X Web Viewer

Last week, @SpaceXAI released Grok 4.6, its new frontier model designed for coding, agentic tasks, and knowledge work. Independent evaluations suggest that SpaceXAI has returned to the AI frontier competition based on both performance and cost. The economics of Grok 4.6 seem particularly compelling for AI agents. Artificial Analysis estimates that Grok 4.6 costs $0.84 per task on its Intelligence Index, placing it on the Pareto frontier for intelligence-versus-cost. More from @MattarARK in this week's newsletter ⬇️

Starlink is taking over aviation. The number of airlines announcing deals with Starlink is surging, and the reason is simple: airlines without Starlink now risk losing passengers to competitors offering a much better onboard connectivity experience. We are already seeing evidence of this on 𝕏, with many passengers saying they are choosing one airline over another specifically because it offers Starlink.

The entire world wants US stocks. Foreign investors outside the US purchased +$181.4 billion of US equities in June. The private sector alone bought a record +$144.7 billion, bringing an annualized rate to +$1.74 trillion, an all-time high. Over the past 3 months, private purchases have totaled +$351.1 billion, equivalent to an annualized pace of +$1.4 trillion, also an all-time high. This surpasses the 2025 annualized record by ~$400.0 billion. By comparison, foreign official institutions, such as central banks and sovereign wealth funds, purchased +$36.7 billion in June. As a result, private investors accounted for ~80% of all foreign equity purchases over this period. The appetite for US stocks is stronger than ever.

Cost per task varies widely, with top performing models from Anthropic ranging from $0.3 to $1 per task - this is up to ~6x higher than the cost of Kimi K3, which sits behind the recent Claude models on the MLCR-AA Score vs. Cost per Task Pareto frontier. While OpenAI’s GPT-5.6 family does not top scores due to low completeness, their strong accuracy comes with relatively lower cost and both GPT-5.6 Terra and Luna sit on the Pareto frontier.

MLCR-AA is our implementation of the Wisedocs benchmark, with the following specifications: ➤ We run only the private Expert and Compound tiers, text-only and with clean context. We do not currently pad case files with filler documents, so this measures reasoning over long meaningful context rather than noise tolerance ➤ Grading uses a three-model judge panel (Gemini 3.1 Pro, Claude Opus 4.8, and GPT-5.5) and aggregates decisions via majority vote for each task and criterion. The top level score requires a response to be correct and complete by majority vote, and to pass the concision test ➤ Every question is run three times, and we report the mean across repeats ➤ Wisedocs has made various revisions to the dataset, and we use the latest version of the holdout set which was revised in August 2026 Scores on our leaderboard are therefore not directly comparable to Wisedocs' original announcement. MLCR-AA is not a component of the Artificial Analysis Intelligence Index.

The latest models stay faithful to source information, but their findings and responses miss key details experts include. MLCR-AA scores each answer on completeness, accuracy, and concision. Accuracy is the easier of the two deciding dimensions, and completeness is where models separate: Claude Fable 5 reaches 73.8% completeness at 90.1% accuracy, while GPT-5.6 Terra (max) records 93.7% accuracy and 33.9% completeness. For medical record review, an answer that is accurate but incomplete can still be unsuitable.

Announcing MLCR-AA, our leaderboard for Wisedocs' MLCR (Medical Long Context Reasoning) benchmark for reasoning over long medical case files. We run the hardest, held-out question tiers, with Claude Fable 5 achieving the top score of 64.4% MLCR-AA tests models on realistic synthetic medical and insurance case files built by the team at @wisedocsai, based on the record review work their platform specializes in. Our leaderboard runs the private hold-out set of 60 questions from the two hardest categories: Expert, which requires specialist medical reasoning across a full case file, and Compound, which packs several independent questions into a single query. Each question is answered against a complete case file of ~70-150 pages, and graded by a three-model judge panel on completeness and accuracy alongside a concision test limiting verbose responses compared to expert answers. Accuracy verifies that the response is well-grounded in the source documents and case context while completeness assesses whether the model produced the same essential details that were included in expert-annotated responses. The concision test verifies that models are not producing excessively verbose content (>5x the length of expert responses). Recent Claude releases from @AnthropicAI lead MLCR-AA, with Claude Fable 5 at 64.4%, and Claude Opus 5 (scoring 53.9% to 59.4% across efforts). and Kimi K3 (max) from @Kimi_Moonshot is the leading open weights model at 38.3%. Key results from MLCR-AA: ➤ Medical record review is partly achievable with AI today, but at high cost and with room to improve: the leading model (Claude Fable 5) achieves a score of 64.4% at a cost of $1 per task, the median model scores <15% ➤ Models stay faithful to source documents but miss key details required for a complete response: nearly 40% of models tested score above 80% for accuracy, the vast majority score below 50% for completeness. Models are largely right about what they do report, and omit a lot of information. GPT-5.6 Terra (max) records the highest accuracy at 93.7% and still places 10th, held back by completeness ➤ Anthropic models lead due to completeness rather than accuracy: the highest-scoring models are from Anthropic, but their accuracy is comparable to or below the strongest OpenAI models. The separation comes from covering the full scope of the expert reference answer We would like to thank Wisedocs for building MLCR and for their collaboration in bringing it to Artificial Analysis!