The Intake — Saturday, August 15, 2026

If you've been designing multi-agent systems and trying to work out what a functional role-differentiated swarm architecture looks like in practice, Anthropic published a concrete example. An unnamed Anthropic research model moved a mathematical record that had stood since 1989. The record is the lower bound on the proportion of Riemann zeta function zeros satisfying the Riemann hypothesis. The previous bound was 41.6%; the new result is 67.2%.

The run used 60 specialized subagents across five role classes. Two developed core ideas and thirteen developed supporting ideas. Thirty attempted new approaches. Thirteen validated arguments and two handled write-up. An Anthropic employee with limited mathematical training initiated the run. The model then coordinated autonomously for roughly a day and a half. The run spanned two sessions and approximately 31 million output tokens.

Four mathematicians independently verified the result. Two were Anthropic-internal. Brian Conrey and Dan Goldston reviewed the work externally. Anthropic formalized the output as a machine-checkable Lean proof. Anthropic states it does not expect the techniques to lead to a proof of the Riemann hypothesis.

If you're designing long-horizon agent runs, the five-class role structure — core development, supporting development, approach attempts, validation, write-up — is the architecture this result puts on the record.

If you're evaluating models for long-horizon agent workflows, SpaceXAI released Grok 4.6 on August 12. The model is a post-training upgrade over Grok 4.5. The training includes a supplemental run and regenerated SFT trajectories. The reinforcement learning component targeted agentic environments and long-horizon tasks specifically.

The context window is 500,000 tokens. Pricing is $2 per million input tokens and $6 per million output tokens. The model is available through the API and Cursor. The Grok Build tool is also supported.

On the Artificial Analysis Intelligence Index, Grok 4.6 scores 61. That ties it with GPT-5.6 Sol Max for third place. On DeepSWE v1.1, Grok 4.6 posts 65.9%. GPT-5.6 Sol Max scores 73% on the same benchmark. On Terminal-Bench v3.0, Grok 4.6 posts 26%. GPT-5.6 Sol Max and Fable 5 Max each post approximately 34%.

If you're evaluating models specifically for long-running agent capacity, the DeepSWE and Terminal-Bench scores are the task-specific numbers from this release.

---