Good Enough Wins

Top-Line Summary

Claude Opus 5.5, GPT-6 Sol and Luna, and Kimi K3 all reached Amazon Bedrock in the same week, and the best of them got cheaper at $4 in and $20 out per million tokens. At the same time, Meta’s Muse agent beat ChatGPT in downloads on a model nobody calls frontier, and Xiaomi’s MiMo became Travers’ default for paid API work at a tenth of the price. So does the best model still matter, or does good enough win?

Show Video

The “Pre-Show” Context

Brett has spent two weeks teaching, and his token consumption has “fallen off a cliff” as a result. He did get stumped in class by an obscure Auto Scaling group setting, so at the break he told Claude not to speculate and to use the AWS documentation MCP server, and it came back with the answer and a reference.

The Engineering Rundown

• Claude Opus 5.5 on Amazon Bedrock (00:32)

  Opus 5.5 was on Bedrock on launch day. Anthropic says it matches Fable 5.1 on most work, costs about 40% less to run than Opus 5 and is roughly 30% faster, and the price drops from $5/$25 to $4/$20 per million tokens. Brett’s Auto Scaling question was his first real test and it felt snappier and more concise, with an answer instead of a thesis. Both hosts noticed the writing reads better too, with less of the academic style and fewer em-dashes.

• GPT-6 Sol, Luna and Kimi K3 on Bedrock (03:07)

  GPT-6 Sol and Luna also arrived on launch day, while Kimi K3 took about 63 days to get to Bedrock. That gap is growing: Kimi K2 took 26 days and K2.5 took 14. The hosts’ guess is compute, with Amazon’s investments in Anthropic and OpenAI getting first call on capacity and those companies wanting token revenue back. The old “your data stays in your AWS account” pitch might explain some delay for third-party models, though that argument is weaker now that some Anthropic and OpenAI models keep logs anyway. Travers wants GLM 5.3 on Bedrock and something newer than GLM 5 in Kiro.

• Getting started with tokenomics on AWS (07:08)

  “Tokenomics” is FinOps for AI spend, and the post covers CUR 2.0 with IAM principal allocation, Bedrock application inference profiles, invocation logging, Budgets and Cost Anomaly Detection. Brett already uses principal allocation on his own Bedrock-heavy project to see which functions drive the bill. Application inference profiles were new to him: they group inference for a model so you can track and alarm on it. The post claims the same task cost 1 cent on Kiro against 7 cents on Claude Code, which got an eyebrow raise because it’s an AWS blog grading AWS’s own tool. The likely reason is Kiro’s auto mode routing work to cheaper models, and Brett’s weekend projects on auto mode used almost nothing.

• Lambda MicroVMs for self-hosted agent sandboxes (11:43)

  Lambda MicroVMs are Firecracker VMs with full OS access for up to 8 hours, each with its own memory, disk and network. They boot from a memory and disk snapshot, scale up to 4x while running, and idle policies suspend and then kill unused VMs, so you pay for run time only. This is the isolation layer the software factories from EP60 need: if Steve Yegge is burning something like $87,000 a month in tokens across a fleet of agents, every one of them needs somewhere safe to run code. Cloudflare has a similar option, with some differences in how it spins up.

• Amazon EC2 T8i instances (14:13)

  This was parked on the radar list, but Brett couldn’t resist. T8i is the new burstable family on a custom Intel processor, with up to 30% better price performance than T3, and the micro and small sizes are Free Tier eligible. It’s also in Canada Central. Travers’ question stands: what happened to T5, T6 and T7? And yes, it’s one more instance type on a list that is already absurdly long.

• Amazon CloudWatch Omni (16:11)

  Omni is generally available only in N. Virginia, Oregon and Ireland. It runs as its own web experience (plus IDE extensions for VS Code, Cursor and Kiro), uses the DevOps Agent for root cause analysis, can span multiple accounts and pulls in Azure. Pricing stacks data ingestion (the hosts recalled around 50 cents a GB) on top of Bedrock evaluation charges, and new sign-ups get a $1,000 credit for 30 days. Brett’s rule: when AWS hands you $1,000 to try something instead of $50, it’s going to be expensive, and the pricing examples ran to thousands a month. Neither host is sold, since an agent with read-only access to your accounts, possibly on an open-weights model, could probably get you most of the way to root cause already.

• Travers’ projects: Melee netplay and an Office clone in Rust (19:58)

  Travers burned through his GPT subscription plus all three of the Astra resets OpenAI handed out. The Super Smash Bros. Melee project is working, with the last netplay issues getting tested this weekend, and he has started building his own Microsoft Office clone in Rust with his agent setup. That’s pushing him into program performance and, apparently, toward C. He’s also been tuning verification loops, on the theory that you only get to a Yegge-style agent fleet once your checks earn enough trust to let the agents run unattended. Results so far: mixed.

• Nightly code reviews with Kiro schedules (23:17)

  Brett moved a nightly job he used to run in the Codex app over to Kiro’s scheduled jobs, with his laptop set to never sleep. Every night an agent reads the last 24 hours of commits across three of his projects, looks for security issues and bugs, and files GitLab issues, then posts a summary to Slack in the morning. He doesn’t always agree with its findings, but it’s interesting to see what another agent picks up. Kiro also lets you define dedicated agents with their own instructions and model, so next up is two reviewers on different models looking at the same commits.

• Amazon Quick: out of credits and stuck at sign-in (26:20)

  Brett connected Quick to Gmail, Calendar and Slack on a schedule and ran out of free-plan credits in about three days. Your data sits in an AWS account Quick creates on your behalf, and you can export or delete it. Hooking it up to his organization was the problem: the agentic features are only in us-east-1 and one other Region, his IAM Identity Center is in ca-central-1, and fixing that means a multi-Region Identity Center deployment with a replicated KMS key. The Android app wouldn’t even sign him in, with an error about his ID being used by another authentication method, and builder IDs have always been rough. It did surface useful things, like flagging that Travers was waiting on a reply, and he might pay $20 a month for a while to see what it can do.

• Meta’s Muse out-downloads ChatGPT (33:04)

  Muse launched Sept 8 as an agent that sends email, books travel and makes purchases, the same pitch as Quick. Sensor Tower numbers put it at 2.5 million downloads in 13 days and ahead of ChatGPT over a comparable period, and it took #1 free iOS app in the US; Brett’s live check of the Play Store showed Quick at about 50,000 and Muse at 500,000. Travers had heard Amazon is blocking Muse from ordering on its shopping sites, which is rich from a company that scraped everyone else’s data. Muse doesn’t run on a frontier model, and it doesn’t need one when the job is reading your calendar and your friends’ activity. The worry is lock-in: the longer an agent learns who you talk to, the more it costs to leave, and Meta starts with more human interaction data than anyone except maybe Google.

• OpenAI agents used a U of T link shortener as a message board (42:14)

  Back in June, OpenAI agents used a University of Toronto link shortener to leave links for themselves and other agents to pick up. OpenAI told U of T, the university says there was no breach, and the shortener use has been shut off. Brett’s question: how many times does a model that “wasn’t supposed to have internet access” get out before someone unplugs the Cat5 cable? The answer from the other chair: it was a cloud misconfiguration, so call your local cloud experts. Expect more of these stories for months.

• Xiaomi MiMo V2.6 Pro (45:00)

  Yes, Xiaomi the phone company: MiMo V2.6 Pro is open weights under MIT with a 1M context window, at $0.43 in and $0.87 out per million tokens. It is now Travers’ default for anything that costs real API dollars, running as the sub-agent model in his Pi harness through OpenRouter after his OpenAI subscription ran dry. On the price-to-performance chart nothing beats it until you step up to something like Opus 5.5. Brett admitted he’s “so married to Anthropic” that he still uses Claude for coding when it should be planning with Opus and handing the programming tasks to something cheap. He also has API keys scattered across Moonshot, Z.ai and others, which is probably the push to finally set up OpenRouter.

• Claude finds a possible new gene-editing system (50:11)

  Posted the day of the show: about 950 Claude agents searched a DNA sequence database for 21 hours and found an unknown enzyme system that looks somewhat similar to CRISPR, which Anthropic then tested in its own wet lab. It’s a preprint, not peer-reviewed, and Anthropic doesn’t know yet what it does. It’s the counterweight to the Muse argument: good enough may win for consumers, but science still needs the frontier. Brett’s problem is the messaging, since last week the story was that AI needs slowing down because it might wipe out humanity, and this week it gets a wet lab. Both hosts want this research to continue, but “that whole posture is bullshit” if the fear only shows up when it helps raise money or scares people off cheap open-weights models.

• TypeSafe’s Jev classifier model (57:17)

  Jev comes from TypeSafe, a company started by an ex-OpenAI engineer. You send it structured JSON describing some state and it returns a yes/no or a confidence score, and the launch demo had it playing Doom in real time. It’s pitched as far cheaper than frontier models ($0.042 per million input tokens, free output), and Travers says it’s a fine-tuned Qwen underneath. Neither host has built with it yet, but the obvious use is a smart switch statement, like deciding whether an alert is worth waking someone up for. Sean Goedecke’s post makes the next point: once a model like this works, you can collect its inputs and outputs and train your own cheaper replacement.

Off-the-Clock Recommendations

• Resident Evil (2026): Travers recommends it. Very little carries over from the earlier films or the games apart from the general idea, and it’s quite good.
• The Wraith (1986): Charlie Sheen comes back as a mysterious racer in a very fast car and loses a bit of his armor with each act of revenge. Two and a half stars on Letterboxd, and it’s bad in the way you want.
• Tremors (1990): Five stars, which under Brett’s rule means you watch it every single time it comes on TV.
• The Bag Man (2014): John Cusack as a delivery guy for Robert De Niro’s gangster. Two and a half stars, and a strong contender for the next bad movie night.