Need to know
- Google Gemini 3.7 Flash ships with major coding benchmark gains while the flagship Pro model remains delayed.
- OpenAI and Cerebras launch Ultrafast mode for GPT-5.6 Sol at 750 tokens per second, 14x faster than standard.
- Z.ai unveils a new coding model aimed directly at Anthropic and OpenAI's enterprise positions.
- Cerebras reports Q2 revenue of $210M, up 103% year-over-year, driven by the OpenAI deployment ramp.
- Microsoft SCCM chain of four CVEs lets a domain user seize SYSTEM control for $58 until patches arrive in October.
New Releases
Google launched Gemini 3.7 Flash, a lower-cost model for coding and multi-step agent workflows, three weeks after 3.6 Flash shipped.
- Benchmark jumps are substantial: FrontierCode 1.1 rises from 34.4% to 43.6% and DeepSWE v1.1 from 48.6% to 65.3%, priced at $0.75 input and $3.75 output per million tokens.
- Pro model absence dominates coverage: Google confirmed Gemini 3.5 Pro is still in partner testing with no public release date, leaving the flagship slot open as rivals push capable models.
OpenAI began previewing Ultrafast mode for GPT-5.6 Sol, delivering up to 750 output tokens per second via Cerebras wafer-scale chips without quality degradation.
- Speed without tradeoff: Cerebras WSE-3 stores model parameters entirely in on-chip SRAM, eliminating the memory bandwidth bottlenecks that force GPU clusters to choose between latency and intelligence.
- API-first rollout: Ultrafast launches in the OpenAI API before consumer surfaces, targeting developers building latency-sensitive agentic pipelines.
Z.ai released a new coding-focused AI model explicitly targeting Anthropic and OpenAI's positions, adding to a series of Chinese open-weight competitive releases.
- Chinese lab pressure on pricing: The release continues a pattern where Chinese labs use open-weight models to compress margins in coding, the category where enterprise AI spending is most measurable.
- Timing is deliberate: The launch coincides with Google's Flash-only posture and OpenAI's speed-tier focus, targeting enterprises evaluating alternatives during a busy model week.
Palo Alto Networks expanded its AI-based exposure analysis capability by integrating an advanced OpenAI model for deeper threat contextualization.
- Practical security use case: The integration applies frontier reasoning to attack surface data, moving exposure analysis from rule-based scoring toward model-driven prioritization.
- Platformization play: The addition reinforces Palo Alto's strategy of embedding AI into Cortex rather than offering it as a standalone product, increasing switching costs for customers already on the platform.
Funding
Cerebras reported $209.9M in Q2 core revenue, up 103% year-over-year, raised full-year guidance to $880-890M, and projects core revenue to more than triple in 2027, with $25.4B in remaining performance obligations tied largely to the OpenAI Ultrafast deployment.
Above Security received a strategic investment from the CrowdStrike Falcon Fund, signaling CrowdStrike's intent to extend its ecosystem into adjacent detection and response capabilities rather than build every layer internally.
Seoul-based Algorix raised $11.5M led by K2 Investment Partners to expand its Akashic platform, an AI-native multi-model database built for the end-to-end data pipeline that agentic AI execution requires, signaling investor conviction that purpose-built agent data infrastructure is a distinct category from existing vector or relational databases.
Case Studies
Tanium published survey data from 333 senior IT and security professionals across Southeast Asia showing that endpoint management failures cost some organizations over $1M per incident.
- Scope of disruption is broad: 62% of surveyed organizations reported endpoint-related operational disruption, with 63% of affected firms experiencing downtime, 41% data exposure, and 31% direct revenue loss.
- Patching delays are the proximate cause: Weak device visibility and slow patch cycles were the leading factors cited, which maps directly to the 48-hour CVE exploitation window documented in CrowdStrike data covered earlier this week.
Trending on X
- Gemini Pro delay frustrates enterprise buyers Practitioners and analysts are calling out Google for shipping another Flash iteration while Gemini 3.5 Pro remains in limbo, with several threads arguing the delay is handing Anthropic and OpenAI another quarter to close enterprise deals.
- Ultrafast Sol changes agent economics Sam Altman and Greg Brockman both posted about GPT-5.6 Sol Ultrafast within hours of launch, triggering debate about whether 750 tokens per second finally makes frontier-model agents viable for real-time user-facing products without falling back to smaller models.
- CIOs unhappy with OpenAI and Anthropic enterprise terms A widely shared thread citing private CIO conversations argues that despite headline adoption numbers, enterprise buyers are quietly frustrated with both labs' pricing opacity and support posture, with Chinese alternatives gaining serious evaluation cycles.
- Fchollet: LLM-guided symbolic world models François Chollet drew significant engagement arguing that the most promising AI direction is LLM-guided on-the-fly synthesis of symbolic world models, framing it as the architecture behind leading ARC 3 results and the long-run trajectory of the field.
- Anthropic vs OpenAI IPO race pricing Prediction market data showing a 94% probability that Anthropic goes public before OpenAI is circulating on X, with traders debating whether Anthropic's cleaner governance structure and watermarking moves are deliberate IPO positioning.