AI Learning -- Day 192

1) GPT-6.1 SOL: NEAR-FLAGSHIP AGENT CODING AT ONE-FIFTH THE PRICE

Search-based reports from OpenAI's DevDay (announced Tuesday, October 6) say OpenAI released GPT-6.1 Sol, which nearly matches its top model GPT-6 Astra on agentic coding, computer use and professional work. Reported API pricing is $2 per million input tokens and $10 per million output tokens, with cached input at $0.10 per million. On DeepSWE v1.1, a test built from real software bug fixes, it is reported to score 75.2% against Astra's 74.8%. It is reported as available in ChatGPT Work and Codex for paid tiers, and in the API as gpt-6.1-sol. These figures come from secondary coverage; confirm against OpenAI's own pricing page.

The concept, explained simply: a frontier model is like a senior consultant, expensive per hour. A 'distilled' or tuned mid-tier model is a strong associate who handles 90% of the work for a fraction of the fee. Vendors get there by training the cheaper model on the stronger model's outputs and by tuning it specifically for agent loops (read code, run tests, fix, repeat). Cached input matters because an agent re-sends the same long context on every step; a cache discount of 95% makes long agent sessions affordable.

Why it matters: it continues the tiering story from earlier days (Sol/Luna/Astra, Sonnet/Opus). The practical lesson is that benchmark parity at one-fifth the cost changes which model you default to. Treat a single benchmark with caution: test on your own tasks.

2) DEEPSEEK V4.1 FLASH: A HUGE MODEL THAT ONLY USES A SLICE OF ITSELF

Reports say DeepSeek launched V4.1 Flash in September, replacing earlier Flash variants and, after September 14, routing V4 Pro requests to Flash at Flash's lower price until a V4.1 Pro arrives. Coverage lists roughly 748 billion total parameters (a 552B backbone plus 196B 'Engram' parameters), only about 8B active per token during prefill and 16B during decoding, a 1 million token context, up to 384K output tokens, and native image input. A newer DeepSeek-V4.1-Flash release item was also listed on October 4 by one tracker; details differ between sources, so treat version specifics as unconfirmed. The concept, explained simply: in a 'mixture of experts' model, the network is a large building of specialist rooms, but each word only walks through a few of them. You pay compute for the rooms you visit, not the whole building. Engram-style parameters can be thought of as a big lookup memory that is consulted cheaply rather than computed every time. A smaller KV-cache (the model's working notes about the conversation so far) means more users fit on the same GPU, which is how prices fall. Why it matters: together with item 1, cost per task is falling from two directions, better tuning by US labs and leaner architecture from Chinese labs. Cheap long context makes whole-codebase and whole-document agents realistic.

3) MICROSOFT AUTOPILOT AND THE 'DIGITAL COWORKER' WITH CONFIGURABLE PERMISSIONS

Reports say Microsoft unveiled new Copilot capabilities, including a Code feature that builds apps from natural-language prompts and Autopilot, an updated version of its Scout agent, which acts as a digital coworker with configurable permissions. This follows OpenAI's always-on 'dots' agents (Day 190-191 coverage) and shows every major vendor shipping agents that run continuously inside Slack, Teams and Office.

The concept, explained simply: 'configurable permissions' means an admin decides which apps, files and actions the agent may touch, the same way a new employee gets access badges. This is the least-privilege principle from Day 191 turned into a product setting. Why it matters: adoption will be decided less by raw intelligence and more by whether IT teams can scope, audit and switch off an agent.