. When each frontier mannequin scores within the ninetieth percentile on a given check, the check loses its usefulness as a decision-making instrument for enterprises attempting to decide on which AI system to deploy. The brand new methodology makes an attempt to resolve this by weighting 4 classes equally — Brokers, Coding, Scientific Reasoning, and Genera l— whereas introducing evaluations the place even probably the most superior methods nonetheless wrestle.
The outcomes beneath the brand new framework present OpenAI’s GPT-5.2 with prolonged reasoning effort claiming the highest spot, adopted carefully by Anthropic’s Claude Opus 4.5 and Google’s Gemini 3 Professional. OpenAI describes GPT-5.2 as “probably the most succesful mannequin collection but for skilled data work,” whereas Anthropic’s Claude Opus 4.5 scores increased than GPT-5.2 on SWE-Bench Verified, a check set evaluating software program coding skills.
GDPval-AA: The brand new benchmark testing whether or not AI can do your job
Essentially the most important addition to the brand new index is GDPval-AA, an analysis based mostly on OpenAI’s GDPval dataset that checks AI fashions on real-world economically precious duties throughout 44 occupations and 9 main industries. Not like conventional benchmarks that ask fashions to resolve summary math issues or reply multiple-choice trivia, GDPval-AA measures whether or not AI can produce the deliverables that professionals really create: paperwork, slides, diagrams, spreadsheets, and multimedia content material.
Fashions obtain shell entry and internet looking capabilities by way of what Synthetic Evaluation calls “Stirrup,” its reference agentic harness. Scores are derived from blind pairwise comparisons, with ELO rankings frozen on the time of analysis to make sure index stability.
Beneath this framework, OpenAI’s GPT-5.2 with prolonged reasoning leads with an ELO rating of 1442, whereas Anthropic’s Claude Opus 4.5 non-thinking variant follows at 1403. Claude Sonnet 4.5 trails at 1259.
On the unique GDPval analysis, GPT-5.2 beat or tied prime trade professionals on 70.9% of well-specified duties, based on OpenAI. The corporate claims GPT-5.2 “outperforms trade professionals at well-specified data work duties spanning 44 occupations,” with corporations together with Notion, Field, Shopify, Harvey, and Zoom observing “state-of-the-art long-horizon reasoning and tool-calling efficiency.”
The emphasis on economically measurable output is a philosophical shift in how the trade thinks about AI functionality. Fairly than asking whether or not a mannequin can cross a bar examination or resolve competitors math issues — achievements that generate headlines however do not essentially translate to office productiveness — the brand new benchmarks ask whether or not AI can really do jobs.
Graduate-level physics issues expose the boundaries of at present’s most superior AI fashions
Whereas GDPval-AA measures sensible productiveness, one other new analysis referred to as CritPT reveals simply how far AI methods stay from true scientific reasoning. The benchmark checks language fashions on unpublished, research-level reasoning duties throughout fashionable physics, together with condensed matter, quantum physics, and astrophysics.
CritPT was developed by greater than 50 energetic physics researchers from over 30 main establishments. Its 71 composite analysis challenges simulate full-scale analysis initiatives on the entry degree — akin to the warm-up workouts a hands-on principal investigator would possibly assign to junior graduate college students. Each downside is hand-curated to supply a guess-resistant, machine-verifiable reply.
The outcomes are sobering. Present state-of-the-art fashions stay removed from reliably fixing full research-scale challenges. GPT-5.2 with prolonged reasoning leads the CritPT leaderboard with a rating of simply 11.5%, adopted by Google’s Gemini 3 Professional Preview and Anthropic’s Claude 4.5 Opus Pondering variant. These scores counsel that regardless of exceptional progress on consumer-facing duties, AI methods nonetheless wrestle with the form of deep reasoning required for scientific discovery.
AI hallucination charges: Why probably the most correct fashions aren’t all the time probably the most reliable
Maybe probably the most revealing new analysis is AA-Omniscience, which measures factual recall and hallucination throughout 6,000 questions masking 42 economically related subjects inside six domains: Enterprise, Well being, Regulation, Software program Engineering, Humanities & Social Sciences, and Science/Engineering/Arithmetic.
The analysis produces an Omniscience Index that rewards exact data whereas penalizing hallucinated responses — offering perception into whether or not a mannequin can distinguish what it is aware of from what it does not. The findings expose an uncomfortable fact: excessive accuracy doesn’t assure low hallucination. Fashions with the best accuracy usually fail to guide on the Omniscience Index as a result of they have a tendency to guess moderately than abstain when unsure.
Google’s Gemini 3 Professional Preview leads the Omniscience Index with a rating of 13, adopted by Claude Opus 4.5 Pondering and Gemini 3 Flash Reasoning, each at 10. Nevertheless, the breakdown between accuracy and hallucination charges reveals a extra advanced image.
On uncooked accuracy, Google’s two fashions lead with scores of 54% and 51% respectively, adopted by Claude 4.5 Opus Pondering at 43%. However Google’s fashions additionally show increased hallucination charges than peer fashions, scoring 88% and 85%. Anthropic’s Claude 4.5 Sonnet Pondering and Claude Opus 4.5 Pondering present hallucination charges of 48% and 58% respectively, whereas GPT-5.1 with excessive reasoning effort achieves 51%—the second-lowest hallucination price examined.
Each Omniscience Accuracy and Hallucination Charge contribute 6.25% weighting every to the general Intelligence Index v4.
Contained in the AI arms race: How OpenAI, Google, and Anthropic stack up beneath new testing
The benchmark reshuffling arrives at an particularly turbulent second within the AI trade. All three main frontier mannequin builders have launched main new fashions inside only a few weeks — and Gemini 3 nonetheless holds the highest spot on a lot of the leaderboards on LMArena, a extensively cited benchmarking instrument used to match LLMs.
Google’s November launch of Gemini 3 prompted OpenAI to declare a “code purple” effort to enhance ChatGPT. OpenAI is relying on its GPT household of fashions to justify its $500 billion valuation and over $1.4 trillion in deliberate spending. “We introduced this code purple to actually sign to the corporate that we wish to marshal assets in a single explicit space,” mentioned Fidji Simo, CEO of purposes at OpenAI. Altman advised CNBC he anticipated OpenAI to exit its code purple by January.
Anthropic responded with Claude Opus 4.5 on November 24, reaching an SWE-Bench Verified accuracy rating of 80.9% — reclaiming the coding crown from each GPT-5.1-Codex-Max and Gemini 3. The launch marked Anthropic’s third main mannequin launch in two months. Microsoft and Nvidia have since introduced multi-billion-dollar investments in Anthropic, boosting its valuation to about $350 billion.
How Synthetic Evaluation checks AI fashions: A take a look at the unbiased benchmarking course of
Synthetic Evaluation emphasizes that every one evaluations are run independently utilizing a standardized methodology. The group states that its “methodology emphasizes equity and real-world applicability,” estimating a 95% confidence interval for the Intelligence Index of lower than ±1% based mostly on experiments with greater than 10 repeats on sure fashions.
The group’s printed methodology defines key phrases that enterprise patrons ought to perceive. Based on the methodology documentation, Synthetic Evaluation considers an “endpoint” to be a hosted occasion of a mannequin accessible by way of an API — which means a single mannequin might have a number of endpoints throughout completely different suppliers. A “supplier” is an organization that hosts and gives entry to a number of mannequin endpoints or methods. Critically, Synthetic Evaluation distinguishes between “open weights” fashions, whose weights have been launched publicly, and actually open-source fashions—noting that many open LLMs have been launched with licenses that don’t meet the total definition of open-source software program.
The methodology additionally clarifies how the group standardizes token measurement: it makes use of OpenAI tokens as measured with OpenAI’s tiktoken package deal as a typical unit throughout all suppliers to allow truthful comparisons.
What the brand new AI Intelligence Index means for enterprise expertise choices in 2026
For technical decision-makers evaluating AI methods, the Intelligence Index v4.0 gives a extra nuanced image of functionality than earlier benchmark compilations. The equal weighting throughout brokers, coding, scientific reasoning, and common data signifies that enterprises with particular use instances might wish to study category-specific scores moderately than relying solely on the combination index.
The introduction of hallucination measurement as a definite, weighted issue addresses one of the persistent issues in enterprise AI adoption. A mannequin that seems extremely correct however incessantly hallucinates when unsure poses important dangers in regulated industries like healthcare, finance, and legislation.
The Synthetic Evaluation Intelligence Index is described as “a text-only, English language analysis suite.” The group benchmarks fashions for picture inputs, speech inputs, and multilingual efficiency individually.
The response to the announcement has been largely constructive. “It’s nice to see the index evolving to scale back saturation and focus extra on agentic efficiency,” wrote one commenter in an X.com submit. “Together with real-world duties like GDPval-AA makes the scores way more related for sensible use.”
Others struck a extra bold observe. “The brand new wave of fashions that’s nearly to return will go away all of them behind,” predicted one observer. “By the tip of the yr the singularity might be simple.”
However whether or not that prediction proves prophetic or untimely, one factor is already clear: the period of judging AI by how properly it solutions check questions is ending. The brand new normal is less complicated and much more consequential — can it do the work?
