💸 Musk's New Grok Beats GPT-5.6 Sol for a FIFTH the Price

Grok 4.6 finished a long job in 53 steps. Claude Opus 5 took about 103 to answer the same thing, and charges five times more per million output tokens to do it. SpaceXAI put the model out at the price it set for older model. Running an agent all afternoon costs less than it did in July. Let's dive in!

In today’s insights:

LATEST DEVELOPMENTS

Evolving AI: Grok 4.6 launched at the frontier this week and kept Grok 4.5's $2 and $6 pricing.

Key Points:

Artificial Analysis ran the model through its own suite and logged a cost of $0.84 per completed task. SpaceXAI has gained five points on that index in one month and 23 points since Grok 4.3. Its 500,000 token window only bills at the headline rate below 200,000 tokens, and a longer prompt doubles the charge on the whole request. Cached input climbed to 50 cents per million tokens, up from the 30 cents charged on the previous model. On long horizon work the run clocked Grok 4.6 at about 53 turns and half a billion input tokens per task. Claude Opus 5 took around 103 turns and two billion input tokens to reach a higher scoring answer.

Try Grok 4.6 in Cursor

Claude Opus 5 still gives the better answer, and that is the gap anyone picking a model has to price in. Cheap output and short runs let an agent work a full afternoon on one job before the cost gets uncomfortable, and that raises how much a single person can hand over in a day. What a model costs to run now sets the limit on how much work one person can push through it in a day.

Most CX platforms don't own the voice. They orchestrate a workflow, then call a third party for speech and transcription. Every hop adds latency, and latency is what turns a frustrated customer into a churned one.

ElevenAgents is the opposite. Built on the voice models the market already builds on, it runs voice, transcription, chat, and reasoning in one vertically integrated pipeline. Responses come back in under 400 milliseconds and sound human, not synthetic. When a caller gets frustrated, the agent detects it and shifts tone in real time: calm, reassuring, patient.

You keep full control. Plug in any LLM, connect tools, webhooks, and MCP servers, and ground every answer in your knowledge base. Launch in minutes, A/B test with Experiments, enforce Guardrails, and version every change.

More resolved conversations, less infrastructure stitching. Pricing is transparent and flat at $0.08 per minute.

Start building free

Evolving AI: Apple is pitching newsrooms a payment model the rest of the AI industry has avoided.

Siri AI runs on Google's Gemini and answers questions using live information pulled from the web. Apple declined to comment on the terms reported by The Wall Street Journal, and it has signed earlier content deals that covered rights to train its own models. Its Apple News+ subscription has paid participating outlets since 2019, so the company already runs a publisher payment system. Apple Intelligence summaries invented BBC and New York Times headlines on lock screens, and Apple switched that feature off in January 2025 before restoring it in iOS 26.

Google's Gemini will produce the answer your phone reads back, and a newsroom earns from it only when its reporting is what the assistant used. That gives the money a reason to follow verified work, which reaches anyone who stopped trusting an AI summary of the day's events. Apple is attaching a price to the sourcing itself, and that bill grows every single time somebody asks.

Ever caught an AI stating something totally wrong with total confidence? This video from Anthropic breaks down the two real culprits, hallucination and sycophancy, and reframes trust as a dial you adjust per question, not a switch you flip once. It explains why confident-sounding answers aren’t necessarily correct, and why uncertainty, verification, and transparency become even more important as AI systems start making decisions and taking actions on our behalf.

Evolving AI: Google’s SL2T Model reads Sign Language and writes it out in English as the signing happens.

Google DeepMind built SL2T to read the hands, torso, head and face at once, because meaning in sign languages comes from all of them at the same time. On the FLEURS-ASL benchmark it scored 70 BLEURT for translation quality, and no earlier model has come close to that. An on-device model tracks points on the signer's body and sends only those coordinates to Google's servers, so the phone can delete the recording straight away. Google also tuned it for the one in ten signers who are left-handed and for people who sign with one hand while the other holds a phone.

Google’s Live Transcribe uses the same model, so a Deaf commuter can sign one-handed on a train and the stranger opposite reads the words as they arrive. Hearing users have had voice dictation for well over a decade, and signers spent those same years typing in a language that was not their own. Google is pointing AI at a signed language for the first time, and the people who use it no longer have to work in English to be understood.

QUICK HITS

🚪 Brad Lightcap, OpenAI's longest-serving exec, is leaving to "start something new" as the IPO nears.

🌊 xAI co-founder Igor Babuschkin's River AI raised $1.1B two months out of stealth, backed by Nvidia, AMD, and Temasek.

📈 Google's Gemini app crossed 1 billion monthly users, its fastest-growing product ever, with 63% now using voice.

🐧 OpenAI shipped a ChatGPT desktop app for Linux, completing its desktop lineup.

 *partner link