What happened this week in AI by Louie
This week, OpenAI’s GPT-5.6 Sol rewrote the production GPU kernels that serve it, cutting end-to-end serving costs by 20%, while an internal OpenAI model called Astra delivered ten new results in mathematics and theoretical computer science, each formally proved in Lean. The cost of intelligence had its own breakthrough at the cheap tier: OpenAI cut GPT-5.6 Luna’s API prices by 80%, DeepSeek shipped a far stronger V4 Flash at unchanged rock-bottom rates, and Qwen launched a Max-tier model with weights promised next week. The loop is direct: cheaper work per dollar funds longer agents and wider searches, and stronger models lower the cost of the next unit of work.
The cheap tier now sits close behind the frontier. On Artificial Analysis’s Intelligence Index as of August 4, DeepSeek V4 Flash 0731 at max effort scores 50 for about $0.03 per weighted benchmark task; after its price cut, Luna scores 51 for about $0.05, running faster and taking images. Flash 0731 re-post-trains April’s 284-billion-parameter architecture (13 billion active, one-million-token context), and its cache turns long agents into near-free reruns by billing cached input at 2% of the miss rate: a 50-call agent with a stable 150,000-token prefix drops from $1.13 to about $0.12. But you need to plan around announced peak-hour pricing (twice the base rate) and personal data stored in China on the direct service.
Another important level is effort settings. Luna at high effort scores 46 for about $0.02 in 16.5 seconds, against 51 for $0.05 and 135.5 seconds at max. A good rule of thumb is to route each task to the lowest effort that clears its quality bar.
Qwen3.8-Max is the new open-weight pressure point near the frontier: 2.4 trillion total parameters (95 billion active), a one-million-token context, and text, image, and video input at $2 per million input tokens and $6 for output, with implicit-cache input at $0.25. Qwen’s own numbers land near GPT-5.6 Sol: 86.6 versus 88.8 on TerminalBench 2.1, 93.0 versus 90.5 on PaperBench, and 86.1 versus 83.2 on OSWorld Verified. The benchmark to beat is Kimi K3, the reigning open-weight leader at 57 on the Index for $0.86 per task; Qwen is likely to be significantly cheaper.
Sol’s engineering work shows what frontier models can now do to production systems. It analyzed OpenAI’s live traffic, tuned routing and load balancing, and autonomously rewrote production GPU kernels, verified with a Floating-Point Sanitizer inside a human-led process. OpenAI credits the kernel work with the 20% serving-cost cut, and separate Sol-run speculative-decoding experiments with more than 15% higher token-generation efficiency. The economics are very large. The Information put OpenAI’s inference costs at $8.4 billion for 2025 and forecast $14.1 billion for 2026; a 20 percentage point saving here could easily reach the $ billions. Much of the gain likely surfaces as capacity rather than lower spend, but a model generating hundreds of millions of dollars of annual value for its own operator marks AI improving the economics of its own use.
The same capability is also now extending the frontier of human knowledge itself. OpenAI’s unreleased next model, Astra, (maybe GPT-6?) made ten math breakthroughs across sphere packing, coding theory, group theory, operator algebras, circuit complexity, quantum complexity, lattice cryptography, and extremal combinatorics: an explicit non-sofic group disproving the soficity conjecture, a disproof of the Connes rigidity conjecture, the exact asymptotic rate of the Cohn-Elkies sphere-packing programme, stronger lower bounds for the permanent, a quantum parallel repetition theorem, and solutions to three Erdos problems. These are major research findings; several would individually anchor a strong research career, and they arrived as a batch across fields that each take years of specialist training to enter. OpenAI estimates the discovery tokens would cost about $2,000 at Sol API rates. The all-in cost including training and failed paths is far higher, but the marginal price of searching for new mathematics has collapsed.
Where expert review exists, it is strong: Andreas Thom, whose 2019 theorem the non-sofic proof builds on, calls the key construction creative and clever and says he had sought exactly such a mechanism since that work. Mathematics is also precisely where LLM research should land first. Lean makes correctness cheap and mechanical to check, so a model can run enormous searches and surface only verified survivors. These ten examples cover a narrow slice of frontier knowledge creation, but it is genuinely the frontier, and a model produced it.
The field is adjusting fast. Terence Tao expects proof production and formal checking to accelerate faster than the digestion of results into shared understanding, making question choice and exposition the scarce work. Jacob Tsimerman, who is joining OpenAI, expects interesting mathematical output to rise ten or one hundred times. The entry barrier has already collapsed: in April, a 23-year-old student, Liam Price, cracked a 60-year-old Erdos problem on primitive sets with a single prompt to GPT-5.4 Pro, after he and Cambridge undergraduate Kevin Barreto had spent months feeding randomly chosen open problems from the Erdos database into ChatGPT, with mathematicians then confirming the solution was genuinely new. Another result was recently found simply by asking GPT-5.6 Sol to “make a breakthrough,” followed by several “continues”. Mathematical work will now follow the path software took: less time producing every step by hand, more time choosing questions, directing searches, checking that the formal target matches the intended claim, and turning proofs into ideas others can build on. The binding constraint is moving from generating results to reviewing, understanding, and using them.
Why should you care?
The cut in serving costs and how it was achieved changes the routing playbook for tasks. In my experiments, Luna now works very well as worker threads managed by Sol or Claude Fable 5, particularly for routine data collection and structuring tasks. Codex sub-agents cannot run Luna yet, but threads in Codex get you similar results. With a 25x task-cost gap between orchestrator and worker, worker cost rounds to zero, so the design question shifts to how much structure and verification the orchestrator imposes on each thread. Keep stable instructions at the front of prompts so caches reuse them, escalate only the judgment-heavy share, and measure the total cost of a trusted answer, including retries and review.
A key takeaway is that AI progress concentrates wherever verification is cheap. Kernels have sanitizers and latency benchmarks, proofs have Lean, code has test suites, so those fields get the recursive gains first: models improving their own serving economics, models producing new mathematics, etc. If your domain lacks a cheap verifier, building one is now the highest-leverage move available, because search has become nearly free and checking is the bottleneck. Expect open-problem databases to be swept systematically; curating good problem lists and verification infrastructure is becoming valuable work in itself. And as generation costs collapse, volume of work output in many fields will inflate: it gets more and more important to track cost per validated result and whether another expert can understand the mechanism. The groups that select, verify, and explain will beat the groups that simply produce.
— Louie Peters — Towards AI Co-founder and CEO
Introducing Towards AI Mentorship, giving everyone on call access to senior AI engineers for production support and career guidance, at a price that isn’t locked behind enterprise budgets.
For anyone building with AI or working toward an AI role: post technical or career questions any day and get written answers from our 15-person engineering team. Live Q&A twice a week across two time zones. Resume and portfolio reviews in 48–72 hours. A production blueprint every month with real engineering decisions explained.
Includes Python for LLMs and LLM Fundamentals courses. $99/month. Cancel anytime.
Hottest News
1. DeepSeek Releases DeepSeek-V4-Flash-0731
DeepSeek released V4-Flash-0731 as the official successor to its Flash preview, while the API remains in public beta. The model keeps the same 284B-parameter MoE architecture, with 13B active parameters, and focuses on stronger agent and coding performance. DeepSeek reports scores of 82.7 on Terminal-Bench 2.1, 54.4 on DeepSWE, and 76.7 on CyberGym, though some evaluations use internal tooling. The API now supports the Responses API and Codex. Pricing remains $0.14 per million uncached input tokens and $ 0.28 per million output tokens, while the MIT-licensed weights are available on Hugging Face.
2. OpenAI’s Unreleased Astra Produces Solutions to 10 Long-Standing Math and Theory Problems
OpenAI says an internal version of Astra, described as its “next major model,” resolved or substantially advanced ten long-standing problems in mathematics and theoretical computer science. Results include an explicit construction of a non-sofic group, a counterexample to Connes’s rigidity conjecture, the sharp Ehrhart volume bound, and progress on sphere packing and three Erdős problems. OpenAI published a 249-page manuscript and Lean 4 formalizations that can be independently checked. It estimates that the solution-finding tokens would cost roughly $2,000 at Sol API rates. OpenAI has not announced Astra’s final name or release plans.
3. OpenAI Cuts GPT-5.6 API Prices and Adds Sol Fast Mode
OpenAI reduced standard short-context API pricing for GPT-5.6 Luna by 80%, to $0.20 per million input tokens and $1.20 for output. Terra fell 20% to $2 and $12, while Sol remains at $5 and $30. Priority Processing has also been renamed Fast mode, offering up to 2.5 times faster processing at twice Sol’s standard price. OpenAI says model-assisted kernel optimizations helped lower Sol’s serving costs by 20%. ChatGPT and Codex subscription prices remain unchanged, though Luna and Terra now consume fewer credits against applicable usage limits.
4. MiniMax Releases MiniMax H3
MiniMax released H3, a multimodal video model that accepts text, images, video, and audio and generates clips of up to 15 seconds with native stereo sound. It supports text-to-video, image conditioning, reference generation, motion transfer, and instruction-based editing. The model produces 768p video first, with a separate regeneration stage used for 2K output. API pricing is $0.08 per second at 768p and $0.13 at 2K. MiniMax published the weights on Hugging Face under its custom Community License, which includes regional restrictions.
5. Google DeepMind Ships Three Physical AI Models
Google DeepMind introduced Gemini Robotics 2, Gemini Robotics ER 2, and Gemini Robotics On-Device 2. The main Robotics 2 model controls a humanoid’s full body under one learned policy, while ER 2 handles video understanding, planning, and coordination. On-Device 2 runs locally and can adapt to new dual-arm robots using fewer than 200 examples, according to Google. Only ER 2 is publicly accessible through Google AI Studio and the Gemini API; the other two remain limited to selected partners. Google also released ASIMOV-Agentic, a benchmark for testing robot safety and human-escalation decisions.
6. Onton Releases Ontology 1: A Neurosymbolic Search Model
Onton, a San Francisco-based search and discovery company, released Ontology 1, a neurosymbolic model for conversational and multimodal product search. It combines learned representations with symbolic reasoning over a custom knowledge graph to interpret intent, infer product properties, and identify questionable claims. In Onton’s own 90-query benchmark, it recorded a precision@10 score of 0.630, compared with 0.543 for Google Shopping and 0.469 for Amazon. The evaluation used three LLM judges and has not been independently replicated. Ontology 1 currently powers Onton.com, but no public developer API or downloadable checkpoint is available.
7. Alibaba’s Qwen Team Announces Qwen3.8-Max
Alibaba released Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model with 95 billion active parameters. It supports text, image, and video inputs, a context window of around one million tokens, and outputs of up to roughly 131K tokens. At launch, it ranked fifth overall in Arena’s text leaderboard and second in vision, though these positions change over time. Alibaba also demonstrated the model working autonomously on a software project for roughly 16 days. International API pricing starts at about $2 per million input tokens and $6 for output. Open weights were announced but had not been released as of August 4.
AI Tip of the Day
When we built the “Research and Writing Agent with MCP” lesson for our Agent Engineering course, we ran into a deceptively simple evaluation question: which articles should the judge score?
The easy option was to ask another LLM to generate a batch of articles and use them as the test set. But those articles would never pass through the system we actually wanted to evaluate. The judge might measure writing quality while missing failures in research, context transfer, or the handoff between agents.
So every evaluation example now follows the production path. We begin with the same brief a user would submit, run the research agent, pass its saved findings into the writing workflow, and score the final article.
This turns the output into a test of the entire system. If the research is weak, an important source is lost, or the writer receives incomplete context, the failure appears in the article, where the judge can detect it.
The takeaway: generate evaluation outputs with the workflow you plan to ship. Otherwise, you may be testing the quality of a substitute model rather than the reliability of your own system.
Five 5-minute reads/videos to keep you learning
1. Persistent Memory for Claude Code on MongoDB Atlas
Long Claude Code sessions silently lose constraints when compaction drops tool outputs and summarizes earlier exchanges. This article builds a plugin that captures those statements into MongoDB Atlas before they vanish, using a PreCompact hook and two MCP tools. Retrieval runs entirely inside the database: Automated Embedding generates Voyage vectors, $rankFusion blends vector and keyword pipelines through reciprocal rank fusion, and a hosted reranker orders results.
2. I Self-Hosted Langfuse so My LLM Traces Would Stop Living On Someone Else’s Bill
Langfuse crossed 100K monthly traces and pushed managed pricing into painful territory, so the author self-hosted the open-source observability platform instead. Docker Compose spins up Postgres, ClickHouse, Redis, and MinIO in minutes, though production requires TLS, SSO, and eventually Kubernetes with externally managed dependencies. The article flags a notorious UTC timezone bug, treats Redis queue depth as the real health signal, and lays out honest cost math: self-hosting secures data control immediately, but only beats managed pricing at scale.
3. How Spark Manages Memory — The Unified Memory Model, Spills, and AQE
A daily 2.1TB aggregation job kept dying at 85% completion, and the author traced the failure to Spark’s Unified Memory Manager rather than raw heap size. This article walks through how execution memory always wins over storage in the borrowing contract, why protected cached blocks starved a skewed aggregation of room to grow, and how memoryOverhead, not executor.memory, killed the YARN container. Adaptive Query Execution’s skew-join splitting fixed the root problem, cutting cluster cost while pinpointing exactly which memory region overflowed.
4. Agent Memory Is the Real Moat
Agent memory becomes dangerous the moment it is treated as an unbounded vector store rather than a governed lifecycle. This article maps four distinct memory classes: working checkpoints, episodic outcomes, semantic facts, and procedural lessons, each with its own ownership and retention rules. Benchmarks from TraceRetain and LoCoMo show unbounded memory collapsing under noisy writes while selective retention holds steady. It closes with a practical adoption blueprint.
5. Road to Bedrock AgentCore, From a Single API Call to a Production Agent
Building the same weather-fetching agent four times across AWS’s tooling stack exposed exactly what separates each layer. This article implements the same core function on the Bedrock Converse API, Bedrock Agents, the Strands SDK, and AgentCore, keeping the logic identical to isolate each layer’s tradeoffs. Converse demands hand-written orchestration. Bedrock Agents hands the loop to AWS at the cost of visibility. Strands restores control with model portability. AgentCore deploys the same agent unchanged, adding Gateway, Memory, Identity, and Observability as managed infrastructure.
Repositories & Tools
1. Evals is a benchmark framework that runs AI coding agents against real Supabase tasks like building schemas, debugging Edge Functions, and fixing RLS policies, scoring results to a public leaderboard.
2. MoonEP is an Expert Parallelism communication library that keeps token loads perfectly balanced across ranks via dynamic redundant experts.
3. Token Saver is a one-click Claude Desktop extension that runs local hybrid RAG over PDFs on your machine, sending only relevant passages to Claude and cutting token consumption by 90–99%.
4. Instella MoE is AMD’s fully open 16B-parameter MoE model (2.8B active) trained from scratch on Instinct GPUs.
Top Papers of The Week
1. Kimi K3: Open Frontier Intelligence
This technical report presents Kimi K3, a 2.8T-parameter Mixture-of-Experts model with 104 billion activated parameters, native vision capabilities, and a 1M-token context window. K3 introduces three architectural innovations: Kimi Delta Attention (KDA), which models linear retention recurrence for position-sensitive mixing; Attention Residuals, which improve information flow across model depth; and Stable LatentMoE, which activates 16 of 896 routed experts per token. Together, these yield an approximately 2.5x improvement in overall scaling efficiency over Kimi K2. Post-training applies reinforcement learning across general, agentic, and coding domains with multiple reasoning-effort levels.
2. Visual Prompt Engineering for Video Models
Text-based prompt engineering is standard for LLMs, but as video models become foundation models for visual tasks, this paper asks whether automatically modifying the task image can similarly improve performance. For example, converting an abstract sketch-like physics scene into a photorealistic version with a single call to an image editing model. The authors find that visual prompt engineering (VIPE) improves video reasoning performance across tasks, and for video models, it can be more effective than classic text-based prompt engineering or test-time scaling.
3. Shieldstral
Mistral introduces Shieldstral, a 3B-parameter policy-adaptive multimodal safety classifier built on Ministral-3B that matches or outperforms models nearly 7x its size on text safety benchmarks and sets a new state of the art on multimodal safety classification. The key design choice is formulating content moderation as a binary question-answering task: given a natural-language query describing a safety concern and a piece of content (text and/or image), Shieldstral outputs a yes/no answer rather than fixed category labels.
RLVR has driven recent progress in reasoning LLMs but remains limited to domains like math and coding, where correctness is deterministically verifiable. Open-ended tasks instead rely on reward models or LLM judges, introducing evaluation bias and capability bottlenecks. This paper proposes RLSVR, which transforms open-ended tasks into verifiable proxy environments whose internal rules automatically generate reward signals. The concrete instantiation, SpyRL, is a multi-agent self-play environment inspired by “Who Is the Spy?”: agents receive asymmetric information, complete the same target task, and vote to identify a designated spy. Because the spy identity is predetermined, voting outcomes provide fully verifiable rewards, while successful identification depends on output quality.
5. AISPA: User-Centric System Prompt Auditing for Large Language Model Applications
System prompts govern how LLMs behave in commercial products but are rarely disclosed to users or regulators, creating a trust and accountability gap. This paper introduces AISPA, a framework that audits system prompts along eight user-facing dimensions (identity transparency, information truthfulness, data privacy, action safety, user agency, unsafe request handling, harm prevention, and fairness), each grounded in corresponding articles of the Universal Declaration of Human Rights. The authors audited 3,249 instructions from system prompts across 88 commercial AI products, classifying each as protective or problematic.
Quick Links
1. Cogent AI team releases VR-1, a frontier reasoning model trained specifically for cybersecurity and designed to operate inside live enterprise environments. Unlike general-purpose models that find individual bugs in isolated codebases, VR-1 is trained for multi-step attack-chain composition, chaining minor weaknesses across cloud infrastructure, identity systems, and internal tools to prove viable paths to an objective. On IntrusionBench, a new enterprise attack-chain benchmark released alongside the model, VR-1 achieved 2x the performance of Kimi K3, Opus 4.8, and GLM-5.2 at roughly a quarter of the cost. Available only to vetted organizations through the Cogent Frontier Access Program.
Who’s Hiring in AI
Senior AI Engineer / Forward Deployed Engineer @Towards AI (London, UK)
Fullstack AI Engineer @Microsoft Corporation (Redmond, WA, USA)
Lead AI Engineer @UnitedHealth Group (Remote)
AI Product & Delivery Lead @ThermoFisher Scientific (Mississauga, Canada)
AI Software Developer @Leidos (Gaithersburg, MD, USA)
AI Engineer @Micron Technology, Inc. (Multiple US Locations)
Jr Gen AI Developer @NTT America, Inc. (India)
Interested in sharing a job opportunity here? Contact sponsors@towardsai.net.
Think a friend would enjoy this too? Share the newsletter and let them join the conversation.
https://dashmasters.io starts with simple controls, making it easy to understand, but each level quickly demands better timing as obstacles become harder to predict and mistakes send you back for another attempt again each time around.