The Bleeding Edge Weekly — W39: GPT-6 halves the price of a token, four frontier models ship in seven days

This edition combines the three newsletters we published separately this week (LLM Weekly, Devices & Robotics and Executive Roundup). Their text is unchanged.

The week in models

Four frontier models shipped in seven days, and the number that actually changes your roadmap isn't on a leaderboard — it's OpenAI cutting API pricing roughly 50%. Capability moved a step. The cost of an autonomous unit of work moved further than that.

OpenAI ships GPT-6 Sol and Luna, and cuts API pricing roughly 50%

Sol is the frontier reasoning tier, Luna the fast and cheap one, and they landed with published benchmarks, model docs and a refreshed prompt-caching guide. The pricing cut is the part that rewrites plans: every enterprise business case written in H1 2026 was costed against a per-token line that just halved. Marginal use cases that failed the ROI test six months ago now pass — which is how deployment volume steps up with no change in capability at all. Caching stacks on top of it, and that's a config decision, not an engineering project. MarkTechPost, OpenAI model docs, prompt caching guide

Three more frontier models land in the same seven days

Anthropic released Claude Opus 5.5, opening the 5.5 family, two days after GPT-6 — the fourth consecutive quarter the two labs have shipped flagships inside the same week. Google pushed Gemini 3.8 Live at the real-time voice tier, and Alibaba shipped Qwen 3.8 Omni, which is the detail worth keeping: Chinese open-weight releases now arrive in the same news cycle as US closed launches rather than a quarter behind. Opus 5.5 is trade-press-sourced in our sweep so far; worth checking anthropic.com directly. MarkTechPost, AI Search

Jev becomes the reaction tier of the agent stack

Jev's ultrafast browser demo did 31.4 million views and 66,000 likes on X inside 48 hours, but the interesting part is what builders did next: pair it with something slower and smarter. One developer ran a real-time Minecraft agent with Jev handling frame-rate reaction and GPT-6 Astra handling planning, fighting several mobs at once. That's the two-tier pattern — cheap fast perception, expensive slow planning — graduating from experiment to default. "Which model" is becoming "which model for which tier." Creators' AI playbook, Creators' AI weekly

Context engineering had its week: 49% less traffic, 92% less context

NVIDIA Research open-sourced SoL-Pi, which wraps a coding agent in an automated research loop that retrieves and caches what it needs instead of re-reading context — up to 49% less token traffic in reported benchmarks. Separately, a Claude Code compaction plugin compressed a 1M-token session to 86K in about a second, still usable. Caching, retrieval and compaction are three attacks on the same problem, and all three had a moment in seven days. MarkTechPost, NVlabs/SoL-Pi, Creators' AI

A browser agent booked a flight for $0.0039

Browser Use founder Gregor Zunic demonstrated an agent completing an end-to-end flight booking in seven seconds for under half a cent of inference. It's a demo, not a product. It's also the first credible public cost-per-transaction figure for agentic booking, and it's the number every intermediary layer now has to argue against. Creators' AI

Stack the week up: a 50% price cut, a 49% traffic reduction, a 92% compaction. Compounded, the cost of an autonomous coding task fell roughly an order of magnitude in seven days, and not one leaderboard measured it. Watch for the first serious eval that reports cost-per-completed-task instead of accuracy — that's the benchmark this week made necessary.

Devices & robotics

Four frontier models landed in seven days and not one of them was aimed at a device. The physical-world story this week isn't hardware at all — it's that the two-tier agent stack stopped being a whiteboard diagram and started fighting zombies in real time.

The reflex-and-planner split gets its first convincing demo

Developer Wuyang Zhou built a real-time Minecraft agent that splits the workload across two models: Jev handles frame-rate reaction, GPT-6 Astra handles planning. It fought multiple zombies simultaneously — which is the part that matters, because holding a plan while reacting at frame rate is exactly what breaks single-model embodied systems. Minecraft is a toy. The architecture is not: it's the same cheap-reflex-plus-expensive-planner pattern warehouse picking, inspection drones, and anything else with a real-time control loop will end up using.

Via The Creators' AI.

Jev is the reaction tier, and it went supernova

Jev launched with a speed demo that pulled 31.4 million views and 66,000 likes on X inside 48 hours. Strip out the virality and what's left is a model optimised for latency rather than reasoning depth — the missing cheap layer in every embodied stack that currently pays frontier prices for a decision that needed to happen in 40 milliseconds. Builders paired it with slower planners within days of launch, which tells you the gap was real.

Via The Creators' AI's agent-speed playbook and AI Search.

Google ships Gemini 3.8 Live at the voice tier

Google released Gemini 3.8 Live, aimed at more capable real-time voice conversation. Reporting is thin and we haven't matched it to a Google primary source yet, so hold it loosely. But voice latency is the entire product for smart glasses, earbuds, and the dedicated-button category — a dedicated real-time voice tier is the upstream dependency those devices have been blocked on.

Reported by AI Search; single-source.

Qwen 3.8 Omni extends the open multimodal line

Alibaba shipped Qwen 3.8 Omni, with Dream RSI and Bonsai 2 landing alongside it. Detail is scarce on all three. The structural point for hardware people: open multimodal weights are the only route to running perception locally without a per-inference bill or a round trip to someone's datacentre, and these are now arriving in the same news cycle as the US closed-model launches rather than a quarter behind them.

Reported by AI Search; detail unconfirmed.

SpeakON puts a microphone on a MagSafe puck

A magnetically-attached hardware button with its own onboard mic, built to trigger voice AI without unlocking the phone. Cheap dedicated AI input hardware keeps producing entrants despite the Humane and Rabbit graveyard, and the pitch has quietly improved: this one isn't trying to replace your phone, just to shave the three seconds before you can talk to it. Attaching to the device instead of competing with it is the only version of this category that has ever made sense.

Via the MarkTechPost newsletter.

What to watch

A week with no humanoid, no autonomous-vehicle milestone, and no new edge accelerator is worth naming as a gap rather than papering over. The thing to watch is whether the reflex-tier model shows up in something with actuators. A two-model stack that works at frame rate in a game engine is one physics gap away from working on a factory floor — and the first vendor to ship it with a robot arm attached sets the reference architecture for everyone else.

What it means for leaders

Four frontier models in seven days, API pricing cut roughly 50%, and the first-ever frontier-lab briefing to the UN Security Council — all in the same week Washington said it wants AI policy left exactly where it is. The cross-role theme: the cost of an autonomous unit of work fell an order of magnitude, and no benchmark, ruling or press release measured it.

If you're a CEO this week...

Every AI business case your team wrote in H1 was built on a cost-per-token line that just halved. The marginal use cases that failed your ROI threshold six months ago now clear it — which means your competitors' deployment volume steps up without any of them getting smarter. That's the competitive-position story, and it's the one your CFO should be re-running before the next board pack.

On regulatory risk: governance moved venue, not teeth. The Security Council briefing is the highest-status AI governance moment to date and produced no binding obligation, while Beijing's line ahead of the Xi talks was shared responsibility and mutual loss from confrontation. Assume no external rule forces your hand in the next 18 months — your timing is your own decision.

Watch the capital side too: a16z is building its own university and General Catalyst is restructuring toward an operating-holding model. Smart money has concluded that owning the operation beats funding it.

The board question: if our AI cost base halved this week, what did we defer last quarter on cost grounds that we should now be shipping?

If you're a CIO/CTO this week...

Standardising on a model is no longer a durable procurement decision. GPT-6 Sol and Luna, Claude Opus 5.5, Gemini 3.8 Live and Qwen 3.8 Omni all shipped inside seven days — the fourth consecutive quarter the two Western leaders have landed flagships within 48 hours of each other. Your defensible investment is the routing and abstraction layer, not the endpoint.

Context management is where the actual money is. NVIDIA's SoL-Pi cuts coding-agent token traffic by up to 49%, a Claude Code compaction plugin squeezed 1M tokens to 86K in about a second, and OpenAI's refreshed prompt-caching guide is a config change, not a project. Stack those with the price cut and agentic coding costs roughly a quarter of what it did last month.

Two security notes: Vercel now runs a model as the production safety reviewer gating fx auto mode — model-reviews-model as a real control, worth stealing. And OpenAI was reportedly hacked: single-source, unconfirmed, no primary statement. Don't act on it; do ask your vendor-risk lead to watch for disclosure.

The read: buy the models, build the routing layer, and put caching and compaction on this sprint — it's the highest-return work available to you right now.

If you lead AI transformation this week...

Your pilot for the next two weeks is a two-tier agent stack. The pattern went mainstream this week: Jev pulled 31.4M views in 48 hours as a reaction-tier model, and builders immediately paired it with GPT-6 Astra as the planner — a cheap fast model perceiving and acting, an expensive slow one thinking. Point it at one workflow with a clear latency constraint. Ops and procurement have the sharpest case: Browser Use booked a flight in 7 seconds for $0.0039. That's your first real cost-per-transaction number for agentic work — use it to reset what "too expensive to automate" means internally.

On change management, read Lenny Rachitsky's account of working inside an AI-native company. It's the only thing this week that describes the destination state concretely: far smaller teams holding far larger scope, and managers spending their time on verification rather than assignment. Most transformation decks stop at tooling and never touch the org chart.

Governance: Amodei called publicly for labs to slow down and was rebutted on the record by Altman and Musk, while Huang put existential risk at zero. Your framework can't inherit a consensus that doesn't exist.

The experiment to run this month: adopt Vercel's pattern — put an automated reviewer in front of one agent workflow already running in production, and measure what it catches.

All three roles are looking at the same gap: capability announcements are public and loud, while the economics that actually change your plan are buried in a caching guide and a GitHub repo. The question worth asking in every function this week is the same — if autonomous work just got ten times cheaper, what are we still doing by hand because it used to be expensive?

---

This post is also published on our Substack newsletter at edge-ai.forum. Subscribe for the weekly roundup direct to your inbox — fresh AI news, executive context, and devices + robotics every Friday morning.