Tech • AI • Robotics • Game

ENFR
TodayPlay for freeFury BloxLeaderboardShortsTop StoriesFor youTopicsVideosYT channelsArchivesSearchFavorites

OpenAI launches GPT-6.1 Terra as Google debuts Gemini 4 Carbon

AIWednesday, October 28, 2026· 26 videos

Briefing

Audio player

OpenAI prices GPT-6.1 Terra aggressively

OpenAI formally launched GPT-6.1 Terra on October 26, turning earlier code sightings into a full release across the OpenAI API, ChatGPT Work and Codex. The model is priced at $1 per million input tokens and $5 per million output tokens, about half the rate of GPT-6.1 Soul. OpenAI is positioning Terra as a fast, lower-cost workhorse for agentic and computer-use tasks rather than a pure flagship intelligence play. The move sharpens competitive pressure on Anthropic, Google and open-model vendors as inference price becomes a central battleground.

Terra nears Astra at lower cost

On agent-style benchmarks, GPT-6.1 Terra is presented as approaching GPT-6 Astra-level results at roughly one-tenth the price per token. OpenAI says it runs more than three times faster than Astra and 6.1 Soul on professional agent tasks, while Terminal Bench 4.0 reportedly rose from 21.5% to 50.3%. A cited hallucination drop from 88% to 38% suggests material reliability gains, though the strongest improvements appear concentrated in practical automation rather than deep reasoning. Editorially, the significance is that cheaper models are getting good enough for always-on workflows once reserved for premium systems.

Terra shines in computer use

Early usage signals point to GPT-6.1 Terra being strongest in PDF extraction, general research and low-cost software control. It appears notably weaker than ChatGPT 6.1 Sol in complex reasoning, production-grade coding, exact invoice automation and high-fidelity 3D generation. That tradeoff makes Terra attractive for repetitive business processes where speed and token cost outweigh perfect precision. For legal, economic or other error-sensitive domains, higher-end models still appear safer despite the added expense.

Google unveils Gemini 4 Carbon

Google introduced Gemini 4 Carbon, a new flagship positioned above Gemini 4 Argon only two weeks after Argon’s release. The company is pitching Carbon as a model for long-horizon work, with internal tests claiming it can persist for days on difficult engineering assignments. Cybersecurity is the headline use case, with the model designed to find vulnerabilities, validate them, exploit them in controlled settings and assist with patching. On Google’s shared results, Carbon reportedly outperformed GBT6 Astra and Claude Opus 5.5, though access remains limited.

Google leak hints at Hydrogen

Separate signs from Google backend logs pointed to a model slug called Hydrogen, widely interpreted as a possible Gemini 4 Flash variant. Early reports described speeds near 300 tokens per second, implying a strong low-latency focus rather than a frontier-capability push. Google also expanded its interactive world-model effort with Genie 4, which generates explorable environments with real-time prompting, native audio and character interaction on Gemini Omni. Taken together, the updates suggest Google is pressing simultaneously on fast inference, interactive simulation and top-tier reasoning.

PHP flaws expose AI security shift

Two critical PHP vulnerabilities, rated about 9.2 out of 10, triggered outsized alarm because they affected PHP 5.4 and later and sat in the runtime itself. Researchers described the bugs as pre-authentication memory-corruption flaws, with PHP2Shell reports linking them to remote code execution from a single request in some cases. The larger story is that AI-assisted variant analysis is accelerating discovery of old bug families across widely deployed software such as WordPress and Drupal stacks. That compresses the timeline between discovery and weaponization, making patch lag the core defensive weakness.

Anthropic economics point to demand boom

Fresh investor commentary around Anthropic painted stronger-than-expected unit economics and lower cash burn than many had assumed. Annualized revenue was cited near $6.7 billion, with roughly 250 customers spending more than $10 million a year and about 11,000 to 12,000 above $100,000. At the same time, AI still accounts for only about 0.1% of customer spend in many large enterprises, implying significant room for expansion. Demand for compute remains so tight that GPU capacity extending into 2027 is still commanding a premium.

OpenAI targets robots by 2027

Sam Altman said OpenAI will "definitely" build a humanoid robot and aims for a public demo in 2027. The company is pursuing both humanoids and more specialized machines for data centers and industrial settings, with home use appearing farther off. OpenAI’s revived robotics group, formally launched on May 31, is built around World Sim and led by Aditya Ramesh of DALL-E and Sora fame. The strategic shift is notable because it moves OpenAI from supplying core models to competing directly in robotics hardware and embodied AI.

Videos covered

Previous briefings · AI