Field Notes

GPT-6.2 gets cheaper, Copilotforce in preview, and a clear lever for lower AI costs

Cheaper models, new Copilot hooks, and a simple cost trick you can use today.

Big AI news

OpenAI brings GPT-6.2 Astra and Sol to the API and ChatGPT Work

OpenAI launches GPT-6.2 Astra and Sol. Better pro outputs, coding, computer use, and higher efficiency, at lower cost in ChatGPT Work, Codex, and the API. Available in the API as gpt-6.2-astra and gpt-6.2-sol. An independent report puts Astra at 8/40 USD per MTok (input/output). That is 80% of the predecessor. Cache pricing is 0.40 USD per million tokens. Sol stays at 2/10 USD. Early user reports show mixed benchmarks and some unclear comms on Astra. You cut runtime costs and plan longer agent runs for less, especially with cache.12

Microsoft and Salesforce launch Copilotforce in preview

Microsoft and Salesforce announce Copilotforce. Salesforce context and actions show up directly in Copilot, Teams, and Fabric. Two-way zero-copy integration. So analysis across platforms without data copies. It allows queries across Data 360 and Fabric without duplicates. The “Salesforce in Copilot” plugin is in public preview with 38+ predefined sales skills. Accenture is a pilot customer. Access and actions stay within Salesforce permissions. Microsoft applies its compliance controls to comms and files. If your sales team lives in Microsoft 365, you pull Salesforce data into Copilot without app switching.3

Ignite 2026: decision model GA, security agents GA, and faster Azure VMs

Microsoft says Decision-1.1 is generally available. A fast, low-cost decision model for routing, classification, prioritization, and workflow control. Microsoft Perception for Security is now GA. Also Cobalt-200 VMs with up to 50% more performance than the prior generation and broad rollout by year end. Copilot Workspaces, Autopilot, and Copilot Code bundle teamwork and long-running agents. Salesforce in Copilot is a plugin in preview. If you are on Azure, you can plan Decision-1.1 and Perception for productive pilots now.4

AWS: rules before models cut cost and latency

AWS shows a decision path for when you even need a model. Check first: can you test the step as a rule? If yes, write it as code. If needed, use a model once to help with regex or schema. If the answer comes from a fixed list, use a bounded decision and only escalate uncertainty. Only open tasks go to a suitable model. That cuts cost, reduces latency, and creates auditable behavior. For example in support assistants with format checks, fast-path keywords, queue selection, and a generative answer only at the end.5

Highlights for your workday

  • Send sales artifacts via agent: deals with a sent agent artifact, a bundled result package for customers, moved to the next stage 47% of the time. Matched deals without one moved 16% of the time. 46% followed with a customer call within 14 days. 1,540 artifacts in 12 weeks. 337 hours saved. Figures are from the vendor. Create a sample artifact for your top 3 offers and test it for a week.6
  • Run long agent sessions cheaper: use cache pricing for reusable context, the lower rate for stored prompts. With Astra the cache drops from 1.00 to 0.40 USD per million tokens. This pays off for reports, code reviews, and research pipelines.21
  • 12 months of Snowflake CoCo free for Streamlit Community Cloud: apply by November 30, seats are limited. You get an AI coding partner that knows your data stack.7

Tools and updates

  • Replit DataCat: free observability alternative per the vendor, to watch apps and catch issues. The post claims feature parity with a leading platform. This claim is not independently verified. Useful if license costs hurt.8
  • Datasette Lite in the browser: load SQLite, CSV, JSON, or SQL by URL, install plugins, share demos without a server. Files need open CORS headers. Handy for data demos and bug repros.9
  • n8n 2.49.0/3.3.0: better support for OpenAI-compatible models and more stable agent runs, including clean tool call history. Helps if you use n8n for AI workflows.1011
  • Tonality Audio-Edit: change pitch, duration, pauses, energy, and pronunciation word by word. Free playground with limits, API available. For podcasts, support prompts, or learning content.12

Try it in five minutes

1) Pick a simple routing goal: three support queues like billing, tech, cancelation. Open your AI chat of choice and ask for fast-path keyword rules plus fallback, using these three sample messages: “I need a copy of my invoice”, “Login is not working”, “Please cancel”.5 2) Ask the AI to generate a regex list and a JSON schema for the decision once. You will then use them as fixed rules. Mark uncertain cases as “unsure”.5 3) Test with five of your own sample messages with no customer data. Check: correct queue, marked uncertainty, no unnecessary model generation. Note latency and whether a model call was skipped.5

Expected result: a small rule routing table plus an “unsure” signal that solves two to three cases per minute without an LLM. Acceptance check: correct queue assignment in at least four out of five tests and noticeably lower latency.5

Sources

  1. Introducing GPT-6.2 Astra and Sol (openai.com)
  2. GPT‑6.2 Astra and GPT‑6.2 Sol (simonwillison.net)
  3. Microsoft and Salesforce Announce Copilotforce, Bringing Trusted Customer Context and Action into Microsoft AI Experiences (news.microsoft.com)
  4. Microsoft Ignite 2026: Empowering human ambition, intelligence and trust (news.microsoft.com)
  5. Cut generative AI costs: Code for rules, models for judgment (aws.amazon.com)
  6. @langchain: Deals where a rep sent a GTM Agent artifact moved to a later stage 47% of the time. Matched deals without one: 16%.

✅ 1,540 artifacts in 12 weeks ✅ 46% followed by a customer call (x.com) 7. @streamlit: Community Cloud users can now apply for 12 months of @Snowflake's CoCo at no charge! 🤯

Write, debug, and ship apps faster with an AI coding partner that understands your data stac (x.com) 8. @replit: The leading observability platform wants $83,700/yr just to monitor your app and catch issues.

So Replit Missions built DataCat: https://t.co/vMAEEUYjf4

100% free.

This isn't a (x.com) 9. Running the full server-side Datasette Python app in the browser with service workers and Pyodide (simonwillison.net) 10. n8n@2.49.0 (github.com) 11. n8n@3.3.0 (github.com) 12. @datachaz: .@NadavKeyson and Riverside Labs just launched Tonality.

pretty mighty stuff:

you can finally edit any word in an existing recording: pitch, tone, duration, pauses, energy, and p (x.com)