Get tomorrow's brief in your inbox
Today: OpenAI launched GPT-6 Sol and Luna at half the price of their predecessors, kicking off a real price war with Anthropic's new Claude Opus 5.5. Zapier's survey of 500+ SMB owners shows 76% now use automation but most lack governance guardrails. Meta patched a zero-day in its Muse AI agent that let attackers hijack the tool through undocumented settings.
The AI price war just got serious, and your API bills are about to drop
OpenAI released GPT-6 Sol and GPT-6 Luna, cutting API prices by 50% compared to their GPT-5.6 equivalents. GPT-6 Luna now costs $0.10/$0.50 per million tokens, making it one of the cheapest capable models ever offered. Hours earlier, Anthropic dropped Claude Opus 5.5 with a 20% price cut and a 60% reduction in cache-read costs. (OpenAI announcement, Simon Willison's analysis)
If you're building anything on top of AI APIs, whether that's customer support bots, content generation, or internal tools, this matters directly. GPT-6 Sol outperforms Claude Opus 5 on business workflow benchmarks at roughly 9% of the cost per task. GPT-6 Luna improved on its predecessor while costing 58% less.
What to do: If you're paying for API access through any provider, review your model selection this week. GPT-6 Luna at $0.10/$0.50 is now cheap enough to run high-volume automation that would have been cost-prohibitive six months ago. If you're on Claude, the cache-read discount on Opus 5.5 makes long agentic sessions significantly cheaper. The models are live now across both platforms.
Adobe Premiere mobile lands on Android (free) Adobe released Premiere mobile for Android with frame-accurate editing, multi-track timeline, 4K export, and Firefly-powered generative AI features (fill, image-to-video, sound effects). No login required, no watermarks, no ads. The free tier covers core editing; generative AI credits cost extra at higher volumes. If you're creating product demos, social clips, or tutorial videos on your phone, this eliminates the need to transfer footage to a desktop. (Source)
Rabbit OS3: AI agent that runs on any computer Rabbit ditched the hardware-only approach and released OS3, a cloud-based AI agent that operates locally across Windows, Mac, and Linux. You can connect up to five devices, link your preferred AI models, and access it through Telegram or iMessage. Rabbit claims it won't store, copy, or sell your data, though chat logs and "memories" do stay on their servers. Worth watching, but give it a few weeks before trusting it with anything sensitive. (Source)
OpenAI ships better prompt caching for GPT-6 New caching tools include a dashboard to monitor cache hit rates, a diagnostics tool to figure out why cache misses happen, and explicit cache breakpoints for developers. You can now change reasoning effort between responses without breaking cache. If you run persistent AI agents that make sequential API calls, this can cut costs significantly. The practical upside: less latency, lower bills for agentic workflows. (Source)
LLM plugin ecosystem keeps pace Simon Willison's LLM tool updated to support GPT-6 Sol, GPT-6 Luna, and Claude Opus 5.5 on launch day. A new plugin for TypeSafe AI's Jev model adds structured yes/no, choice, and scoring queries, useful for automated triage and classification tasks without building custom logic. (LLM 0.36, llm-typesafe)
76% of SMBs are automating, but most are flying blind on governance
Zapier surveyed 500+ SMB owners and found that while 76% use some form of automation, and 30% have AI making decisions across multiple systems with limited human oversight, very few have governance frameworks in place. Among automation users, 94% report real benefits: fewer errors (44%), cost savings (44%), and time freed for higher-value work (43%). (Source)
The gap worth paying attention to: 55% of businesses still have manual tasks they know they should automate but haven't. The top blockers are not knowing which tools to use (43%), unclear ROI (34%), and perceived technical complexity (30%). Meanwhile, 35% had someone work overtime to keep a manual process running, and 27% had a manual error that cost them money.
Action steps: - Pick your single most repetitive weekly task and automate it this week using a no-code tool - If you have AI making decisions without human review, document what it can and cannot do; that's your minimum governance framework - Start with rule-based automation before jumping to AI-powered decisions; the ROI is clearer and the risk is lower
AI in HR is getting practical For businesses with even a small team, Zapier's breakdown of seven real HR automation use cases is worth reading: automated resume screening, AI-drafted job descriptions, benefits FAQ bots, onboarding workflows, and survey analysis. The key insight is that most of these combine traditional automation triggers with a single AI step in the middle, not full AI agents. (Source)
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Change |
|---|---|---|---|
| GPT-6 Sol | $2.00 | $8.00 | -50% vs GPT-5.6 Sol |
| GPT-6 Luna | $0.10 | $0.50 | -50% vs GPT-5.6 Luna |
| Claude Opus 5.5 | $4.00 | $20.00 | -20% vs Opus 5.0 |
| Claude Opus 5.5 (cached) | - | - | -60% on cache reads |
GPT-6 Luna at $0.10/$0.50 is now cheap enough for bulk classification, tagging, and content generation tasks that used to require careful token budgeting. For comparison, GPT-6 Astra and Claude Fable 5.1 both sit at $10/$50, ten tiers above Luna. If you're on the free tier of ChatGPT or Claude, these price drops won't affect you directly, but they signal that paid API access is becoming practical for solo operators building custom tools.
Meta patches Muse zero-day, but the security story is bigger than one bug Security researcher Patrick Wardle found that any local app could redirect Muse's transcription to a third-party server, take pictures, and write files to disk, all without alerting the user. Meta patched it within hours but downplayed the risk since it required local access. The lesson for businesses: AI agents that can see your screen, access your files, and take actions on your behalf need the same security scrutiny as any other privileged software. If you're evaluating Muse or similar tools, wait for the security posture to mature. (Source)
Qualcomm's new Snapdragon chips push AI onto phones The Snapdragon 8 Elite Gen 6 can run 30-billion-parameter models on-device. Qualcomm and Google are positioning this as the shift from "phone-centric" to "agent-centric" mobile experiences. For businesses, this means AI features in your mobile apps will increasingly run locally rather than hitting cloud APIs, reducing latency and costs. Not actionable today, but worth factoring into 2027 mobile strategy. (Source)
Snorkel AI triples to $3.5B valuation The AI training data company hit $375M annualized revenue, an 18x increase in 12 months. This signals that the bottleneck in AI development has firmly shifted from model architecture to training data quality. Not directly actionable for SMBs, but it explains why every AI tool you use is about to get noticeably better. (Source)
Audit your AI agent permissions this week. The Muse vulnerability is a reminder: check what your AI tools can access. If an AI assistant has access to your screen, files, or accounts, verify it only connects to the vendor's servers and review what permissions you've granted. Five minutes of checking beats one data leak.
Test GPT-6 Luna for your bulk tasks. At $0.10 per million input tokens, try routing your highest-volume, lowest-stakes AI tasks (email categorization, content tagging, first-draft generation) to Luna. If the quality holds for your use case, you could cut that line item by 50-90%.
The price war between OpenAI and Anthropic is compressing faster than anyone expected, and the mid-tier is where most businesses actually operate. AI agent security is emerging as a real concern now that tools like Muse are shipping to millions of users before their security models are battle-tested. If you're budgeting for AI tools in Q4, the cost curve is moving in your favor; lock in usage patterns now while pricing is aggressive.