It was a busy week for frontier AI. Google shipped a new workhorse model and an ambitious open-source push to make AI practical on encrypted data; China’s DeepSeek rolled out peak/off-peak API pricing to match the explosion in demand; Anthropic published a detailed cost-efficiency playbook for Claude Code; and a widely-discussed essay asked whether AI’s edge on mathematics is less about smarter reasoning and more about an almost limitless working memory. Here are the top five AI stories of the day.
Google unveils Gemini 3.7 Flash, its “most intelligent workhorse model”
Just three weeks after Gemini 3.6 Flash, Google released Gemini 3.7 Flash (model ID gemini-3.7-flash), calling it the most intelligent model yet in its high-volume “Flash” line and framing it as a coding-and-agent model first, a chat model second. The update is a direct result of developer feedback and algorithmic optimizations, and Google has positioned it to compete on price as well as quality.
Benchmarks show a modest but steady improvement over 3.6 Flash, with “strong gains” in debugging and issue resolution, better design adherence in UI generation, and improved reasoning in knowledge-dense fields like finance, law, and bioscience. Behaviorally, Google says 3.7 Flash adapts better to roadblocks, asks clarifying questions when intent is ambiguous, and follows instructions with greater fidelity. It maintains the same 1M-token input context and 64k output limit, with multimodal input across text, image, video, audio, and PDF. Updated safeguards ship against misuse in CBRN (chemical, biological, radiological, nuclear) and cyber-offense domains.
Under the headline of the week is the price. Google is offering an introductory price of $0.75 per 1M input tokens and $3.75 per 1M output tokens through the end of 2026 — half the cost of the original 3.6 Flash pricing. The model is live in the Gemini API, Google AI Studio, Android Studio, Google Antigravity, the Gemini Enterprise Agent Platform, and consumer-facing via Spark. Customers including Box, Databricks, Harvey, LangChain, and Stanford’s Department of Biology offered testimonials. Why ship Flash before the long-awaited 3.5 Pro? Improving the model most production clients touch daily keeps Google in the release-cycle conversation during a burst of rapid shipping across every major lab.
Google open-sources HEIR to make private AI practical
Google’s Security Blog (by Staff Software Engineer Jeremy Kun) detailed how the company is making homomorphic encryption practical for AI. Homomorphic encryption allows computation directly on encrypted data, letting a cloud service process user inputs without ever decrypting them. Google is building the open-source HEIR (Homomorphic Encryption Intermediate Representation) compiler to convert pre-trained AI models that run on plaintext into versions that operate on ciphertexts.
HEIR is designed as a “one-click” solution so non-experts can add encrypted inference to production without needing a team of cryptographers. Google teamed with hardware accelerators Belfort, Niobium, Cornami, and Optalysys, and the project has become a research platform with collaborations at Georgia Tech, Carnegie Mellon, UC Santa Barbara, Purdue, Tsinghua, the University of Edinburgh, and others. The team released demonstrations including encrypted credit card fraud detection, network-threat detection, and a hotword detector. Critics on Hacker News noted fully homomorphic encryption has long been “horrifically slow” and that demos must be significantly stripped down, but Google argues the computational cost is falling rapidly — moving the privacy/security trade-off from “impossible” to “how much will you pay for it.”
DeepSeek introduces peak and off-peak pricing as V4 hits GA
Chinese AI leader DeepSeek updated its API pricing with a new peak/off-peak model, announced alongside the GA (general availability) release of its V4 lineup. Under the new structure, off-peak rates are 50% lower than peak rates, a pricing framework already familiar from cloud services but rare in the LLM API market.
Peak hours are 01:00–04:00 and 06:00–10:00 UTC. For deepseek-v4-flash, off-peak effective prices are $0.22 per 1M cache-miss input tokens and $0.66 per 1M output (half of the peak rates); for deepseek-v4-pro, the corresponding off-peak rates climb to $0.66 and $1.98. The expanded peak/off-peak model is aimed at shifting flexible workloads to cheaper off-peak windows, and mirrors a broader trend observers say is inevitable as inference demand scales and datacenter utilization becomes a key competitive lever.
Anthropic publishes an efficiency playbook for Claude Code sessions
Anthropic’s Lydia Hallie published “Maximizing the Value of Your Claude Code Sessions,” a practical field guide to getting the most out of every token spent on agentic coding. Why it is central to the AI conversation: with subscription and API pricing, a finished task has a price tag, and prompt engineering at the session level now directly shapes developer cost.
Key recommendations: run /clear between tasks to avoid til irrelevant context being re-sent; set your model and effort level (/model, /effort) before starting, since switching mid-conversation busts the prompt cache and forces a full re-prefill at full price; @-mention files instead of naming them to skip extra read calls; add quiet flags to noisy commands or run them in a subagent; and use /compact before a long break since the prompt cache expires after an hour on subscription (five minutes on an API key). Hallie explains the underlying token pricing mechanics (input vs output, prefill vs decode) and why one “square fix” can cost different amounts of tokens depending on how many files the model read along the way.
Is AI’s math edge really about working memory large humans can’t match?
A popular essay by Davide Piffer (titled “AI Isn’t Outthinking Mathematicians. It’s Out-Remembering Them.”) argued that the main AI advantage on math may not be superior reasoning but a virtually unlimited symbolic working memory — the giant context window and impressive external “notebook” that lets a model hold the whole problem, hundreds of intermediate equations, abandoned branches, and constraints all at once.
The post reviews psychological research on how a limited human working memory constrains mathematical performance — including Alloway and Alloway’s six-year longitudinal study in which early working-memory performance predicted later numeracy even after controlling for IQ, plus other studies showing working memory predicts mathematical achievement beyond general intelligence. Piffer’s argument, which drew more than 400 comments on Hacker News, is that part of what we label “machine intelligence” actually reflects the nearly unlimited symbolic workspace humans didn’t evolve with, much as scratch paper expands effective working memory by letting us externalize what we’re juggling. He concedes the context-window advantage is not equally useful across all forms of reasoning, and that advertised context length is not the same as perfectly usable memory — models can still lose track — but the sheer difference in potential capacity remains enormous.
Together these stories trace the shape of 2026: model vendors competing on cost and cadence, AI hardware and infrastructure, privacy-pressure on approaches like homomorphic encryption, riding agentic tooling, and a deeper, philosophical debate about what we actually mean when we say an AI is “smart.”
