Nvidia hikes AI server prices over 15%

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

💰 Nvidia hikes AI server prices over 15%

📌 OpenAI cuts GPT-5.6 API price over 20%

📊 Nvidia agent aces tough AI test

🕵️ Free frontier coding model emerges

🤖 Opus 5 overtakes Fable 5

Plus: 💡 5 strategies & tactics, 🎁 6 other news you might like, 🧰 6 tools, and 📚 5 papers.

💰 Nvidia hikes AI server prices over 15% LINK

Nvidia is raising prices on servers built around its AI chips by more than 15% in many cases, driven by soaring memory costs, with the increases hitting systems shipped early next year.

The hikes apply to configurations using the flagship Vera Rubin and Grace Blackwell chips, varying by chip generation and DRAM setup, with contract server builders already notifying customers like Microsoft, Google, and Oracle.

Surging demand has handed memory makers Samsung, SK Hynix, and Micron unusual pricing leverage, and even Nvidia, running a 75% gross margin, couldn't absorb the cost, though whether rivals gain an opening depends on customers securing their own memory supply.

📌 OpenAI cuts GPT-5.6 API price over 20% LINK

OpenAI is dropping GPT-5.6 Sol API prices by more than 20% for the next three months, a temporary cut aimed at competition from Anthropic and Chinese frontier models.

The discounted rate lands GPT-5.6 Sol at $4 per 1M input tokens and $20 per 1M output tokens for standard short-context use, rolling out across eligible API plans plus ChatGPT Work and Codex credits.

That undercuts Anthropic's Claude Fable 5 ($10 input/$50 output) and roughly matches Claude Opus 5 ($5 input/$25 output), though the reduced pricing only holds for three months.

📊 Nvidia agent aces tough AI test LINK

NVIDIA's Agentic Variation Operators (AVO) agent system hit a perfect 100.00 RHAE score on the ARC-AGI-3 public benchmark, solving all 183 levels across 25 environments with the same architecture built for GPU-kernel optimization.

Driving AVO with Claude Opus 5, the system cleared the public set in 6,624 environment actions, about 12% fewer than VISTA's 7,542-while running text-only on 64x64 grids, and its earlier attention-kernel run beat FlashAttention-4 by up to 10.5% over seven autonomous days.

NVIDIA credits AVO's persistent memory and supervisor loop rather than raw model strength, since ARC Prize separately reports ~30% for Claude Opus 5 alone, though the team stresses the VISTA comparison isn't a controlled ablation and these are public-set, not private-competition, results.

🕵️ Free frontier coding model emerges LINK

A free stealth model called Ox Alpha appeared on OpenRouter Thursday, pitched as a reasoning model built for coding, sustained agentic work, and production workloads.

The listing says Ox Alpha is "developed and operated by a third-party provider who has chosen to remain anonymous," with Stripe CEO Patrick Collison, whose company is acquiring OpenRouter, calling it "very impressive."

Speculation about the builder has centered on China's Z.ai GLM models and Microsoft's unreleased MAI, though the provider stays anonymous during the preview and no benchmark numbers have been published.

🤖 Opus 5 overtakes Fable 5 LINK

Anthropic's cheaper Opus 5 overtook its flagship Fable 5 in corporate spending within a month of launch, as businesses increasingly judge models by cost per completed task rather than raw capability.

Released July 24 at $5 per million input and $25 per million output tokens, half Fable's rates, Opus 5 matched or beat Fable on bounded coding and office work in Anthropic's evaluations, while Fable stays aimed at days-long autonomous projects.

Ramp put Fable near 11% of attributed Anthropic spending by August 23 but did not disclose Opus 5's exact share, and Vercel's July gateway data conflicts, placing Fable at 13.2%-though that sample predates Opus 5's near month-end launch.

💡 Strategies & Tactics

> EP223: Ollama vs vLLM vs SGLang: Pick Ollama for local prototyping, vLLM for high-traffic serving, and SGLang for AI agents with overlapping prompts, since each engine handles requests differently.

> Why real-time AI at scale is so hard: Real-time AI breaks at scale because slow feature lookups, stale data, and rotting search indexes, not the model, degrade speed and accuracy, so isolate workloads to prevent it.

> Study explains why AI agents benefit from "skills" and when they fail: AI agents improve mainly because skills give them a reliable step-by-step process to follow, not extra facts, but oversized skill libraries make finding the right one nearly impossible.

> How Claude Watermarks AI-Generated Text: Anthropic marks Claude's text with an invisible pattern woven into word choices, letting only Anthropic detect whether text is AI-generated.

> The Evolution of the Agent Harness: As AI models absorb the scaffolding that once surrounded them, the harness's real future is managing when agents interrupt or defer to scarce human attention.

Other news you might like

• IBM’s next-gen mainframe chip is the first to run Arm and Z workloads on the same coresLINK

• Anthropic Expands Mythos 5 Access to More Defenders, Unveils $35M Open Source FundLINK

• Windows could soon let you control how your PC splits memory for gaming and AILINK

• Inherent, founded by DeepMind alumni, says its AI ‘teammate’ just outperformed Anthropic and OpenAI at replicating researchLINK

• Anthropic’s best AI model struggles to attract users as cheaper tools thriveLINK

• Ramp launches its own AI model router, called RouterLINK

🧰 Trending tools

Offloop: shared workspace where teammates and AI agents plan and track multi-step work in channels, with clear stage ownership across handoffs.LINK

Treebar: monitors Git worktrees and Codex agents from your MacBook notch, showing which branches are active and what each agent is changing.LINK

whatsapp-claude-plugin: a plugin that lets you run Claude AI from WhatsApp, offering voice transcription, remote tool approval, and access control.LINK

emisar: an MCP that lets AI tools securely connect to infrastructure, write IaaS code, debug issues, and assist during incidents.LINK

emulo: mines your Claude Code and Codex logs to build a local you.md agent profile capturing your coding patterns.LINK

Decawork: adds an approval layer to internal AI agents, letting you control their access and permissions while logging every action they take.LINK

📚 Trending papers & reports

Human feedback training can teach AI models the same quality using roughly 10x less human-labeled data, matching results from 200K labels with fewer than 20K, sharply cutting the cost of aligning models.LINK

AI agent memory can be turned into a compact image instead of ever-growing text, cutting token and memory costs by over half while keeping more than 95% of task performance.LINK

Self-teaching for reasoning models lets a model improve itself without a stronger outside teacher, fixing the instability and reasoning decay that plagued earlier self-training methods by rewarding better answers over merely copying its own hints.LINK

Diffusion text generators can now grade their own answers by re-checking how confidently they would rewrite each word, letting them flag low-quality output and adjust response length automatically, closing a key reliability gap in this newer model type.LINK

AI agent workflows get tested on whether letting them revisit earlier steps helps or hurts, showing it aids recovery in messy tasks but wastes time and cost in straightforward step-by-step ones.LINK

See you tomorrow for a new dose of ☕️ AIpresso!