GPT-6.2 drops, Microsoft's $0.01/min model, and agents that build their own models
6 min read · 15 sources
- OpenAI released GPT-6.2 Astra (frontier) and GPT-6.2 Sol (efficient) to ChatGPT Work, Codex, and the API.
- Microsoft unveiled seven MAI models, with MAI Cyber 2 Flash beating GPT-6 Astra quality at under 2% of the cost per task.
- MAI Realtime 1 starts at $0.01 per audio minute, and Maia 200 delivers 1.7x more tokens per watt.
- A Horizon Labs agent ran 24/7 for seven weeks, built 38 Hugging Face models, and saw 50,000+ downloads.
- Runta's FrontierHarness Eval v1.1 found harness choice alone changes cost per task by up to 5x.
OpenAI fired first with GPT-6.2 Astra and Sol, but the real story is the price war Microsoft just declared. At Ignite, Redmond shipped seven models that undercut OpenAI’s frontier line by 50x or more on cost per task, and it’s not just marketing math. If you run agents at scale, the numbers below change your infrastructure bill overnight.
The second big thread today is that models are starting to build other models. Two independent experiments - one from a researcher poking at weights by hand, one from an autonomous agent with a blank Hugging Face account - produced working, adopted ML artifacts without a single gradient descent step. That’s not a lab curiosity anymore; it’s a new deployment pattern you need to plan for.
Microsoft’s MAI Cyber 2 Flash beats GPT-6 Astra’s quality at less than 2% of the cost per task on CyberGym-E2E.
Microsoft's Seven-Model Answer to GPT-6.2
Microsoft AI dropped seven new MAI models at Ignite 2026, and the pricing is the headline. MAI Cyber 2 Flash beats GPT-6 Astra’s quality at less than 2% of the cost per task on the CyberGym-E2E benchmark - that’s a frontier-level security model for pocket change. MAI Code 2, now in GitHub Copilot, hits 98% of GPT-6.1 Sol’s coding quality at roughly 40% of the price.
The rest of the lineup rounds out the attack: MAI Image 2.7 for editing, MAI Realtime 1 for voice starting at $0.01 per audio minute, Dragon Copilot Transcribe, and Decision 1.1 - 50x faster and 79x cheaper than GPT-6.1 Sol for calibrated decision-making. Underneath it all sits Maia 200, Microsoft’s second-gen accelerator now in production at 1.7x more tokens per watt.
For engineers running 24/7 agent workloads, this is the cost curve breaking. Frontier quality at commodity prices means the economics of always-on agents just flipped. If you’ve been rationing model calls, you can stop.
Replit Missions: Agent Swarms With a Budget Cap
Replit Missions lets you define an outcome and let a swarm of agents plan, build, test, and iterate toward it - within a budget you set. The internal proof point is DataCat, a monitoring platform with dashboards, alerts, and tracing, built in two days for under $3,000.
The interesting part for operators is the guardrails: configurable spending limits and approval requirements for actions like contacting prospects. This isn’t a toy. Replit is positioning it for enterprise sales and legacy software replacement - multi-step work that previously needed a team. The swarm handles the delegation; you handle the boundaries.
Weight-Space Programming: Building Networks Like It's 1995
This write-up of weight-space programming experiments is the most fascinating thing today. The author had Claude Fable 5.2 write all 5,994 weights of a fixed CNN for MNIST classification from scratch - no fitting, no regression, no search. Only training loss, accuracy, and confusion matrix feedback. Claude beat Opus 5.5 considerably, and the resulting network used hand-crafted filters and templates straight out of the pre-deep-learning era.
A second replication confirms the pattern with GPT-6 Astra in a Docker container with no internet, Python, and NumPy only. The model could identify a network’s task from raw weights, write a game-playing policy network from scratch, and improve models by editing weights directly.
This is “LLM Descent” - a training loop where frontier models author weights without gradient descent. It’s early, but it points at a future where you don’t train models, you edit them. That changes what “training” means for anyone running a model stack.
Horizon Labs: One Agent, 38 Models, 50,000 Downloads
Horizon Labs ran the adoption test. An Opus 5.5-based agent, running 24/7 in Claude Code for seven weeks, autonomously created 38 models on Hugging Face. The bill: 600+ GPU-hours, ~1.9 billion tokens processed (96% from prompt cache), at an API cost of ~$1,030.
The models - 141M to 568M parameters - focused on narrow tasks like prompt-injection classifiers and PII redaction. They’ve been downloaded over 50,000 times, with real use in 143 GitHub repositories and two arXiv papers. That’s not vanity publishing; that’s an agent producing genuinely adopted ML artifacts at scale. The risk profile is real too - an agent with a blank account and compute can ship a lot of code, and not all of it will be good.
Vercel's eve 1.0 Goes GA
eve 1.0 is Vercel’s open-source framework for production agents, now generally available after 230+ releases in public preview. It’s a harness that calls models, runs tools, and delegates to subagents, plus a Workflow SDK for durable execution, multiplayer sessions, and agent-to-agent calls. Managed hosting is on Vercel.
The adoption numbers matter: Rillet tripled shipping rates, and Clay saw 30-40% weekly usage. For engineers, eve’s pitch is model-agnosticism with built-in observability and security controls - the missing pieces that keep agent frameworks in demos instead of production.
FrontierHarness Eval v1.1: Your Harness Is a 5x Cost Lever
Runta’s FrontierHarness Eval v1.1, built with Baseten, is the first thing to read before you pick an agent harness. It’s a leaderboard from 83B tokens across 27,280 agent runs - 22 harness configurations × 5 models × 62 tasks × 4 attempts.
The findings are brutal for anyone who standardizes on one harness. The best-value harness varies by model: Codex CLI pairs best with GLM-5.3, ZCode with DeepSeek V4.1 Flash. The median harness moved 11 of 22 places in cost rank between models, and Codex CLI dropped from 1st to 16th. Harness choice alone changes cost per task by up to five times.
The takeaway: switching models should trigger a harness re-evaluation, not a swap-and-hope. Your “standard” setup is likely leaving money on the table.
NVIDIA's Sparse Attention Primer for Long Contexts
Source: developer.nvidia.com ↗
NVIDIA’s post on co-designing sparse attention is a practical deep dive for anyone pushing past 128K-token contexts. Each query attends to only the k most relevant keys instead of the full context, cutting compute and memory traffic.
The catch: the indexer still scores all N keys, and it takes 47% of prefill time in DeepSeek V3.2. NVIDIA sweeps the design knobs - query heads, compression ratio, speculative-decoding depth - against a production FP8 kernel. If you’re building long-context inference, this is the hardware-aware guidance you need to keep indexer and Top-k costs from eating your gains.
Cerebras Goes Production With Disaggregated Inference
Cerebras’ disaggregated inference is now in production: GPUs handle prefill, Cerebras Wafer Scale Engines handle decode. The result is five times more tokens per megawatt at ultrafast speeds, available via Cerebras Inference Cloud and for infrastructure partners.
This is a direct answer to the inference capacity crunch from agentic apps like Muse and Copilot. When decode speed is the bottleneck for interactive agents, splitting the pipeline across specialized hardware is the pragmatic move. Cerebras is betting its decode advantage is worth the architectural complexity.
Redwood's Capabilities Definition Could Shape Compute Allocation
Source: blog.redwoodresearch.org ↗
Redwood Research’s proposal for defining “capabilities activities” is policy work with direct engineering consequences. The goal is pacing frontier development by restricting compute used for capability improvements, via legislation or voluntary commitments.
The definition aims to be robust to adversarial reframing and simple to adjudicate, excluding narrow safety and product development. If this sticks, AI companies will need to report compute allocation by category - which means someone has to build the telemetry for it. That’s a compliance burden landing on your infrastructure team.
Gemini Can Now Call Businesses and Pay Bills
Google’s Gemini app, running on Gemini 4 Hydrogen, can now make phone calls and pay bills - free, rolling out to hundreds of millions of users in the U.S., Japan, Korea, Canada, Australia, and India. It books appointments with AI disclosure, analyzes utility bills via Google Password Manager to find savings, and manages calendars.
This is agentic capability for real-world tasks, integrated with Google’s ecosystem. The engineering significance is the pattern: an agent with access to personal data, a spending limit, and a phone line. That’s the same architecture you’re building for enterprise agents, just pointed at consumer workflows.
You May Also Like
Gemini 4 Carbon ships at frontier prices, Anthropic files for $60B IPO, and 35k Grok apps leak data
Google released Gemini 4 Carbon, its strongest model yet, at $2/$10 per million tokens with a #1 ranking on CursorBench 4.0 and a new /swarm multi-agent mode. …
Anthropic’s Public Benefit Fund, OpenAI’s Aurora Test, and DeepSeek’s New APIs
Anthropic locked 5% of its equity into a Public Benefit Fund, with a formula that could push the stake to ~20% as valuation climbs. OpenAI quietly surfaced …
BootLoops reframes Claude for science as the decision model ecosystem goes open-source
Anthropic and Matthew Schwartz rolled out BootLoops, aligning LLMs directly to quantitative calculations to map mathematical overlaps across ecology and …




