Friday, August 14, 2026

Google shipped Gemini 3.7 Flash — a coding/agent workhorse at half the previous Flash price — while OpenAI previewed an "Ultrafast" GPT-5.6 Sol tier on Cerebras hitting 750 tokensA 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 ↗/sec. US labs are cutting prices (OpenAI dropped GPT-5.6 Luna 80%) even as DeepSeek launched V4-Pro and raised some API rates up to 1,100%. And OpenAI paused its Astra model after it couldn't rule out "Critical" cyber capability under its own framework.

TLDR

Google ships Gemini 3.7 Flash — coding/agent model with big DeepSWE/FrontierCode gains, 1M context, at an intro $0.75/$3.75 per M (half the prior Flash) (Aug 13) Trending
OpenAI previews "Ultrafast" GPT-5.6 Sol on Cerebras — same model, up to 14x faster (~750 tok/s), limited API preview (Aug 13) Infra
US price cuts vs. DeepSeek hike — OpenAI cut GPT-5.6 Luna 80% and Anthropic pitched Opus 5 at "half of Fable 5," while DeepSeek's new V4-Pro raises some rates up to 1,100% (Aug 13–14) Pricing
OpenAI pauses Astra after tests couldn't rule out first-ever "Critical" cyber capability under its Preparedness Framework (Aug 7) Security
Databricks raises $5B at $190B as its revenue run-rate tops $7B (Aug 13) Trending

Action Items

Things your engineering team should look at or do this week.

Bake-off Gemini 3.7 Flash for coding/agent workloads. It posts large coding-benchmark jumps (DeepSWE v1.1 49%→65.3%, FrontierCode 1.1 34.4%→43.6%), 1M context, and an intro $0.75/$3.75 per M through year-end — worth a head-to-head against your current default. It's live in the Gemini API, AI Studio, and Antigravity. (Reuters)
Re-forecast token spend — direction is now mixed. US frontier models are getting cheaper (OpenAI cut GPT-5.6 Luna 80%; Opus 5 at ~half Fable 5) while DeepSeek's V4 price changes hit Aug 17. If you route across providers, re-check assumptions before then. (Financial Times via SL Guardian | Caixin)
Treat orchestration as a cost lever, not just the model. Writer's new research shows a rebuilt agent "harness" cut token spend ~41% and ran ~44% faster across third-party models too — audit your own retry/context/sub-agent logic before assuming you need a bigger model. (TechCrunch)

Models & Tools

Google — Gemini 3.7 Flash (Aug 13) Trending — Released just three weeks after 3.6 Flash, Google's "most intelligent workhorse" for coding, knowledge work, and agents. Gains on nine benchmarks (DeepSWE v1.1 49%→65.3%, FrontierCode 1.1 Main 34.4%→43.6%), 1M-token context, better first-pass code. Intro pricing $0.75/$3.75 per M through year-end; live in the Gemini API, AI Studio, Antigravity, and Spark. (Reuters | SiliconANGLE | 9to5Google)
OpenAI — "Ultrafast" GPT-5.6 Sol on Cerebras (Aug 13) Infra — A new API tier running the same GPT-5.6 Sol up to 14x faster (~750 output tok/s) on Cerebras wafer-scale chips, aimed at latency-sensitive real-time apps and agents. Limited preview for select API customers, expanding with capacity; OpenAI reports ~5.6x end-to-end speedup on GDP-Val with no quality loss. (OpenAI | Cerebras)
Writer — Palmyra X6 + rebuilt agent harness (Aug 13) Agents — A new flagship (post-trained on Z.ai's GLM-5.2) at $2/$8 per M, paired with an orchestration layer Writer says cuts agent cost ~52% and runs ~48% faster with X6. The harness alone cut token spend ~41% and ran ~44% faster across third-party OpenAI/Anthropic models — a bet that orchestration, not just model choice, drives agent economics. (TechCrunch | VentureBeat)
Z.ai — GLM-5.3 (Aug 14) Open Source Security — Open-weightsA model whose trained weights are published for anyone to download and run.NTIA ↗ update built on the same 743B base as GLM-5.2, with gains from post-training alone: tops open coding leaderboards and edges Claude Mythos 5 / GPT-5.6 Sol on the CyberGym vuln-detection benchmark (84.5). Notably, cyber capability "outgrew" its intended training, surfacing 1,097 critical bugs in Linux/WebKit/FreeBSD; weights to be released in ~2 weeks after safety hardening, with higher-risk capabilities gated. (Unite.AI | MarkTechPost)

APIs & Pricing

US labs cut prices as Chinese rivals gain (Aug 14) Pricing Trending — Per the FT, OpenAI cut its "fastest and most affordable" GPT-5.6 Luna by 80%, and Anthropic is positioning Claude Opus 5 at ~half the price of Fable 5, as cost-conscious buyers (DoorDash, and testers like Siemens/Airbnb) route work to cheaper Chinese models. The competitive frame is shifting from benchmarks to cost-per-task. (Financial Times via SL Guardian)
DeepSeek launches V4-Pro and raises API prices (update) (Aug 13, effective Aug 17) Pricing — Following last week's warning of a "significant" hike, DeepSeek formally shipped its V4-Pro flagship and set peak/off-peak pricing with increases of 50%–1,100% depending on model/token type/time. V4-Pro output hits $3.96/M at peak (from a $0.87 flat rate); cheaper V4-Flash remains. Even at peak it undercuts most Western frontier tiers — but re-check if you depend on DeepSeek. (Caixin | Reuters via SRN)

Industry & Business

Databricks raises $5B at $190B valuation (Aug 13) Trending — Its second round this year (up from $134B in February), led by Coatue with Blackstone, MGX, and T. Rowe Price, as revenue run-rate tops $7B (80% YoY growth). Funds target Lakebase (serverless Postgres for agents), the Genie "AI coworker," and Unity AI Gateway for multi-model governance/cost control. (Bloomberg | CNBC)
IBM + OpenAI form enterprise alliance (Aug 13) Customer AI — IBM will embed OpenAI's frontier models plus Codex and ChatGPT Work into IBM Consulting Advantage, spin up a dedicated OpenAI practice, and certify tens of thousands of consultants — targeting financial services, government, telco, and retail. Signals AI distribution increasingly running through systems integrators. (IBM Newsroom | TechCrunch)
Apple trains a China-specific model with Alibaba (Aug 14) Customer AI — Apple built a custom LLMLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ for its China Apple Intelligence rollout with Alibaba's help, having registered its on-device generative service with Chinese regulators — a concrete example of geopolitics fragmenting the model/infra stack by market. (The Verge)
Lovable raises $400M Series C at $13.3B (Aug 12) Trending — The Swedish "vibe-coding" platform's round (led by Menlo Ventures, co-led by Scaleup Europe; CapitalG, EQT among others) values the app-builder at $13.3B, ~$730M raised to date — another marker of capital pouring into AI software creation. (TechCrunch | Crunchbase News)
Pony.ai + Uber to deploy 2,000+ robotaxis across Europe (Aug 14) Robotics/IoT — Expanding beyond their Zagreb service to five European cities (plus a Middle East ambition), with Pony.ai supplying L4 tech and Uber handling booking/payments — a move from pilot-scale to fleet-scale autonomy. Cities and timeline unspecified. (Uber Investor | TechCrunch)

Policy & Safety

OpenAI pauses Astra over "Critical" cyber capability (Aug 7) Security — OpenAI said internal tests couldn't rule out that its Astra model reaches the "Critical" cyber tier under its Preparedness Framework — i.e., independently finding unknown vulns or planning sophisticated attacks with little human input. It paused non-compliant internal Astra work pending stronger controls (isolated testing, restricted access, monitoring) — the first model to hit that threshold. (TechCrunch | OpenAI | Axios)
EU AI Act transparency rules kick in; Anthropic watermarks Claude output (Aug 11) Policy — With Article 50 transparency obligations now enforceable (as of Aug 2), Anthropic began adding machine-readable/invisible watermarks and signed provenance to text and files from new Claude models — applied globally, not just in the EU. Worth checking if watermarking affects any downstream copy/paste, detection, or content pipelines. (TechCrunch | Euronews)

Paper Spotlight

AI designs 16 working viruses from scratch (Stanford / Arc Institute) (Aug 6, Science) Policy — Using the Evo 1/Evo 2 genome language models, researchers generated complete, viable bacteriophage genomes — the first fully AI-designed viruses — that killed bacteria resistant to their natural counterparts. Human/animal/plant-infecting sequences were excluded from training to limit risk, but the result crystallizes a real biosecurity governance gap alongside a promising path for phage therapy. (betanews | BioPharmaTrend) (This is a sensitive dual-use topic; included for its governance relevance.)

A price-war-and-throughput day.