The best AI platform for factual accuracy

Claude Opus 5.5 at max effort leads the AA-Omniscience Index at 46.4, the Artificial Analysis score that rewards correct answers to closed-book factual questions and penalizes wrong ones. We crown Anthropic for factual accuracy on that run, which also answered a larger share of all the questions correctly than GPT-6 Astra did at high or max effort.

OpenAI suits a workflow where a wrong answer costs more than an unanswered question. GPT-6 Astra at high effort had the lowest hallucination rate of the runs compared here, and it attempted fewer questions than the Opus run.

What a buyer of factual accuracy should measure

AA-Omniscience asks 6,000 questions across 42 topics in six domains, and the Artificial Analysis methodology specifies open answers with no tool use. Accuracy on it is the proportion of questions answered correctly out of all questions, whether or not the model chose to answer. That makes accuracy the measure of coverage when the model has nothing to look up.

The benchmark's hallucination rate is incorrect answers divided by all responses that were not correct, a group that includes partial answers and questions the model did not attempt. A lower rate means a smaller share of those non-correct responses were incorrect answers and a larger share were partial answers or questions left unattempted. It is a rate within the non-correct responses, so it cannot be read as the share of all questions a model got wrong.

A wrong answer lowers the AA-Omniscience Index and an abstention leaves it unchanged, under the benchmark's scoring. Correct answers raise it, so the index is the one figure here that weighs coverage and wrong answers together.

Anthropic's effort documentation sets Claude Opus 5.5 to medium effort by default, and the Opus run that leads the index used max. Artificial Analysis scored GPT-6 Astra at more than one effort and got different results at each, so a published score describes one setting of a model. Stanford's HELM framework holds that the strategy for adapting a model to a scenario should be controlled for when models are compared.

Vectara's hallucination leaderboard gives each model a passage and asks for a summary using only the information in that passage, across more than 7,700 articles. Its rate measures factual consistency with a supplied document, which is a different task from closed-book recall. Check it if your work is summaries or answers drawn from retrieved documents.

Artificial Analysis publishes a cost per task for each run on its Intelligence Index, a broader benchmark suite that includes AA-Omniscience. The figure prices a run across that suite and does not give a cost per factual question. List price per million tokens is a second basis, and a lower token rate does not settle the task cost, because a run that spends more tokens is billed for each of them.

NIST's AI RMF Playbook asks how performance metrics such as accuracy will be monitored after the AI is deployed, and it calls for evaluation in conditions similar to the deployment context. For this purchase that means scoring a set of your own domain questions at the exact model and effort you plan to run, and counting wrong answers and abstentions separately.

Anthropic and OpenAI on the accuracy benchmarks

Anthropic: Claude Opus 5.5 at max effort leads the recall index

Artificial Analysis lists Claude Opus 5.5 (Max, Default Fallback) first on the AA-Omniscience Index, and the same run answered 66.2% of all questions correctly. Its hallucination rate was 58.6% and it attempted 89.0% of the questions. The Default Fallback label marks a run made with Anthropic's default fallback enabled, in the words of the Artificial Analysis article on the model.

Claude Fable 5.1 at max effort answered more of the questions correctly than the Opus run. Artificial Analysis's article on Fable 5.1 says it "records the highest accuracy we have measured at 67.2%." The same run attempted 93.4% of the questions and had a hallucination rate of 72.6%, so a larger share of its non-correct responses were wrong answers. Artificial Analysis places it third on the index, behind the Opus run and GPT-6 Astra at high effort.

Anthropic's effort documentation says Opus 5.5 defaults to medium and that omitting the effort parameter gives the default, so a team that wants the scored run sets effort to max. The same documentation tells teams to run an effort sweep on their own evaluations.

The Artificial Analysis leaderboard puts Opus 5.5 at $5.98 per Intelligence Index task at max and $1.34 at medium, a cost measured across that broader suite. Anthropic's pricing page lists Opus 5.5 at $4 per million input tokens and $20 per million output tokens, and Fable 5.1 at $10 and $50.

Vectara's leaderboard lists Claude Opus 4.7 at a 12.0% hallucination rate on summaries. That row measures Opus 4.7, an earlier release, and does not score Opus 5.5. Anthropic's own guide to reducing hallucinations says its techniques "don't eliminate them entirely."

OpenAI: GPT-6 Astra at high effort and the lower hallucination rate

GPT-6 Astra at high effort posted a 44.8% hallucination rate on AA-Omniscience and attempted 80.3% of the questions. It answered 61.1% of all questions correctly and scored 43.7 on the index, second on the board behind the Opus run.

At max effort Astra's accuracy rose to 62.6% and its hallucination rate rose to 51.3%, with an attempt rate of 83.4% and an index score of 43.4. More effort added coverage and raised the share of non-correct responses that were wrong answers. A team whose concern is wrong answers has reason to test the high setting first.

OpenAI's model page for GPT-6 Astra lists low, medium, high, xhigh and max as its reasoning efforts, and the reasoning guide says supported values are model-dependent. Name the effort in your request, because the two Astra scores above came from two settings of one model.

The OpenAI pricing page lists Astra at $10 per million input tokens and $50 per million output tokens for short-context requests. The model page prices a prompt with more than 272K input tokens at 2x the input rate and 1.5x the output rate for the full request, so the higher rates apply to the whole of a request that exceeds that input threshold. On the Artificial Analysis leaderboard, Astra costs $1.73 per Intelligence Index task at high effort and $3.26 at max.

Vectara's summary leaderboard lists GPT-6 Astra at an 8.7% hallucination rate and GPT-6 Sol at 6.5%, with the older GPT-5.5 at 9.3%. Sol posts the lower rate of the two GPT-6 models on that task, so a summary workload is a reason to test Sol alongside Astra.

Which AI platform is best for factual accuracy?

We crown Anthropic for factual accuracy because Claude Opus 5.5 at max effort leads the index that weighs correct answers against wrong ones, and it answered a larger share of the questions correctly than either Astra run. OpenAI holds two advantages: Astra's hallucination rate was lower at both efforts, and its cost per Intelligence Index task was lower at max. We weigh coverage and the index above those because a buyer of factual accuracy is paying for correct answers first, and the index already penalizes the Opus run for its wrong answers and still places it first. The crown is for closed-book recall with effort set to max.

OpenAI is the better pick when a wrong answer costs your workflow more than an unanswered one. Start with GPT-6 Astra at high effort, which had the lowest hallucination rate in this comparison and attempted the fewest questions. Score Astra on your own questions at that setting before you rely on it to pass on the right ones.

Other AI platforms in this category