Google ships Gemini 3.7 Flash at half price

Overview

9 stories in this issue. The first 3 are today's priorities.

Hot model developments

1. Top · Google ships Gemini 3.7 Flash at half price, three weeks after 3.6
2. Top · OpenAI ships two on the same day: GPT-5.6 Sol at 750 tok/s and cross-app memory
3. Top · DeepSeek open-sources its Harness agent framework under MIT
4. Writer launches Palmyra X6, post-trained on Z.ai's GLM-5.2

Global AI news 5. Anthropic set AI agents loose on one task — they started a turf war 6. Nvidia backs up to $500B in AI buildouts, guaranteeing aging GPUs 7. Someone compiled Doom into LLM weights, with a Hugging Face checkpoint

Regional & early signals 8. Singapore's Acrab reaches mass production on an on-device AI chip 9. On-device AI shifts from technical preference to supply-chain reality

Jiufeng graphic based on the sources cited in this issue.

Hot model developments

Google ships Gemini 3.7 Flash at half price, three weeks after 3.6

Google compresses frontier updates to a three-week cadence, using a temporary discount to push coding agents.

Google DeepMind released Gemini 3.7 Flash on August 13, replacing Gemini 3.6 Flash after just 23 days in production, with a focus on coding and agents. Product lead Tulsee Doshi described the model as the result of developer feedback and algorithmic changes, priced at an introductory rate of half the original 3.6 Flash cost per million tokens and running through year-end.

Limitations: the discount is temporary; from January developers must evaluate 3.7 Flash at its higher price. Ars Technica notes the update lands while developers still wait for Gemini 3.5 Pro, which Google said was in partner testing; RuntimeWire cites Google's own evaluation table showing the gains were uneven.

Image source: Google; mirrored on Jiufeng R2.

Source: Google DeepMind · Ars Technica · RuntimeWire

OpenAI ships two on the same day: GPT-5.6 Sol at 750 tok/s and cross-app memory

On August 13, OpenAI shipped an Ultrafast inference tier and cross-app Mac memory.

Cerebras and OpenAI previewed Ultrafast Mode on August 13, a new OpenAI API service tier powered by Cerebras. OpenAI says GPT-5.6 Sol on Ultrafast delivers up to 750 output tokens per second with no quality compromise; measured against output speeds reported by Artificial Analysis, Sol Ultrafast runs 11x faster than Fable 5 and 5x faster than Opus 4.8 on Fast mode.

The same day, OpenAI launched Computer History, an opt-in ChatGPT feature that records a user's activity across apps and websites on a Mac and draws on that record in later conversations. It is rolling out globally through the ChatGPT desktop app for Pro, Business, and Enterprise subscribers, is enabled under Settings > Integrations, and includes a timeline for reviewing prior work.

Limitations: Ultrafast is available initially to a select group of customers, with this preview powered by Cerebras for GPT-5.6 Sol; Computer History's availability in the EEA, UK, and Switzerland will follow in the coming weeks, and the feature is off by default, leaving users and employers to decide how much desktop activity an AI assistant should retain.

Source: Cerebras · OpenAI · RuntimeWire

DeepSeek open-sources its Harness agent framework under MIT

DeepSeek ships its in-house agent software as an MIT-licensed alternative to Codex and Claude.

DeepSeek released Deepseek Harness v0.1 as an MIT-licensed Developer Preview. Per The Decoder, Harness turns language models into autonomous agents through a modular plugin system, is pitched as an alternative to OpenAI's Codex and Claude, and is built on the newly released Cordis plugin system; a concurrent V4-Pro-0813 build scores higher on agent benchmarks but still trails Claude Opus 5 (63 points) and Kimi K3 (60). Chinese outlet Leiphone reports a leaked partial list of 17 open-source projects selected for the beta — roughly 70% near-zero-star individual or small-team work — concentrated in MCP plugins, coding agents, and agent runtimes; one example, the 224-star open-managed-agents, is described as an open replica of Claude Managed Agents focused on gateway-level credential injection and sandboxing.

Limitations: Harness is a Developer Preview; DeepSeek has not published weights for the new build (the April preview remains on Hugging Face). Leiphone's selection list is an unconfirmed leak, and is a Chinese-language source.

Source: The Decoder · Hugging Face · X/Deepseek · Leiphone

Writer launches Palmyra X6, post-trained on Z.ai's GLM-5.2

Writer post-trains the open GLM-5.2, claiming up to 50% lower cost on basic tasks.

Writer, which builds AI tools and agents for marketers, launched its flagship Palmyra X6 model on August 13, built as a post-training variation on Z.ai's open-source GLM-5.2. Writer says the model, combined with changes to its harness infrastructure, can cut customer costs by as much as 50% for basic tasks, aiming to balance open models' lower per-token cost against the difficulty of picking the right model for a job.

Limitations: the 50% figure is Writer's own estimate for basic tasks, not an independent benchmark, and rests on a single report from TechCrunch.

Source: TechCrunch

Global AI news

Anthropic set AI agents loose on one task — they started a turf war

Anthropic's red team found multi-agent systems clash, collude, and coordinate in ways current tests may miss.

Anthropic's Frontier Red Team published research on August 13 examining how groups of AI agents behave when they encounter one another in a shared environment. The tests found agents can clash, collude, and coordinate in unexpected ways, leading the researchers to question whether today's safety tests capture multi-agent risks as companies and governments move to run agents autonomously across shared codebases, markets, and computer systems.

Limitations: these are early red-team observations rather than a live incident; the work itself questions whether single-agent evaluation methods cover multi-agent interaction risks.

Source: Anthropic · TechCrunch

Nvidia backs up to $500B in AI buildouts, guaranteeing aging GPUs

Nvidia pledges its own money to keep pledged GPUs' value, unlocking financiers for data centers.

Nvidia said this week that Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR are willing to commit up to $500 billion to build AI data centers. Per TechCrunch, the bigger story is Nvidia's effort to create a secondary market for aging GPUs: to convince those financiers, Nvidia agreed to guarantee, with its own money, that its chips used as collateral will retain their value.

Limitations: TechCrunch notes the guarantee is risky for Nvidia itself — if the GPU collateral loses value, Nvidia is on the hook.

Source: Nvidia · TechCrunch

Someone compiled Doom into LLM weights, with a Hugging Face checkpoint

No training involved — a custom compiler writes Doom's renderer straight into transformer weights.

A developer (Hugging Face namespace physicsrob) used a compiler they wrote, torchwright, to translate Doom's original rendering algorithm directly into transformer weights — every weight computed, none learned. The result is a standard Hugging Face checkpoint that loads with no custom code: given a prompt containing level data, the player's position, and the viewing direction, the model generates the frame Doom would have drawn.

Limitations: this compiles a fixed algorithm into a network as a demo, not a trained or playable product; it is backed by the author's blog and the Hugging Face checkpoint.

Source: ood.dev · Hugging Face

Regional & early signals

Singapore's Acrab reaches mass production on an on-device AI chip

Under three years old, Acrab puts its first on-device AI chip, GΞLIX 1, into mass production.

Per Chinese outlet QbitAI citing DealStreetAsia, Singapore-based AI compute infrastructure company Acrab closed a Series B (Vertex Growth among investors), bringing total funding above $480 million, with $130 million this round. Founded in 2024, the company released its first-generation on-device AI chip GΞLIX 1 and an "Agent Box" personal AI hub built on it less than a month ago; the products have entered customer onboarding and mass-production preparation. Acrab is betting on rising edge-compute demand as models compress and high-frequency, privacy-sensitive, low-latency tasks migrate to endpoints.

Limitations: single Chinese-language source; the first-generation product is still in customer onboarding and production prep, with no independent shipment or performance data.

Source: QbitAI

On-device AI shifts from technical preference to supply-chain reality

Phones' and cars' demands for low latency, offline use, and power budgets are repricing small models.

Per TMTPost's English edition, the loudest LLM competition of the past three years centered on scale — more parameters, data, and cloud compute — while a parallel track focused on delivering capability within a device's memory, power, and thermal limits. As phones, cars, and other endpoints increasingly need AI that works with limited or no connectivity, local inference cuts latency, protects some data, and shifts inference cost onto hardware the customer already owns. The piece cites ModelBest — the Beijing company behind the MiniCPM series, spun out of a Tsinghua University lab in 2022 — whose researchers describe a "density law": by their measurements, the capability extracted per parameter roughly doubles every few months.

Limitations: this is an industry-trend synthesis; the "density law" is the firm's own empirical claim ("under their measurements"), not an independently verified result.

Source: TMTPost