Hey there 👋 Every so often a launch buries its own headline. OpenAI shipped a new model this week and most of the coverage led with benchmarks. I went hunting for the benchmark that mattered and found it on the price list instead. The cost of a finished agent task fell by roughly ten times. I sat with that for a second, because a cost drop that size changes which jobs are even worth handing to an agent. The rest of the week lined up behind it: a place to run all those new jobs without getting burned, and a discount that hands agents to a set of buyers who could not afford them a month ago. One theme, three rooms. - Ron
The Big ThingOpenAI's GPT-6.1 Terra cut the price of running an agent about tenfoldOpenAI shipped GPT-6.1 Terra this week, and the headline is not a benchmark. It is the invoice. API input dropped to $1 per million tokens, down from $2 on the last Terra. Output runs $5 against the old $12. Cached input fell to a nickel, a quarter of what it was. On Artificial Analysis's own index the model finishes a task for about $0.40 where its predecessor charged $1.40, and it matches GPT-6 Astra, a much larger model, at roughly one-tenth the token price. Call it a tenfold cut on the thing you pay for, which is a finished run. It is no toy, either. Terminal-Bench 4.0 jumped to 50.3 percent of tasks solved at full effort, up from 21.5 on the prior version. When it does not know an answer it now guesses 38 percent of the time instead of 88, so it invents far less. It runs about three times faster than the bigger Astra and Sol models. You can use it today through the API and inside ChatGPT Work and Codex. Now walk the math forward. For two years the ceiling on how much work you handed a fleet was cost per finished task. Drop that by an order of magnitude and a whole shelf of jobs you parked last month comes back into budget: classification across your entire backlog, an agent that reviews every pull request instead of only the scary ones, overnight research runs you would never have paid to kick off. The cheaper the run, the more of them you set loose. One honest caveat. Terra is the cost and speed play, built to do ordinary agent work for pennies. OpenAI still points you at GPT-6.1 Sol and GPT-6 Astra for the hardest coding, and Terra's coding score trails Sol's. Its safety numbers held, and its defense against indirect prompt injection even climbed to 98.8 percent, so the tradeoff is raw horsepower on the toughest problems, not trust. Why it matters: the cost per finished agent run just fell about tenfold, so the work that was too expensive to automate last month is back on the table this week. Ship it? DEPLOY for agent, computer-use, and high-volume work where cost is the cap. Keep a bigger model wired in for the gnarly coding, then go dig up the workload you priced out last quarter. Sources: OpenAI, The Decoder, TechRepublic, gHacks.
Tour de Headlines📦 Akamai put its Agent Sandbox into tech preview this week, and the setting worth copying is the default. Code an agent generates gets no outbound network access unless you grant it. The agent itself runs as a WebAssembly app, and the Python or JavaScript it writes runs in a locked-down runtime nested inside that, with execution limits the agent cannot raise on its own. Call it a blast radius of zero until you decide otherwise. For now it suits short jobs, each step finishing in about a minute, with longer-running agents promised later. Access is request-only, so treat it as a pilot. Even if you sandbox somewhere else, steal the deny-by-default rule. Verdict: PILOT. Akamai 💸 Anthropic widened its GSA deal this week, and if you build agents for the public sector your pricing conversation got easier. It is a 40 percent discount on Claude Enterprise, Claude for Government, and the Claude API, available today through GSA's Multiple Award Schedule with no minimum spend. It does not stop at federal. The discount reaches state, local, tribal, and territorial governments too. Pair that with Terra's price cut and the cost of adopting agents fell on two fronts in seven days. The model got cheaper to run, and the contract got cheaper to sign. A whole buyer class that could not clear procurement last quarter is suddenly reachable. Verdict: NOTE. Anthropic 🌍 Google DeepMind previewed Genie 4, and it earns the delight slot: something you cannot deploy and will want to anyway. It is a world model that generates an explorable scene in real time. You steer it with text, change the weather or drop in an object, and it keeps the physics roughly honest while characters move and react. It generates the ambient sound and dialogue as you go, which the team calls a first for a real-time world model. Scenes hold together over short sessions and drift on longer ones. It is a research preview for trusted testers, with a waitlist and no ship date. Watch this one, you cannot run it yet. Verdict: WATCH. Google DeepMind
Sponsor Your team's calls are full of signal. RapportScore reads it. RapportScore scores how your team communicates on real recorded calls and coaches them on it, so you can see where rapport holds and where it slips. Built for B2B teams who live on the phone. See your team’s score → |
Tool of the Day📨 FrancoAn outbound sales agent that runs on the Claude Code or Codex subscription you already pay for. Franco runs on your own machine, on a subscription you already have. Point it at your accounts and it drafts a personalized email sequence with the model, or sends copy you wrote word for word if you would rather it not improvise. It reads replies, sorts them by intent, logs opt-outs, and catches bounces. Every outgoing message waits in an outbox for your sign-off unless you turn on autopilot, and every fact it cites is stored with a source link and a date. It is MIT-licensed and new, around 300 stars, so treat it as a sharp utility to test before you lean on it. Keep autopilot off until the drafts earn it, and mind your own sending reputation. Verdict: TRY. Get Franco on GitHub →
Worth a Click- GitHub Agentic Workflows shipped six point releases (v0.93.0 through v0.93.5) on October 26. They tighten static analysis, handle network and credential use in integration jobs more safely, sharpen failure diagnostics, and make Pi CLI 1.4.0 the default. If you run these in CI, test the bump in a throwaway repo first.
- gitai is version control built for AI coding agents. It runs on top of Git but lands changes through compare-and-swap instead of a merge queue, so a fleet of agents writing in parallel does not clobber each other. MIT-licensed, around 1,000 stars, currently v0.4.0.
Last week an agent got a meter. This week it got cheap. GPT-6.1 Terra finishes a task for about a tenth of what a comparable model charged, and in the same seven days a sandbox showed up to run the overflow safely and a government discount opened a fresh set of buyers. A price cut that big never gets spent doing the same work for less. It gets spent doing far more of it. So the real budget question is how many agent runs you can afford to turn loose, and where they execute. Settle the blast radius before you settle the volume. Go find the one workload you priced out last quarter. It just got cheap enough to automate. Future-you, reading a smaller bill, will thank you. - Ron |
