OpenAI’s GPT-6.1 Terra Cuts Costs, Google Runs AI in Orbit, Anthropic Files for IPO

FEATURED INSIGHT💡

GPT-6.1 Terra halves token prices for computer-use work

OpenAI's GPT-6.1 Terra gives developers a lower-priced option for work that combines reading images, using software and writing. Its standard API token prices are half those of GPT-5.6 Terra and GPT-6.1 Sol: $1 per million input tokens and $5 per million output tokens. The company points to jobs such as interpreting technical drawings and entering invoice details into accounting software. For recurring work, that changes the price at which an automation might become useful.

OpenAI reports that Terra scored 15 percentage points above GPT-5.6 Terra on Agents' Last Exam at medium reasoning effort, at about one-third the estimated task cost. That benchmark includes professional tasks such as converting architectural drawings into 3D models. These are the company's evaluation results, and a lower token price does not guarantee the same reduction in a particular workflow's bill. Failed attempts and human review still count toward the cost of getting a job done.

The API documentation lists a 1.05-million-token context window and requires the Responses API for tool calling; requests above 272,000 input tokens carry higher rates. Teams evaluating Terra should compare completed work on the same inputs, including corrections, rather than choosing by token prices alone. An invoice pilot, for example, should check the saved accounting record against the original document before counting the task as successful.

Stop losing deals in between meetings.

Aligned is changing how B2B teams run their most important deals.

Most reps guess where a deal really stands. They forecast on a gut feeling and the buyer’s word vs what's actually happening. Then they realize the deal slipped through only after it's too late. Aligned surfaces what buyers are actually doing inside every deal: the risks, the openings, and the exact next step to keep things moving.

Start increasing your close rates and stand up your first deal room in minutes.

Start Now For Free

ON THE HORIZON 🌅

Google's AI chip answers from orbit

Google says its Project Suncatcher prototype produced an AI response in orbit, running the Gemma 4 open model on a data-center-class tensor processing unit, or TPU. During a ground-station pass, engineers asked for a haiku about space and received the output as the satellite passed Svalbard, Norway. Google's October 30 report describes follow-up tests with limericks and prompts about the payload's technical status.

The prototype launched on October 1 through a partnership with Planet. Google describes the project as research into whether space could eventually support scalable machine-learning infrastructure. The new result tests a concrete part of that proposal: operating an AI chip, running a model and returning its output while the hardware is in orbit. It gives the team working equipment to study under conditions that a ground-based simulation cannot fully reproduce.

Google's next tests will examine TPU performance under thermal stress and radiation exposure, ahead of a planned pair of interconnected satellites in 2027. Those are substantial remaining engineering questions. A successful model response establishes an early operating milestone; deciding whether orbital computing can become useful infrastructure will require evidence about sustained workloads and reliability. For teams tracking future compute supply, the development is worth following as a hardware experiment with further tests ahead.

LATEST IMPORTANT NEWS 📰

Anthropic files for a Nasdaq debut

Anthropic filed its public S-1 registration statement on October 26 for a proposed initial public offering, with plans to list on Nasdaq under the symbol ANT. The company said the share count and offering price range had not been set, and the registration had not yet become effective. That moves the Claude developer's listing plans into a public filing process, while leaving the size and terms unresolved. The announcement establishes a proposed offering, rather than a completed stock-market debut.

Mistral Large 4 opens its weights, with a licensing catch

Mistral released the improved Large 4 model's weights, giving organizations a way to download, customize and deploy its trillion-parameter multimodal model on their own infrastructure. The company reports further coding gains after additional reinforcement learning since its earlier preview. Its modified MIT license permits commercial use, modification and fine-tuning, but businesses operating a model-as-a-service offering need a separate commercial license. Self-hosting expands deployment choices; the specific license condition matters for developers planning to sell model access.

Google and Microsoft back one AI training opt-out

An October 30 Internet-Draft proposes a shared robots.txt token, __usage_ai-training, for publishers to express training preferences across providers. Microsoft plans to begin honoring its new preferences as pages are crawled starting February 1, 2027, while Google targets early 2027. The draft remains an individual proposal with no formal IETF standing, and the signal expresses a usage preference rather than enforcing access control. Publishers should retain their existing crawler rules while checking each provider's rollout.

The AEO work, without the AEO hire.

Most AI visibility tools show you where you rank, then leave the work to you.

ChatFeatured is an AI search agent that does it: finds the questions competitors win, drafts the content in your voice, and tracks the lift. 

Your team decides what ships. Month to month.

Try the agent, today.

FOR THE TECHNICALLY INCLINED 🛠️

Pika cuts video generation from 50 steps to six

Pika H3 Flex, released October 29, post-trains MiniMax H3 to generate synchronized video and audio in six sampling steps instead of 50. Pika reports that a 15-second, 768p clip can take seven seconds of transformer computation in a controlled benchmark; that figure excludes the rest of the serving pipeline and is not end-to-end API latency. The distinction matters when estimating how quickly a production tool will return a usable clip. The released model package supports native 480p and 768p output, mixed stereo audio and up to 12 reference inputs. Higher-resolution upscaling and separate audio stems are hosted API features. Developers can test the local generation path against their hardware and creative requirements, while treating those finishing services as separate capabilities. Pika says the model's consumer-product rollout is planned for November 2.

AI TOOL OF THE DAY 🚀

Napkin AI turns pasted text into editable diagrams, flowcharts and other visuals that you can customize and export for presentations, documents or posts.

Your Search Bar Is Haunted

Shoppers type exactly what they want and get “0 results found” back. They don’t complain. They just vanish to a competitor.

Gladly understands what shoppers mean, typos and all, and turns dead-end searches into product recommendations.

See how Gladly brings search back to life

Turn Dead Ends Into Sales