Google Cloud adds OpenAI models, and Vercel makes agent harnesses crash-proof
OpenAI's GPT-6 models arrive on Google Cloud in private preview, ElevenLabs ships a drop-in realtime voice model, and Vercel adds durable sessions to existing coding agents.
Most of today’s news is about the layer around the model: who sells it, how it talks, and what happens to it when a machine dies. The model itself barely changed.
1 · Top story
OpenAI’s GPT-6 models are coming to Google Cloud
Google Cloud says three OpenAI models, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, are rolling out in private preview to select enterprise customers in Model Garden. Google calls itself the only hyperscaler offering all three leading closed-source model families. They’re sold through Google Cloud Marketplace, so usage lands on one invoice and counts against existing commitments.
If you buy models through a cloud, this changes the procurement conversation more than the architecture. Zero data retention and Confidential Compute on dedicated infrastructure are both “coming soon”, and pricing isn’t published, so regulated workloads should wait. Abstract your model client now and you can compare it with the others on your own evals when access opens.
2 · Models & features
ElevenLabs ships a realtime voice model you can swap in
Eleven Realtime comes in two sizes, Max and Mini, both speech-to-speech models for voice agents, with 90+ languages. ElevenLabs reports time to first audio of 0.57 s for Max and 0.62 s for Mini. On Artificial Analysis evaluations cited in its post, Max scores 98.8% on Big Bench Audio and 87.1% on tool use; Mini drops to 75.9% and 60.3%.
The API is WebSocket-only and follows the OpenAI Realtime format, so migrating can be close to an endpoint change. The gaps matter: no WebRTC, no published pricing, and the benchmark gap says Mini is for simple flows, not tool-heavy ones.
3 · Agents
Walmart opens an MCP gateway so agents can shop for customers
Walmart announced that outside agents will be able to shop on a customer’s behalf. Agent Gateway, built on MCP, starts as a beta for companies and developers with no date for general availability, and Agent Profiles are due “in the next month”. Named integrations are Muse, ChatGPT and Gemini, alongside Walmart’s own Sparky. Walmart’s terms say agents may not act beyond what the shopper intended.
That last rule is the interesting one. A retailer is writing the permission model for third-party agents before there’s a standard for it, and “shopper intent” isn’t something you can enforce in code. Expect every merchant API you depend on to grow something similar.
4 · Research
Train an agent across 18 harnesses, not one
A new preprint, Harness Scaling, splits a harness into five parts (system prompt, tool interface, observation, context management, workflow) with three variants each, giving 243 configurations. It trains on 18 of them, picked by an orthogonal design. On a Qwen3-30B-A3B model, the average score rises from 37.29 to 42.78 with GRPO, and a transfer test to Claude Code scores 41.20 against 38.08 for the best single-harness baseline.
The takeaway for builders: a model tuned to one harness may be quietly overfit to it. No code is linked, and the gains are modest on small models, so treat it as a hypothesis worth a cheap check on your own agent.
5 · People worth following
Claude Mods: plugins that sit inside the agent loop
Dan McAteer’s Latent Space piece describes mods, Claude Code plugins written as TypeScript or JavaScript functions that respond to events like a prompt, a tool call or a render, and can observe, re-render or handle the event themselves. He calls them “middleware for the agent harness”. Six ship built in, including a side agent called You Should Know that flags things you may have missed.
His cost note is the useful part: the Next Steps Supervisor mod is estimated at about 5-6k input-equivalent tokens per check, and it fired on 99 of 111 of his turns. The article also says mods aren’t sandboxed. Middleware is powerful, and every mod is code with the agent’s reach.
6 · Architecture
Vercel makes coding-agent sessions survive a crash
Vercel open-sourced durable-harness on 19 October. It wraps existing harnesses (Pi, OpenCode, Claude Code, Codex) so each session keeps a log of the conversation and every model and tool call, with intent written before a call and the outcome after. If a function dies, another reads the log and resumes. Finished calls are read back, not re-run or re-billed.
The design question is side effects. Tools tagged replay-safe re-run, “never” tools are reported to the model as possibly done, and a call that finished just before the crash may not be logged yet. It’s Apache-2.0 but on 0.0.x and marked experimental.
Also worth knowing
🆕 New on the radar
- MiniMax M3.1 Flash — API-only for now at $0.15 in and $0.60 out per million tokens, with weights promised “later”; the coding and agent benchmarks are MiniMax’s own. Watch. Source
- RecursiveSI Flint — Apache-2.0 coding MoE, 39.8B total and about 2.5B active, trained from scratch; its claimed SWE-bench Pro v2 score (89.7%) sits above its Verified score (74.2%), which is odd. Watch. Source
- IBM Granite Speech 5.0 2B — Apache-2.0 speech model for recognition and translation in six languages; none are Indian languages. Try, if English or European languages are your case. Source
- KiviDB — a Rust, Redis-compatible store with built-in vector search, AGPL-3.0 with a commercial option; its own benchmark shows 9× Redis 8 on pipelined GETs but slower index builds than Dragonfly. Watch. Source
- Mouse-1 — a small “decision model” for agent harnesses that returns a probability for routing-style choices, with no published pricing; claims of 13× to 190× cheaper than frontier models use mismatched baselines. Watch. Source
🏢 Platforms & enterprise
- Ramp Laminar — Ramp built one in-house agent harness for its agentic products after trying provider SDKs, managed agent APIs, frameworks and vendors; the post gives its lesson but no numbers. Source
- Apple on scaling agents with Ray — an InfoQ talk from two Apple engineers, recorded 20 October, covers autoscaling agent replicas when in-flight requests pass five and guardrails against runaway loops. Source
Deadlines & deprecations
- 2026-10-31 — DeepSeek starts billing server-side Search and Fetch calls at $0.75 per 1,000 calls each, from 16:00 UTC; until then they’re free inside DeepSeek Harness, Claude Code and Codex. If your agents use them, add a budget cap before the switch. Source
That’s it for today. If the harness is the part you haven’t looked at in a while, it can wait until the weekend.
Researched and drafted with AI, checked against the linked sources. How we use AI →