7 min read · 14 sources

The AI price war just got a lot more literal. Within an hour of each other, Anthropic and OpenAI both shipped new models: Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna. The headline number is Luna’s price tag - $0.10 per million input tokens and $0.50 per million output, half of what GPT-5.6 Luna cost. That is not a promotional teaser. That is the list price.

The consequence for anyone building on these APIs is immediate: the cost floor for a capable model just dropped by half, and Opus 5.5 is undercutting its own predecessor. If you were waiting for a reason to re-architect that RAG pipeline or spin up more agents, this is it. Here is everything that changed today.

GPT-5.6 Luna scores 75% on GPQA Diamond for $0.0004 per question, a 725-fold price drop from o3’s 30 cents in 18 months.

GPT-6 Luna Makes Cheap Models Actually Good

Source: simonwillison.net ↗

GPT-6 Luna is the budget tier of OpenAI’s new pair, priced at $0.10/M input and $0.50/M output. That is half the price of the promotional GPT-5.6 Luna pricing, and it lands in a territory where “cheap” and “capable” stop being mutually exclusive. For comparison, GPT-6 Sol - the flagship - costs $2/M input and $10/M output, matching GPT-5.6 Terra’s price.

What this breaks is the old assumption that cost-efficient inference meant sacrificing quality. Luna is not a toy model; it is OpenAI’s answer to the question of what happens when you stop charging a premium for intelligence. For engineers running high-volume workloads - classification, extraction, summarization at scale - the math changes overnight. The same budget now buys roughly twice the tokens, which means either twice the throughput or a serious cut to your inference bill.

Claude Opus 5.5 Answers the Communication Complaints

Source: simonwillison.net ↗

Claude Opus 5.5 is Anthropic’s counterpunch, and it is aimed at a different weakness. Opus 5.0 was criticized for being verbose and hard to work with; 5.5 is explicitly built to be more token-efficient and to work across every effort level. It matches the intelligence of Fable 5.1 while charging less per token than Opus 5.0 did.

That matters because token efficiency is a real cost multiplier. A model that says the same thing in half the tokens halves your bill even at an identical per-token price. Anthropic is betting that developers who left for cheaper or more responsive models will come back for a model that is both smarter per dollar and less annoying to prompt. The two releases within an hour of each other are not a coincidence - this is a pricing war, and both sides fired.

The Plunging Price of Thought, Quantified

Source: epoch.ai ↗

Epoch AI’s analysis puts a number on what today’s launches are part of: the cost of achieving a given level of AI performance has fallen about 47% per quarter since 2023, or roughly 13x per year. That is faster than the historical price declines of DNA sequencing, compute, lithium batteries, or electricity.

The concrete example is stark. OpenAI’s o3 cost 30 cents per question to score 75% on GPQA Diamond in January 2025. Eighteen months later, GPT-5.6 Luna scored the same for $0.0004 per question - a 725-fold drop. The fastest price declines happen right after a performance level debuts as state-of-the-art, which means the smart play is rarely to be first, but to be early on the second wave. For teams budgeting AI spend, this is the trend to bet on: capabilities keep rising, and the price of harder tasks keeps falling.

Intrinsic Open-Sources Its Robot Building Blocks

Source: therobotreport.com ↗

Intrinsic, the Google robotics group, released Intrinsic Core at ROSCon 2026 in Toronto. It is an open-source set of ROS-compatible capabilities for building robotic applications, available on GitHub under a permissive Apache 2.0 license. Think of it as pre-configured building blocks for robot behavior - the kind of glue code that usually eats hundreds of hours of a robotics engineer’s life.

For the ROS ecosystem, this fills a real gap: there is no standard, well-maintained set of high-level behaviors you can drop into a robot stack. Intrinsic is betting that sharing its core software builds a community around its platform, and the Apache 2.0 license means no copyleft headaches for commercial products. If you are building on ROS, this is worth a look before you write another motion-planning wrapper from scratch.

LiteParse: Local PDF Parsing at 2-5ms Per Page

Source: github.com ↗

LiteParse is a standalone open-source PDF parser from the LlamaIndex team, focused on being fast and light. It uses PDFium for spatial text extraction at roughly 2-5ms per page, with optional OCR via Tesseract or HTTP servers, and outputs Markdown, JSON, or Text with bounding boxes. It runs fully locally - no cloud dependencies - and supports Rust, Node.js, Python, and WASM, with a worker pool mode for parallel parsing.

For RAG pipelines, this is the cheap, fast option that keeps your documents on your own hardware. The trade-off is that complex documents - heavy tables, unusual layouts, scanned pages without OCR - may still need the cloud-based LlamaParse. But for the common case of extracting text from clean PDFs, LiteParse is a drop-in that cuts both latency and cost.

Bug Blindness Is a Skill You Can Learn

Source: danluu.com ↗

Dan Luu’s argument is that most of us are “bug blind”: we hit the same software bugs repeatedly and do not notice them, because we have internalized the non-intuitive workarounds that get around them. He claims this blindness can be cured in a few weeks by actively pointing out bugs, and notes that when he evaluates products for executives, he finds severe issues that internal comments claim work well.

The operational takeaway is blunt: your product is probably more broken than you think, and the people who notice are the ones who deliberately cultivate the habit. Luu says he now uses LLMs to reproduce issues the way a normal user would, which is a cheap way to simulate fresh eyes. Fixing what you notice is a skill; noticing at all is a discipline.

Xbox Restructures, Sells Studios, Cuts Layers

Source: arstechnica.com ↗

Xbox’s “reset” continues with a dramatic restructuring. Obsidian moves into Bethesda to work on a new Grounded game and a “new Fallout project” with Bethesda Game Studios; Playground Games merges with Turn 10. Microsoft is looking to sell Ninja Theory and Arkane Studios, and has already divested Compulsion, Undead Labs, and Double Fine.

The engineering angle is the management structure: the consolidation aims to reduce layers, which could be up to 14 deep in places. That is part of a planned 3,200 layoffs, now roughly three-quarters complete. Fewer layers means faster decisions, but also fewer people to carry the load.

Apple's Screenless Whoop Rival

Source: bloomberg.com ↗

Apple is developing a screenless health and fitness tracker as part of a rethink of its smartwatch lineup. The project is preliminary and likely won’t debut before 2028, but it signals a direct challenge to Whoop’s subscription-based, screenless approach. Apple is also working on a major smartwatch update for as early as late 2027.

Quick Hits

Source: sciencealert.com ↗

• Alzheimer’s is no longer considered untreatable, with hundreds of clinical trials on medicines, diagnostics, and lifestyle interventions. A cure is far off, but earlier diagnosis and a wider range of interventions make optimism defensible.
• Amazon’s drone deliveries are overwhelming a Texas suburb - residents closest to the flight paths report constant noise.
• AI generalization remains mysterious: training a model to write insecure code made it immoral in general, and training on 19th-century bird names made it behave like a 19th-century person. Some safety researchers see this as hopeful for alignment.
• Fix bugs before writing features, argues a post citing “The Mundanity of Excellence”: superlative performance is a confluence of small wins, and unstable foundations produce unstable products.
• AI will worsen GDP estimates because quality improvements are hard to measure, making real GDP too conservative while nominal GDP stays accurate - a problem for monetary policy and market signals.
• Jev, a new AI model, is not an LLM but a BERT-style classifier claiming to be 193x faster and 444x cheaper than LLMs, with no hallucination risk. Its exact architecture is undisclosed.
• Defensive driving for your career: take public credit by owning 1-3 important announcements, because visibility is often the only way organizations value work.