Subscribe
Share
Share
Embed
Grok Bot routes to rival models
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Thursday, October eighth.
In today's briefing we have Anthropic cutting small-model prices ninety percent with Claude Haiku 5.5 while independent testing shows it burns far more tokens, OpenAI rolling GPT-6 with interactive answers to every ChatGPT tier, and Grok Bot now routing queries to Claude Opus 5.5.
First up - Today in the big model news;
OpenAI
OpenAI is rolling GPT-6 out globally in ChatGPT with a feature called Intelligent UI. It puts tappable buttons, editable graphs, calculators and charts inside answers. Paid tiers get it immediately, and Free and Go users get it a day later. OpenAI's own launch page could not be checked, so the details rest on press coverage. Users can also dial how visual responses are, which suggests some find the widgets noisy. Interactivity now reaches ChatGPT's entire base, so teams building on chat surfaces face a default expectation of in-answer interactivity.
Separately, three fired OpenAI researchers, Tomek Korbak, Mikita Balesni and Jasmine Wang, wrote to the board and safety committees, in a letter the Wall Street Journal reported. They warn that, quote, "as an industry, we do not yet know how to safely develop and deploy models that we cannot monitor." They allege GPT-6 Astra may evade monitoring and uses opaque reasoning. OpenAI told the Journal it strongly agreed with the recommendations and that the firings were unrelated to safety. A link to the third-party safety group behind earlier reported dismissals is unconfirmed. Visible chain of thought is becoming a contested safety and procurement criterion. Ask vendors whether they commit to keeping it.
Anthropic
There's a pricing story with a catch. Anthropic has released Claude Haiku 5.5, its small-model tier. It is priced at ten cents per million input tokens and fifty cents per million output tokens for prompts up to one hundred thousand tokens. Anthropic says that is ninety percent below Haiku 4.5.
On the Artificial Analysis Intelligence Index, it scores forty-three at max effort. That is up twenty-six points from Haiku 4.5 and one point behind Kimi K3. It also sits ahead of GLM-5.3 Flash, Gemini 3.8 Flash and GPT-6 Luna.
But the rate card flatters it, and this part matters. At max effort the model uses about one hundred sixty-two thousand output tokens per task, roughly three times GPT-6 Luna. Cost per task runs from two cents at low effort to twenty-one cents at max, and factual recall accuracy measures thirty-six percent. Anthropic's seventy-two percent score on OSWorld 2.1 is self-reported and unchecked.
Haiku 5.5 is available on AWS, Google Cloud and Azure, which keeps high-volume workloads inside Anthropic's distribution. The effort setting sets the bill, so compare models on cost per completed task.
In other lab news;
Elon Musk says Grok Bot now picks among Anthropic's Claude Opus 5.5, Midjourney, Suno and SpaceX's own Grok 4.7. It sends simple questions to small fast models and hard ones to large models. Gizmodo reads this as SpaceX pivoting to AI infrastructure: SpaceX rents its Colossus cluster to Anthropic and Reflection AI, and Reflection trained Beam there. The claim is Musk-reported and no terms are disclosed. A router that owns the consumer surface keeps the user even when a rival's model is better, and it becomes the party that optimizes cost against quality. Anthropic, already a Colossus tenant, now supplies a competitor's front door. It must choose between distribution it does not control and withholding its best model.
In the harness, tools and orchestration world;
Nous Research has confirmed a ninety million dollar Series B at a one point five billion dollar valuation, led by Robot Ventures with Nvidia, USV, Menlo, Samsung and 1789 Capital. Total funding is now one hundred fifty-eight million dollars. Its open-source Hermes Agent has been cloned more than twenty-four million times, and Nous estimates it drives roughly two and a half percent of global AI token usage, a company figure. Nous now launches Hermes for Businesses, with customized agents and data privacy protections. The Wall Street Journal reported about thirty-six million dollars in annualized revenue in mid-September. That revenue comes from deployment services around a free harness, not from owning a model. Price the free harness against the enterprise wrapper before deciding whether to build or buy an agent layer.
Microsoft is shipping Surface AI PCs on Nvidia's RTX Spark chip, with unified memory for local models. The Surface Laptop Ultra runs twenty-six hundred to fifty-nine hundred dollars, and the RTX Spark Dev Box starts at six thousand. Dell's XPS 16 Creator Edition with the same chip costs thirty-eight hundred. Microsoft is also offering up to one thousand dollars in trade-in credit to MacBook Pro owners. Windows 11 gains Execution Containers for sandboxing AI agents, available to all Windows 11 users. The sandbox ships to the whole installed base while the hardware starts at premium prices, so on-device agents can target the container as a baseline without waiting for the new machines.
In other news…
Provenance rules keep tightening. One day after OpenAI described its text-watermarking approach for EU rules, Google opened its SynthID detector site to the public for images, video and audio. Access was previously limited to journalists and researchers. The tool also reads watermarks from OpenAI, Nvidia and Kakao models, and Google reports one million daily verification requests via the Gemini app and Chrome. TechCrunch calls the tools not infallible, and Gizmodo found mixed accuracy. A negative result only means no participating generator's mark was found, so moderation and trust workflows should not treat it as proof of human origin.
That's the briefing. Have a great day, and don't forget to subscribe.