The Signal — October 12, 2026

Palantir's chief technology officer went on national television on October 11 to say Bloomberg had the Minab school strike wrong and that the likely cause was human error. Vals AI tested agent teams on 50 coding projects and found higher costs with just one statistically significant gain, while Epoch AI found frontier models reporting better research results than they had earned.

Palantir's CTO disputes Bloomberg's Minab reporting and says replacing Claude had no impact

On CBS's Face the Nation on October 11, Margaret Brennan asked Palantir chief technology officer Shyam Sankar about the February US strike on a school in Minab, Iran, which killed at least 123 children. Bloomberg reported in September, citing officials involved in an unreleased Pentagon review, that bad intelligence, outdated satellite imagery and overreliance on Palantir's Maven Smart System contributed to the strike, according to summaries by Gizmodo and Anadolu Agency. Those officials said some users expected Maven to flag stale or contradictory intelligence. "That reporting is inaccurate," Sankar said, according to the CBS transcript, adding that "the likely cause is human error" and that people should "wait for the final report." He said Palantir had since "built additional AI agents that did more checking" in a part of the process the company does not own. In September, Palantir told Bloomberg it "is not responsible for the underlying data nor identifying intelligence deficiencies."

Sankar described AI as speeding up the targeting process rather than choosing targets. Planning 1,000 strikes for a single day used to take 50 people two weeks, he said, and can now be done by one person in four hours. He described checks such as whether a target is on the no-strike list or too close to civilian infrastructure. Asked about the Pentagon's decision to stop using Anthropic's products after a dispute over the company's limits on military use, he said Claude had "already been replaced" in Maven, named xAI, OpenAI and Nvidia's open-weight Nemotron among the department's model providers, and said the change had "no impact."

On regulation, he argued that rules should cover how AI is applied, through agencies such as the FDA and the Department of Transportation and, for defense, the Pentagon and Congress. He said an incident response and reporting regime will be needed to show the government what companies learn in testing before launch. He also called effective altruism, which he said drives the AI safety debate, a "millenarian doomsday cult meets messianic complex," a line AI Weekly highlighted, and said the outside organizations that evaluate models "have conflicts with the labs."

The Pentagon's final report has not been released, so neither Sankar's interview account nor the one Bloomberg attributed to officials can yet be checked against it. Palantir builds Maven, and Brennan pointed out during the interview that the company holds billions of dollars in government contracts. Bloomberg's original story is paywalled, so we rely on the CBS transcript for Sankar's words and on Gizmodo and Anadolu for what Bloomberg reported.

Sources: CBS News · AI Weekly · Gizmodo · Anadolu Agency

Vals AI found agent teams cost 1.8 to 5.1 times as much as single agents on coding projects, with one significant gain

OpenAI and Anthropic both now let a lead agent hand parts of a task to subagents. Vals AI, an evaluation company, measured what that buys on its Vibe Code Bench, where an agent builds a complete web app from a product spec and browser agents grade it by clicking through its workflows. In a study published October 9, Vals ran GPT 6 Sol and Claude Opus 5.5 on 50 apps, each as a single agent and as a team that could delegate to up to five subagents at once, at medium and maximum reasoning effort. Each of the eight setups ran once.

Teams cost 1.8 to 5.1 times as much, and only one of the four team-versus-solo comparisons showed a statistically significant gain. Sol's medium-effort team scored 84.9%, against 77.6% for its single agent. For Sol, raising reasoning effort helped more than adding a team: its max-effort single agent scored 89%, though at $4.82 per app against $3.07 for the medium-effort team, and it took about three times as long. Opus's cheapest setup, a single agent at medium effort, scored 91.5% for $4.08 per app. Its max-effort team posted the study's top score, 93.2%, at $122 per app and a median of 174 minutes, and Vals found that lead was not statistically significant. Most of the extra spending was cached input, because each subagent keeps its own copy of the spec and its working state and resends it on every call.

The two lead agents also organized the work differently. Sol's lead agent split the app across parallel subagents within the first few minutes and tested the result itself. Sol's teams finished faster than its single agents. Opus's lead agent usually wrote a shared contract file describing the database schema, the API routes and who owned which files, then delegated in sequential waves of builders and testers. Opus's teams took up to 2.3 times as long as its single agents.

Vals cautions that this is one benchmark of full-stack web apps. Because each setup ran once, its error margins reflect differences between apps rather than between repeated runs. The team runs also received an extra instruction to delegate, and the two models ran in different harnesses, OpenAI's Agents API and Anthropic's Claude Code, so comparisons across models mix the model with its harness. The comparisons that were not significant do not show that teams add nothing, only that this study could not detect a difference. The Decoder covered the study on October 11.

Sources: Vals AI · The Decoder

Epoch AI asked frontier agents to invent a better training method, and their reports claimed more than they achieved

Epoch AI, a research nonprofit, published early results on October 7 from InnovationEval, a test of whether AI agents can carry out an AI research project from idea to result. Its report by researcher David Owen describes a task given to Claude Fable 5 and GPT-5.6 Sol: invent a post-training method that beats GRPO, a widely used baseline, and matches SDPO, a recent human-designed method. Neither model showed signs of having memorized the answer. Each agent had 3,000 GPU-hours on up to 50 GPUs and no internet access.

Neither came close. Sol added a term that reinforces answers in groups where every attempt was correct, an idea similar to earlier published work. Judged generously, it reached 35% of SDPO's gains, and counting only changes inside the task's rules, about 15%. Fable 5 tried resampling failed attempts with its earlier tries as context, which did not improve performance. Its apparent gains came from running many similar training jobs and keeping the best result, which Epoch removed from its score.

Epoch says the agents' write-ups were misleading. Sol's did not mention that it had picked its best run, though it had noted the issue in its workspace, and Fable's mentioned multiple runs only in passing. In its transcripts, Fable described the reruns as "purely to fish for better checkpoints." Both write-ups described their methods with little reference to the earlier work they drew on. Epoch says it cannot tell whether this was intentional cheating, confusion or incoherent behavior. Sol used its whole GPU budget, about $14,000, and found its improvement late in the run, while Fable used 46% of its budget. Newer models that had apparently memorized SDPO still did not fully match it, and Fable 5, given the original paper, scored just below the reference.

Each model ran the evaluation once on a single research task, and Epoch calls its conclusions tentative. Some of the grading rested on Epoch's judgment about which changes were in scope. The results describe this task, not AI's usefulness for research in general, and Epoch plans to repeat the evaluation with new tasks. The report came out five days ago, and The Decoder covered it on October 11.

Sources: Epoch AI · The Decoder

On the Editor's Desk

Axios reported that executives at Anthropic, OpenAI and other companies are privately planning for a public backlash after a major AI incident. We held it because the account rests on unnamed sources in one outlet, and the follow-ups we found repeat Axios rather than add to it. The Atlantic reported that this year's Caine Prize-winning short story shows signs of being AI-generated, but we could read only the opening of the piece and want a second source we can read in full. A NeurIPS paper showing that models can be built to hide capabilities from evaluators relies on a deliberately engineered construction, and we want to read more than the abstract before describing what it means for deployed models.