Conversational AI, Part 2
Published by Sam Hardin | Bright Pivot LLC
A few months back, I put together a glossary of the AI terms most people following the news were running into: agent, prompt, hallucination, RAG, the words that were suddenly everywhere but rarely explained in plain English. That post is still good, and it's still worth a read if you're just getting your footing.
This is the second batch. The lightning pace of the AI industry means new terms show up fast. These are the ones that are worth your attention. Some of these are technical concepts that can help you use AI better, the rest are just useful vocabulary for conversations you might already be having.
Multimodal
A multimodal AI model can work with more than just text, it can take in images, audio, or video, and often produce more than one of those back. Ask it to read a photo of a whiteboard and turn the notes into a task list, or describe what's happening in a security camera clip, and you're using multimodal capability.
For a business owner, this matters because a lot of the useful raw material sitting around a business isn't text. It's photos of equipment, scanned paperwork, a video walkthrough of a process. Multimodal models are what make that material usable by AI as isl, instead of needing someone to manually transcribe everything first.
Context Window
Every AI model has a limit on how much information it can actively work with at one time, its context window. Think of it as a snap shot, everything currently "in view" for the model, the conversation so far, any documents you've shared, its own previous responses.
Context windows have grown enormous, some models now handle the equivalent of several long novels at once. But size isn’t the only consideration, and there are two practical things worth understanding.
First, not everything fills the window at the same rate. A page of plain text might use a few hundred tokens. A single image or a scanned PDF page can use many times that, since the model has to process the visual information itself before it ever gets to reading the words on it. If you're uploading a stack of scanned documents or photos, expect the window to fill up faster than the same information typed as plain text would.
Second, what happens when a conversation runs long enough to approach that context window’s limit can impact accuracy and response time. Most AI systems handle this through compaction, automatically summarizing older parts of the conversation to make room for new material. That summary is genuinely useful, but, like any summary you lose detail. An AI summarizing a hundred messages into a paragraph is going to drop some nuance, and it decides what to keep on its own. Research has found that a meaningful share of AI task failures trace back to exactly this, a system gradually drifting off track as its grip on earlier context loosens, not because it ran out of room, but because it was work off of a summary of a summary of a summary.
The practical takeaway: for anything that matters, don't wait until a conversation is forced to compact itself. Periodically wrapping up and starting a clean session, carrying forward just the key decisions and current state, tends to produce sharper, more reliable results than letting one conversation run indefinitely.
And you do that with…
Context Handoff
Related to the point above, a context handoff is the moment a conversation or task moves from one place to another: one AI session to a new one, an AI agent handing a customer over to a human, or one specialized agent passing work to another. The whole question a handoff has to answer is what gets handed off. What's already been established, who the customer is, what's been tried, what the goal is, all needs to transfer cleanly to whoever picks it up. Otherwise, they have to start from scratch.
A good handoff preserves the useful parts and leaves behind the noise. A bad one either loses real information or drags along so much unfiltered history that it slows everything down. If you're building any kind of automation where AI hands work off to a person, or one AI system hands off to another, this is the specific thing worth testing carefully, not just whether the handoff happens, but whether what actually matters survives it.
Context handoffs are important to do regularly when working with chatbots. Even before a conversation hits its technical limit, quality tends to slip as a single session runs long. The model's grip on earlier details gets fuzzy, and letting it compact and summarize on its own means losing some of that detail for good. If you haven’t already, make a habit of wrapping up periodically by asking the chatbot for a context handoff to use to start a new session. It will pull out the key info, decisions, and current state, and give you either a file or a text block to paste into a new chat. Starting fresh with that summary keeps the work sharper than riding one long conversation until it's forced to compact itself multiple times.
Reasoning Model
Most AI models respond quickly, generating an answer in something close to one continuous pass. A reasoning model is built to slow down first, working through a problem in intermediate steps before producing a final answer, closer to how a person might sketch out their thinking on scratch paper before writing a final response.
This tends to produce better results on genuinely hard problems, math, multi-step logic, anything where getting it right requires actually working through it rather than pattern-matching to something similar. The tradeoff is cost and speed, reasoning models are slower and meaningfully more expensive per response, since all that intermediate thinking uses tokens too, even though you typically don't see it. Not every task needs a reasoning model. They’re typically reserved for issues that actually require the extra thinking.
MCP (Model Context Protocol)
If you only take one term away from this post, make it this one. MCP is probably the most important development in how AI actually gets things done.
Here's the plain-English version. AI models are only as useful as what they can actually reach: your calendar, your files, your CRM, your inbox. For years, every connection between an AI system and an outside tool had to be custom-built, one at a time, which made real integration slow and expensive. MCP is an open standard, originally released by Anthropic, that gives AI systems and outside tools a common language to talk to each other. It’s sometimes called the USB-C of AI, a single standard connector instead of a drawer full of incompatible cables for every device you own.
Practically, this is why AI agents can now check your calendar, pull a file from Google Drive, or update a record in Airtable without someone hand-building that specific connection from scratch. It's the plumbing behind the lion’s share of what makes agent-based automation actually work, and it's a big part of why the pace of what's possible has picked up so much over the past year. Adoption has moved fast, most of the major platforms either support it already or are actively building toward it.
Guardrails
Guardrails are the boundaries put around what an AI system is allowed to do, what data it can access, what actions it can take on its own, and where it has to stop and ask a person before proceeding. They are a major part of establishing effective security for AI systems. Good guardrails aren't about limiting what AI can do for the sake of it, they're about making sure autonomy is scoped to exactly what a task actually needs, and nothing more.
This term has become more prominent in recent headlines for a good reason, and it connects directly to the next one.
Sandbox
A sandbox is an isolated testing environment, a space where an AI system can be run or evaluated without real access to live systems, data, or the open internet. It's a standard, sensible safety practice – test something risky somewhere it can't cause any real harm.
At least, that’s what’s supposed to happen. Over the summer of 2026, several AI labs, including OpenAI, Anthropic, and Meta, separately disclosed incidents where models under evaluation reached systems well outside their intended sandbox. Most of these weren't dramatic technical breakouts, they were environments that were never properly isolated in the first place, a misconfiguration rather than a broken lock. One case did involve a genuine exploit, a model chaining together real vulnerabilities to escape a properly built sandbox entirely.
What we all learned is that the word "sandbox" describes an intention, not a guarantee. A sandbox is only as secure as its actual configuration, and that's worth remembering any time a vendor tells you their AI system is "safely contained" without explaining exactly how.
Deterministic vs. Non-Deterministic
This one explains something you've probably already noticed and wondered about. A deterministic system gives you the same output every time you give it the same input, like a calculator or a formula in a spreadsheet. A non-deterministic system can give you a different answer each time, even with the exact same input.
AI models are non-deterministic by nature. Ask the same question twice and you may get two differently worded, sometimes meaningfully different, answers. This is also part of why hallucinations happen, the model isn't looking up a fixed fact, it's generating a statistically likely response each time, which is usually right and occasionally isn't.
For someone running a repetitive process, this can have real implications. Anything you're automating that requires perfect consistency every single time: a legal filing, a regulatory calculation, needs a deterministic system or a human check built in, not an AI model working alone. Anything where a range of good answers is fine: drafting, brainstorming, summarizing, is a much better fit for AI's non-deterministic nature.
There are ways to parse a model’s output so that you get the best of both worlds, but it takes real understanding of what specific outputs are expected and what systems are running around them.
Post-Transformer Architecture
Nearly every mainstream AI model in use today, GPT, Claude, Gemini, is built on something called a Transformer, the underlying architecture that's powered most of the progress in this field since 2017. A Post-Transformer architecture is exactly what it sounds like, a fundamentally different design, not just a bigger or better-tuned version of the same approach.
Worth knowing about right now is BDH, an architecture from a lab called Pathway, loosely inspired by how neurons actually connect and reason in biological brains rather than the token-by-token prediction Transformers use. A specific version of it, tested publicly this August, scored meaningfully well on a difficult reasoning benchmark using a fraction of the size and cost of a comparable frontier model, roughly eleven times cheaper per task. It's not more accurate than the best frontier models yet, and it's a small research model, not a finished product. But it's a real, independently verified demonstration that a different, far cheaper approach can get close to frontier performance on certain problems.
This isn't a "the next big thing has arrived" story, but I’m keeping an eye on it with interest. It’s a sign that the current approach to building these models isn't the only road forward, and that cost curves in this industry can bend in unexpected ways.
Updated Pricing: Where Things Stand Right Now
Model pricing continues to drop sharply as the major labs compete, and the specific numbers from the first glossary post are already out of date. Here's where things stand as of this post, priced per million tokens processed (input / output):
OpenAI:
GPT-5.6 Luna: $0.20 / $1.20 per million tokens (cut 80% on July 30, 2026)
GPT-5.6 Terra: $2 / $12 per million tokens (cut 20% on July 30, 2026)
GPT-5.6 Sol: $5 / $30 per million tokens (flagship tier, unchanged since its July 9 launch)
This pricing is for short-context requests. All three tiers step up meaningfully on longer prompts, Sol rises to 10/45, Terra to 4/18, Luna to 0.40/1.80, once a request crosses OpenAI's long-context threshold. Given how relevant that is to your context window section right above it in the post, it might be worth a one-line note tying the two together.
Anthropic:
Claude Haiku 4.5: $1 / $5
Claude Sonnet 5: $2 / $10 through the end of August, stepping up to $3 / $15 starting September 1, 2026
Claude Opus 5: $5 / $25
Claude Fable 5: pricing generally mirrors Sonnet 5's tier for now but is expected to increase. Several deadlines for this have been extended.Also, it's worth knowing that Fable 5, and its sibling model Mythos 5, briefly had access suspended in June to comply with export controls from the US Government before being restored on July 1. Worth checking current availability and terms directly before building anything that depends on it.
Google:
Gemini 3.5 Flash-Lite: $0.30 / $2.50 per million tokens (Google's budget tier)
Gemini 3.5 Flash: $1.50 / $9 per million tokens (Google's workhorse for coding and agentic tasks)
Gemini 3.1 Pro: $2 / $12 per million tokens for prompts up to 200K tokens, stepping to $4 / $18 beyond that (Google's flagship reasoning model)
XAI:
Grok 4.5 / Grok 4.6: $2 / $6 per million tokens for prompts under 200K tokens, stepping to $4 / $12 beyond that (xAI's current flagship, positioned as an aggressive value play against Claude and GPT on output cost specifically)
The overall trend holds steady, model pricing has fallen roughly 80% over the past year across the industry, and reasoning-specific pricing tiers are becoming more common as models like these get faster at deciding how much "thinking" a given task actually needs. If you built cost assumptions into a project six months ago, it's worth revisiting them. The math has almost certainly gotten better.
That’s it for this round. I’ll keep these going as new important terms rise to the surface. Let me know if an unfamiliar term or concept keeps showing up in your feed. I’m happy to talk about it and see if it’s something you should be interested in.
Sam Hardin is the founder of Bright Pivot LLC, an AI automation and operations consultancy based in Covington, Louisiana. Bright Pivot helps small and medium businesses on the North Shore automate the routine and augment the rest.
Have questions about what this means for your business? I'd love to talk it through. Book a time on my Calendly, or reach me directly at sam@bright-pivot.com.
Source Links
Anthropic, "Introducing the Model Context Protocol": https://www.anthropic.com/news/model-context-protocol
IBM, "What is Model Context Protocol (MCP)?": https://www.ibm.com/think/topics/model-context-protocol
Pathway, BDH-CQ ARC-AGI-1 benchmark announcement (August 11, 2026): https://www.businesswire.com/news/home/20260811268264/en/Pathways-150M-Parameter-Model-Breaks-the-ARC-AGI-1-Cost-Efficiency-Frontier
Pathway, BDH architecture GitHub repository: https://github.com/pathwaycom/bdh
Zylos Research, AI agent context compression and drift: https://zylos.ai/research/2026-02-28-ai-agent-context-compression-strategies/
Anthropic, Claude Platform Docs on compaction: https://platform.claude.com/docs/en/build-with-claude/compaction
OpenAI API Pricing: https://developers.openai.com/api/docs/pricing
Gemini Developer API Pricing: https://ai.google.dev/gemini-api/docs/pricing
xAI (Grok) API Pricing: https://docs.x.ai/developers/pricing
