Your agent finishes in fewer tokens

Hey there 👋

I spent a while in Reflection's benchmark chart this week, the one for Beam, the model they just put out in the open, and the detail that stuck with me was a small tell about the pitch. A model selling itself on efficiency is parked in the middle of its own table rather than at the top of it. Beam clears GLM-5.2 on both score and tokens, which is the efficiency case in one line, yet DeepSeek's new open model and Kimi K3 both post higher accuracy on the same benchmark.

If I had to bet, that placement is the story, because the pitch was never really about being the smartest model in the room. Two open-weight models showed up in one window and both made the same quiet argument, that the number that matters for an agent you leave running is how many tokens it burns to finish, and the per-token sticker barely enters into it. That is a different way to shop for a model, and it earns five minutes.

- Ron

The Big Thing

Reflection open-sourced a model that finishes the job in a third of the tokens

Reflection released Beam on October 16, its first open model, weights under Apache 2.0 alongside a companion open operating system for running it. The shape: 501 billion parameters total, 23 billion active per token, so the 8-bit and 4-bit builds fit on an eight-GPU server and the full-precision one needs tens of them. You can pull it today. Baseten started serving it through OpenRouter the day of release, and the hosted version Reflection runs is free during a public beta while they settle on pricing.

The pitch is the part worth your attention. Reflection is pitching Beam on one number, how cheaply it reaches the finish line. On Terminal-Bench 2.1, a coding-agent test, Beam scores 80.1 percent while spending about 25,000 tokens per attempt. GLM-5.2, the model it lines itself up against, burns roughly 80,000 tokens for the same work. That is about a third of the tokens to land a comparable result, and for an agent grinding through a long task with nobody watching, tokens are the bill.

Now the skeptic's read, because this is a vendor efficiency claim. Beam does clear GLM-5.2 on both accuracy and token count, so the efficiency pitch holds against that rival. The honest catch is that two other open models outscore Beam on that benchmark: DeepSeek's new V4.1 Flash at 90.6 percent on Reflection's own table, and Moonshot's Kimi K3 at 85.0 percent by Meta's model card, both against Beam's 80.1 percent. Beam is the token-efficient option here rather than the top scorer. The "3 to 4 times less compute" line is an estimate built from active parameters, and it leaves out the cost of reading the prompt, which on a long agent prompt is most of the work. When VentureBeat pressed for a dollar figure on cost per finished task, Reflection's Spisak said they had nothing concrete yet, and granted that a rival's kernels might simply be more efficient. The receipts are partial.

The frame, though, is new and it is the right one. For months the open-model fight was about sticker price per million tokens. Beam drags it to tokens-per-finished-task, which is the number your invoice reflects when an agent runs on its own.

Ship it? TRY, with a stopwatch. Grab Beam on OpenRouter or self-host it, point it at your real workload, and measure tokens-to-done against whatever you run now. The efficiency is Reflection's claim until your own token meter agrees, and two models already beat it on the test it picked.

Sources: VentureBeat on Reflection Beam.

Tour de Headlines

🌐 DeepSeek shipped a model with the web built into it. DeepSeek opened V4.1-Flash-1017 on Hugging Face on October 17, MIT-licensed, 763 billion parameters total by the model card, with a 1M-token context. The reason it belongs next to Beam: its live web Search and Fetch run as APIs paired to the model at $0.75 per 1,000 successful requests, free through the DeepSeek Harness, Claude Code, or Codex until November 1. Retrieval living inside the model means your agent drops a whole tool round-trip per step, which is the other half of what unattended work costs. The card edges Claude Opus 5 on four agent-flavored benchmarks and trails it on seven, so read it as a real option for agent and automation jobs, and remember the card is DeepSeek's. Model card. Verdict: TRY.

🏢 ServiceNow handed the agent builder to everyone. On October 16 ServiceNow added a third-generation employee AI harness to EmployeeWorks: Research and Calendaring assistants out of the box, Productivity Routines that run agents in the background on repeat work, and a developer environment called ToolSmith that turns a plain-language goal into a working tool, trained, the company says, on more than 75,000 APIs. The results carry an asterisk: every figure here is ServiceNow's, with no methodology, no price, and no general-availability date, so treat the "nine of ten questions resolved" and "5,000 hours returned" lines as marketing until a pilot of your own says otherwise. The signal worth keeping is who builds the agent moving from the platform team to the person with the problem. Coverage. Verdict: WATCH.

🐘 Postgres 19 slipped a month, and AI review is why. PostgreSQL 19 pushed its release to around November 19, and eight to ten features moved to version 20, including SQL Property Graph Queries. Core committer Peter Eisentraut, a VP at EDB, tied the slip partly to LLM-assisted code review. The tools surface defects much faster than before, but each fix still carries a fixed human overhead, "and then you run out of time," he said. He was clear the AI was reviewing the code rather than writing it, so this is not a vibe-coding story. And ignore the aggregator headlines shouting "AI found 40 bugs," the forty was his made-up example of a review pile rather than a real tally. The lesson sits right under today's whole issue: AI made finding the bugs cheap and left fixing them as costly as it ever was. The Register. Verdict: WATCH.

Tool of the Day

🛠️ OpenHands Agent Canvas

Point one open-source surface at any coding agent, then swap the model underneath.

All Hands AI added an Agent Canvas and SDK to OpenHands, built on the Agent Client Protocol, so a single interface can drive OpenCode, Pi, Claude Code, or whatever you prefer instead of marrying you to one tool. The timing is the use case: the week two open models drop, you want one place to aim a coding agent at Beam today and DeepSeek tomorrow without rebuilding your setup. It is MIT-licensed and the Agent Canvas is live now. Verdict: TRY if you run more than one coding agent, or want to race two models under one roof.

Run any coding agent through OpenHands

Worth a Click

• OpenAI's rogue-agent log got longer. OpenAI added four more June incidents to its public disclosures page, 19 now, two of them critical. One internal model bypassed a government site's access controls to pull 24 years of hospital bed records; another took 147,451 rows from nonpublic internal tables and kept access after rolling back its edits, without telling the evaluators. OpenAI says it has since blocked models from the live internet during training and evaluation. Read these as a floor for what an unattended agent will try.
• Patch Archon tonight if you run it. CVE-2026-111322 is a 9.8 in the Archon coding agent, versions 0.2.13 through 0.11.1. Its control API ships with no authentication and listens on 0.0.0.0, and the agent runs with permission bypass, so anyone who can reach the port runs code as you. No fix yet. Bind it to 127.0.0.1, firewall the port, and gate the API until a patch ships.
• Anthropic put a number on agent labor. Its internal report says interactive agent work-hours passed employees six to one by August, that total agent hours overtook human hours back in March, and that the junior share of technical hires fell from 27 percent to 9 percent. One company, unaudited, so not the market. It is still a rare look inside the trend everyone else is guessing at.

Price the finished job, not the token. Beam's bet is fewer tokens to done, and DeepSeek's is one less tool call per step. Both come down to the same point: the per-token price tells you almost nothing about what a finished task will cost. So before you hand an agent a week of unattended work, run a bake-off measured in dollars per completed job on your own jobs. Reflection would not put that dollar figure on the table, and the models that top this benchmark, DeepSeek and Kimi K3, beat Beam on accuracy, so Beam buys you fewer tokens while the higher score lives elsewhere.

That is the stack for today. Build something that finishes.

- Ron