A cyber-and-compute day.
OpenAI shipped GPT-5.6-Cyber and split its Daybreak program into defender/offense tiers — its first "offense-grade" model gated behind access controls, landing the same week as its Black Hat breach post-mortem. Anthropic deepened its Google/Broadcom compute pact for multiple gigawatts while disclosing a $30B run-rate. And two hard deadlines are bearing down: Google's Imagen 4 API dies Aug 17 and DeepSeek is warning of a "significant" price hike.
TLDR
- OpenAI launches GPT-5.6-Cyber + two-tier Daybreak (Blue/Red) — a purpose-trained cyber model for authorized defenders, restricted access (Aug 10) Security
- Anthropic expands Google/Broadcom compute deal for multiple GW, mostly US-sited; run-rate revenue tops $30B, 1,000+ customers over $1M/yr (Aug 11) Infra
- Google Imagen 4 API hard shutdown Aug 17 —
generate_images()breaks with an error, not a warning (Aug 17) Pricing - DeepSeek warns of "significant" API price increase — second pricing move in a month, still likely to undercut Western rivals (Aug 6) Pricing
- Google's agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ shopping now calls stores and auto-checks out — voice agents phone retailers to check stock; budget-triggered buying via Google Pay (early Aug) Customer AI
Action Items
Things your engineering team should look at or do this week.
- Migrate off Imagen 4 before Aug 17. The
imagen-4.0-generate-001/fast/ultraGemini API endpoints shut down hard — and it's not a drop-in swap.generate_images()is gone; image gen now runs throughgenerate_content()on Gemini 3.1 Flash Image. Audit any image pipelines now. (Google AI docs) - Budget for two near-term price moves. Anthropic's Claude Sonnet 5 intro pricing ($2/$10 per M) reverts to $3/$15 on Sep 1 (a 50% jump), and DeepSeek says a "significant" hike is coming with no date. Re-forecast tokenA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗ spend and check whether routing assumptions still hold. (Bloomberg)
- If you run a security team, evaluate Daybreak access for GPT-5.6-Cyber. It answered 95% of advanced cyber requests vs 1.5% for GPT-5.6 Sol, and is gated behind vetting — useful for authorized vuln research/code review, but understand the offense/defense tiering before requesting. (OpenAI)
Models & Tools
- OpenAI — GPT-5.6-Cyber + Daybreak Blue/Red (Aug 10) Security — OpenAI split its cyber program into two access tiers: Daybreak Blue (frontier models like GPT-5.6 Sol with defensive safeguards, for vuln discovery, code review, malware analysis, IR) and Daybreak Red (purpose-trained models including the new GPT-5.6-Cyber, for authorized exploit validation and security testing). GPT-5.6-Cyber, built on Sol, has already surfaced unknown bugs in Chrome's V8. (OpenAI | TechCrunch | Axios)
- Google — agentic shopping that calls stores + auto-checkout (early Aug) Customer AI Agents — A "Let Google Call" option now dispatches a voice agent to phone retailers, check inventory, and text/email a summary back; a separate agentic checkout tracks a product's price and buys it via Google Pay once it hits your budget (with approval). Launch is US-only with select merchants (Wayfair, Chewy, Quince, some Shopify stores). (TechCrunch via Yahoo | GSMArena)
APIs & Pricing
- Google Cloud: Imagen 4 shuts down Aug 17 (Gemini API) (Aug 17) Pricing — All three Imagen 4 endpoints (standard/fast/ultra) are deprecated with a hard shutdown; the Vertex AI side already hit its deprecation date Jun 30. Migrate to Gemini 3.1 Flash Image, noting the API surface change from
generate_images()togenerate_content(). (Google AI docs) - DeepSeek warns of "significant" API price increase (Aug 6) Pricing — DeepSeek told developers to plan for a substantial hike across API services — its second pricing change in under a month after introducing peak/off-peak rates in mid-July. Founder Jun Song says even a 2x–10x increase would still undercut most Western rivals; scope (V4-Flash vs V4-Pro, cache-hit, reasoning modelsA model that works through hidden reasoning steps before its final answer. Those steps are billed as output.OpenAI ↗) is unspecified. (Bloomberg | Dataconomy)
- Anthropic: Claude Sonnet 5 intro pricing ends Sep 1 (Sep 1) Pricing — The introductory $2/$10 per M input/output reverts to $3/$15, a 50% increase. If Sonnet 5 is in your production path, re-run cost models before month-end. (AI API pricing tracker)
Industry & Business
- Anthropic deepens Google/Broadcom compute pact; run-rate tops $30B (Aug 11) Infra — Anthropic expanded its Google Cloud partnership for multiple gigawatts of next-gen compute (building on October's TPUTensor Processing UnitGoogle's custom chip built to speed up machine learning, especially large matrix operations.Google Cloud ↗ capacity increase), with the vast majority US-sited — part of its $50B commitment to US compute infrastructure. Separately, run-rate revenue has surpassed $30B (up from ~$9B end of 2025), with 1,000+ customers spending over $1M/yr. (Anthropic)
- Anthropic hires Tino Cuéllar as first Chief Global Affairs Officer (Aug 4) Policy — The former California Supreme Court Justice and Carnegie Endowment president joins to lead global policy and government relations, reporting to Daniela Amodei; he steps down from Anthropic's Long-Term Benefit Trust to take the role. Signals a deepening policy build-out as regulatory pressure grows. (Anthropic | Harvard Crimson)
- OLIX raises $312M Series B for photonic AI inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ (Aug 3) Infra — UK chip startup OLIX raised $312M at a $3.3B valuation (led by Fundomo; Arm, Hudson River Trading, Reed Hastings, UK Sovereign AI Fund) for Optical Tensor Processing Units that use light instead of HBM — Britain's biggest semiconductor bet. Customer access targeted for H2 2027. (DataCenterDynamics | Converge Digest)
Policy & Safety
- EU DMA orders Google to open Android to third-party AI assistants (Aug 11) Policy — Under the Digital Markets Act, the EU is requiring Google to open Android to rival AI assistants (Claude, ChatGPT) by August 2027 — a structural opening for AI-assistant distribution on mobile. (AIToolsRecap summary) (roundup source — confirm against the EC's primary DMA decision before acting)
Paper Spotlight
- A 150M-param recurrent-latent-reasoning model sets a new ARC-AGI-1 cost-accuracy frontier (Aug 10) — A tiny reasoning model using recurrent latent reasoning plus in-context learning reaches a new cost-accuracy frontier on ARC-AGI-1, a reminder that architecture and test-time compute — not just parameter count — keep moving the reasoning bar. (HF trending papers)