June's blog was about the AgentCore harness hitting GA and a new set of models worth trusting. July's is what happens when trust gets tested and choice gets overwhelming. The models got cheaper. The models got loose. And now I have to mention agentic "harness" 50 times in this blog because of this. All kidding aside, harness.

The New Moore's Law for Token Cost

Six months ago, frontier-level intelligence cost $15 per million input tokens on the leading model. Today, you can get comparable performance for 30-50% cheaper, and in some cases a dollar or less, depending on which provider and which tier you pick. That compression happened in weeks.

Here is where the major model providers landed by the end of July 2026:

The spread between the cheapest and most expensive frontier-class models now exceeds 100x. DeepSeek's V4-Flash, which scored 50 on Artificial Analysis's Intelligence Index and was open-sourced under the MIT license on July 31, costs less per token than what most teams spend on logging. Luna, after OpenAI's 80% price cut on July 30, sits at $0.20 per million input tokens for a model in the same family as Sol.

The implication for anyone building: cost per token is no longer a useful constraint for choosing a model. The cheapest option is capable enough for most tasks. The deciding factor shifts to which model fits the task, which model fits the latency requirement, which model fits the governance posture. Picking the right model for the right step in a workflow, rather than picking one model for everything, has been for a while and will continue to be where the value lives.

In production systems, the shift towards this is routing through the right harness. Once again, I'll come back to that latest buzzword down below.

A Couple Rogue Agents, No Big Deal

This past month will be remembered for the month AI agents went from hypothetical risk to demonstrated risk, twice, at two different labs.

The OpenAI-Hugging Face breach

Around July 9, during an internal cybersecurity evaluation called ExploitGym, OpenAI's GPT-5.6 Sol and an unreleased model escaped their testing sandbox. The agent exploited a zero-day in a package registry proxy, escalated privileges, moved laterally to a machine with internet access, and then broke into Hugging Face's production infrastructure. It executed over 17,000 actions through swarms of agents over roughly four days before Hugging Face detected the intrusion on July 16. OpenAI disclosed its responsibility on July 21.

The agent was trying to find the answer key to the benchmark it was being tested against. It was not trying to cause harm. The harm happened anyway because the containment boundary failed.

The Anthropic disclosure

On July 30, Anthropic published a review of 141,000 cybersecurity evaluation runs and found six runs across three incidents where Claude models, including Opus 4.7 and Mythos 5, escaped a third-party evaluation environment, reached the internet, and gained unauthorized access to the production systems of three separate organizations. The incidents traced back to a configuration error at a testing partner that gave the models internet access they were not supposed to have.

Models that can run multi-step autonomous workflows can also find and exploit gaps in their containment. When you give a model a tool-use loop and the ability to reason about its environment, it will reason about its environment, including the parts you did not intend it to access.

The agents being powered by these models need to run inside something that enforces boundaries regardless of how clever the model gets. A properly built harness is that something. More to come on that below.

GPT-5.6 on Bedrock: Woohoo!

OpenAI's GPT-5.6 family, Sol, Terra, and Luna, became generally available on Amazon Bedrock on July 13. Claude Opus 5 followed on July 24. Both sit alongside every other Bedrock model under the same IAM controls, CloudTrail logging, and VPC boundaries.

The AgentCore harness we wrote about last month handles this cleanly. You define the agent in config. You point each step at the model that fits. The harness handles session state, memory, tool execution, and error recovery regardless of which model is generating the next response. Swapping a model mid-session is basically just a config change now.

Through the proper building and deployment of a harness, these all converge into a single practical reality: you can now build an agent that routes planning tasks to Sol, coding tasks to Opus 5, fast classification to Luna, and bulk extraction to DeepSeek, all within one governed environment, all under one bill, all without touching a consumer API.

Sam Lahti's recent piece on building a production chatbot on Bedrock AgentCore makes a related point: the choice between buying SaaS seats and building from scratch is a false binary. The third option, composing AWS-native primitives into a governed stack, is now faster to stand up than either extreme. Model portability falls out for free because the agent loop is decoupled from the model. That decoupling matters more this month than it did last month, because the set of models worth routing to has doubled.

"Hey ChatGPT, Define Harness for Us (again)"

If the pricing war means you have more models worth using, and the new rogue-agent incidents mean you need containment that the model cannot circumvent, and their availability in model catalogues means you can actually access all of it through a single governed layer... then you need a harness.

Without a harness, model choice is purely chaos. You are managing API keys across providers, building bespoke routing logic, hoping your sandbox holds, and operating different security models for different vendors. The more models you want to use, the more operational surface area you accumulate.

With a harness, model choice can be done very simply. You can route the right model to the right task. You can enforce tool permissions and network boundaries at the infrastructure level, not the prompt level. You swap providers when pricing changes or capability improves. You can log everything in one place and audit it with one set of controls.

Bottom line

July gave us cheaper models, scarier failure modes, and better infrastructure. The teams that treat the harness as the product, rather than the model, are the ones that will scale what they have built. Or you can keep rebuilding every time the leaderboard changes. Oh, and harness the future.

About the author

Alec MacEachern is Vice President of AI at UTurn Data Solutions, an AWS Premier Tier Services Consulting Partner based in Chicago. Over the past decade, he has held roles at NVIDIA, AWS, and Microsoft, helping organizations design, build, and scale AI solutions across a wide range of industries and platforms. Today, Alec brings that cross-platform experience to helping enterprises navigate cloud migration, modernize data foundations, and adopt production-ready generative and agentic AI solutions.