Two developments this week carry real architectural weight: OpenAI has named its next major model and demonstrated its capabilities at a cost that compresses assumptions about expert-knowledge work, and Microsoft has moved its first purpose-built security model into public preview. A third item deserves attention before the calendar turns.

OpenAI names its next major model — and proves it with unsolved mathematics

On 1 August, OpenAI published ten advances in mathematics and theoretical computer science produced by an internal model it named Astra. The announcement was deliberately understated — embedded in a research post rather than a product launch — but the label is explicit: Astra is described as OpenAI's "next major model" after the GPT-5.6 family.

The headline result is the first-ever explicit construction of a non-sofic group, resolving a question in group theory that has stood open since Mikhail Gromov introduced the concept of soficity in 1999. No mathematician had established whether non-sofic groups exist in the 27 years since. Astra also produced a disproof of Connes's Rigidity Conjecture and new bounds across high-dimensional sphere packing, circuit complexity, lattice cryptography, and extremal combinatorics. Each result is accompanied by a 249-page manuscript and machine-checkable Lean 4 certificates published on GitHub, allowing independent verification.

The compute cost for all ten results was approximately $2,000 at current API rates. That number does not describe the model's deployment cost — Astra is internal and not yet released, and it is expected to go through the US government's frontier model review before any public availability. What the number does describe is the efficiency curve: the class of problems that used to require years of expert labour, and that now fit inside a $2,000 compute budget, is widening.

Astra is described as a multi-agent architecture designed for sustained work over hours or days, which is consistent with OpenAI's "abundance" framing from last week. The next generation is not a faster text interface but a research collaborator. For operators, the relevant question is not the launch date. It is what happens to the competitive moat in any domain that runs on expert knowledge — legal analysis, financial modelling, drug discovery, complex engineering — when the cost-per-insight curve moves faster than planning cycles assume.

Microsoft's Project Perception enters public preview today

Microsoft's Project Perception — an agentic cybersecurity platform — enters public preview on 3 August. At its centre is MAI-Cyber-1-Flash, the first model Microsoft has built specifically for security work. The architecture is a sparse mixture-of-experts transformer: 137 billion total parameters, five billion active per token, and a 256,000-token context window. It scores 95.95 per cent on the CyberGym benchmark, and Microsoft reports the combined system runs at roughly half the operating cost of its current deployed configuration.

Project Perception coordinates specialised agents — red for adversarial simulation, blue for detection, green for remediation — in a continuously learning system designed to hold humans in control of consequential decisions while automating the high-volume triage underneath. Access is through Azure AI Foundry with customer vetting rather than open public availability.

The operational shift here is that security AI has crossed from assistive to agentic. A system that detects a threat, generates a remediation plan, and executes it without a ticket queue changes the staffing and process assumptions of any security operations centre. Organisations evaluating this category should now be asking not just whether to adopt agentic security tools, but which decisions must remain manual and how to instrument those hand-off points.

Four weeks to review your Claude Sonnet 5 workloads

Anthropic's introductory pricing for Claude Sonnet 5 — $2 per million input tokens and $10 per million output tokens — expires on 31 August 2026. Standard pricing from 1 September moves to $3 and $15 respectively, which is the same rate card as Sonnet 4.6.

The less-visible factor is the tokenizer. Sonnet 5 uses a new tokenizer that produces approximately 30 to 35 per cent more tokens for the same input text. At the standard rate, an identical workload can cost 30 to 35 per cent more on Sonnet 5 than it did on Sonnet 4.6 — even though the per-token price is unchanged. Teams that migrated during the promotional window and have been operating on the introductory rate may see a combined step-up that exceeds 100 per cent compared to their pre-migration baseline.

  • Run representative production requests through the Sonnet 5 tokenizer now, not the Sonnet 4.6 token counts you used for earlier estimates.
  • Reprice your unit economics and, where relevant, your downstream pricing to customers before 31 August.
  • Note that Anthropic separately extended the 50 per cent weekly usage boost for Claude Code subscribers through 19 August; that window closes before the pricing transition, so the two timelines do not overlap.

The pattern across all three stories today is compression: of research timelines, of security response loops, and of the window operators have to adapt to infrastructure cost changes. In each case the practical response is the same — narrow the planning horizon and measure what you cannot afford to treat as stable.