SI Agent News

• Inline Tangents for ChatGPT (STILL OPEN): a user on the OpenAI community forum proposes letting you branch a side question inside a ChatGPT conversation without derailing the main thread. It is a community proposal only, with no company response and no detail in the post. Source: OpenAI community forum

Microsoft Copilot reportedly adds OpenAI's GPT 6.1 and Anthropic's Claude Opus 5.5 for Premium users (REPORTED)

What happened: Geeky Gadgets reports Microsoft Copilot is adding eight new features this October, headlined by GPT 6.1 (an OpenAI model) and Claude 5.5 (an Anthropic model) for Premium subscribers. Source: Geeky Gadgets

Check before you rely on it: the report calls the Anthropic model both 'Claude 5.5' and 'Claude Opus 5.5', and doesn't describe the other seven features. Single secondary source. Price: not stated. Confirm on Microsoft's pages.

What changed: not a new model, but one subscription tier reaching two rival vendors' flagships. STILL OPEN: how this affects people paying OpenAI directly. Report doesn't say.

Worth it for: Microsoft-tool users comparing vendors without a second subscription.

What this means for you:

• Small business owner: if you pay for Copilot Premium, test one real task, like a customer email or quote, on both before adding another subscription.
• Job seeker or student: run one resume bullet or essay outline through both; keep whichever sounds more like you.
• Creator: run your script hook through both; pick by voice, not brand.

You're with SI Agent News, OpenAI edition. Ahead: coding agents, a cost cutter, and a prompt fixed line by line. Behind this show: a daily scan of a large number of articles.

Cursor, Claude Code or Codex? What the 2026 coding-tool guide tells you (REPORTED, explainer)

A dev.to guide argues the best AI for coding depends on the work to finish. It points to Cursor (AI code editor) for editor-centered work, Claude Code (Anthropic's terminal tool) for terminal development, and Codex (OpenAI's coding agent) for agent-assisted repository work. Source: dev.to

Our analysis: three working styles. In an editor, you drive and the AI suggests. In a terminal, you steer and watch commands run. With a repository agent, you hand over a task and review the result.

What it lacks: no benchmarks, no prices (not stated), little new for regular users. Treat it as a checklist, not proof one tool wins.

Three questions:

• Where do I want to read changes: editor, terminal, or finished review?
• How big is the task: one function, one feature, or a repository-wide chore?
• How much checking can I afford: line by line, or tests?

Who benefits: developers picking a first or second tool. Skip: non-coders. Limit: guides date quickly, so re-run the questions every few months.

The loop inside every coding agent. Evergreen lesson, not news.

A coding agent is a model in a loop called the harness, the program giving it tools and rules.

1. Goal in: the harness packs task, files and instructions into the context (all the model can see).
2. Action out: the model edits nothing; it emits tool calls: 'read file', 'search name', 'run tests'.
3. Execution: the harness runs it under permission rules (e.g. asking first).
4. Observation back: output (file text, test failure) joins context.
5. Repeat until the model says done, hits a step limit or you stop it.

Why it works: tests, compilers and linters give ground truth, so a model reading failures can self-correct and check its work.

Novel vs incremental: the tool-calling format and harness, not the model. Cursor, Claude Code and Codex differ in tools, permission asks and context management.

Breaks: context fills and crowds out early instructions; loops on one failing fix; no tests, silent errors.

Practical: give a test command, keep tasks small, define 'done'. Analysis: tool gaps come as much from harness design as model choice.

Two-model bake-off: pick by evidence, not brand. Illustrative workflow, not tested by us.

Copilot Premium if your plan offers both (price not stated; confirm on Microsoft's page), or ChatGPT and Claude.

1. Pick three real tasks: writing, analysis, coding or spreadsheet.
2. Same prompt for both, fresh conversation each run; record model names.
3. Paste answers into a document as A and B, shuffled; score blind 1 to 5 on correctness, voice or format fit, editing needed, reasoning or flagged uncertainty.
4. Note winners per task type, e.g. 'drafting: model B'.

Small sample, so run important tasks twice; models update, so redo every few months.

Verify next week: use your note on a fresh task.

Tangents: separate chat for side questions; paste back only the conclusion.

Write a reusable 'house brief' once and stop retyping context.

The number (an estimate, not from today's stories): if you spend about 5 minutes setting context on each of 10 prompts a week, that is roughly 50 minutes, close to a lunch break. A brief that takes about 20 minutes to write once can cut most of that. Your own numbers will differ.

How: write one page covering who you are, your audience, your tone, your format rules and what 'good' looks like. Save it where your tool stores standing instructions. Options include custom instructions in ChatGPT (OpenAI) and project instructions in Claude (Anthropic). Features and plan limits change, so check the current price and limits on each pricing page. Today's stories state no price.

Money angle: before adding a second AI subscription for model choice, run the bake-off above on the plan you already have. Copilot Premium reportedly offers both OpenAI and Anthropic models (price not stated), so check the cost against paying two vendors directly.

What you give up: a stale brief quietly steers every answer. Review it monthly, and drop it for tasks where you want a fresh, unbiased take.

Prompt clinic: delegating coding. Illustrative, untested.

Weak: Review my code and make it better.

Fails: 'better' undefined, scope unlimited, unverifiable, rewrites possible.

Better:

Goal: fix [symptom, e.g. wrong discount totals].
Scope: only [folder]; keep public function names; no new dependencies.
Context: logic in [file]; run [test command] before, after.
Done: existing tests pass; new test fails before, passes after.
Process: plan in three bullets, wait, then smallest fix.
Output: summary of changes, doubts.

Works: tight scope, runnable test, defined done, plan-first.

Adapt: swap folder, test command, symptom; writing: rules checklist.

Check: ask 'which constraints did you skip?' 'None' yet other files changed? Read diff.

Put key instructions first and last. Published research, not today's stories.

A Stanford study, 'Lost in the Middle', found models used starts and ends better than middles. Newer models may differ.

Put your question first and last when pasting long documents; keep rules out of the middle.

Agents lead, then open-source AI and policy. Copilot hosting two vendors' models suggests assistants becoming marketplaces. Analysis, not confirmed: comparing models quickly beats loyalty.

Watch: Microsoft confirming Copilot model names and price; OpenAI answering Tangents (both open).

I asked two AI models to settle a debate. They agreed, then cited each other.

That's it for today. Tomorrow we're watching whether Microsoft's own pages confirm the Copilot model lineup and whether OpenAI responds to the Tangents proposal.

Daily listening is learning with a reason: today it was a bake-off, a house brief and a prompt you can use at work. If this made you smarter, share it with one person who needs it. Thank you for spending your time with us.

SI Agent News — a podcast newsletter.

Made by AI.