Google says frontier models already know the facts they get wrong. That is a budget decision.

Google Research published a post on August 12, 2026 by research scientists Nitay Calderon and Gal Yona summarising a paper called "Empty Shelves or Lost Keys? Recall Is the Bottleneck for Parametric Factuality," which introduces a behavioural framework the authors call knowledge profiling ^[1][2]. Their headline result: on their own benchmark, Gemini-3-Pro and GPT-5 encode 95-98% of facts but still fail to directly recall 26-34% of them ^[3]. If that generalises, the marginal dollar for factual reliability belongs to post-training and inference plumbing, not to another pretraining run.

The framing matters because standard accuracy scores collapse two different failures into one number ^[13]. The paper separates encoding (the fact is represented in the weights), recall (the model produces it with no external cue), and recognition (the model picks it out from alternatives) ^[6], and then classifies each fact into one of five profiles rather than scoring each question ^[7]. The authors are explicit about the consequence: encoding failures argue for more parameters or more data coverage, while recall failures also open the door to post-training and inference-time methods that help a model use what it already holds ^[5].

The measurement scaffolding is substantial. WikiProfile contains 2,150 Wikipedia-derived facts, each paired with ten tasks: two for encoding, four for knowledge evaluation, four multiple-choice variants for recognition ^[8], which is 21,500 questions ^[1]. Thirteen models were run with and without thinking, eight samples per model, fact and task, producing roughly 4.5 million responses graded by prompted autoraters ^[9]. That arithmetic checks out: 21,500 questions at eight samples across thirteen models in two modes is about 4.47 million ^[5].

Now the ratio that should drive planning. If 95-98% of facts are encoded, the encoding gap is 2-5%, while the direct-recall gap is 26-34%, so recall failure is roughly five to seventeen times the size of the problem that scale is supposed to solve ^[2]. Thinking helps and does not close it: failures drop to 11-12% ^[4], which removes about 58-65% of the direct-recall failures ^[3] but leaves a residual still two to six times the encoding gap ^[4]. Thinking here means eliciting intermediate computation before the answer, including chain-of-thought prompting and thinking-optimised models ^[12], so that improvement is bought with tokens and latency on every request, not once at training time.

Two caveats before anyone reallocates a roadmap. This is Google evaluating models including its own: of the four frontier systems named in the post, three are Gemini variants ^[11][6], and the benchmark itself was generated by a prompted Gemini-2.5-Pro with thinking, filtered through a search engine and then manually validated ^[10]. And the facts are Wikipedia-derived ^[8], which bounds the claim to encyclopedic knowledge that pretraining corpora cover well. Nothing here says a model has encoded your contract terms or your inventory table, so retrieval over private data remains a separate expense.

What to watch: whether recall-failure rates move in the next round of post-training releases, since that is now the visible dial; whether vendors start pricing or exposing thinking budgets against measured factual recall rather than reasoning benchmarks; whether anyone applies knowledge profiling to a proprietary corpus, where the encoding-versus-recall split is unknown; and the paper's recognition numbers, because architectures that generate candidates and verify them depend on recognition being much cheaper than recall ^[6].