# A machine that runs all night, and why the labs aren't at the forefront (slim) > Vlad Cealicu on why the big labs might have the latest models but aren't at the forefront of what's being done with agents. The harness moves results more than the model, tokenmaxxing led to nothing useful, loops are brute force, features come from customers and from testing tools, ASI comes from reasoning plus tools, and agent tooling is one of the biggest bottlenecks, which is why Anchor Terminal exists. - Full: https://www.anchorterminal.com/blog/a-machine-that-runs-all-night.md (~3,500 tokens) · this version ~1,480 tokens · JSON https://www.anchorterminal.com/blog/a-machine-that-runs-all-night.json · canonical https://www.anchorterminal.com/blog/a-machine-that-runs-all-night - Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-07 - Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data) By Vlad Cealicu, 7 November 2026. 1,333 words, about 6 min to read. - 7.8x: how much more a coding benchmark score moved with the harness than with the model, in one controlled study - 60 trillion: tokens used in 30 days by staff on Meta's internal leaderboard, which came down two days after the news broke - 4 months: for Uber to burn through its whole 2026 token budget The model, the harness and the tokens. The big labs, Anthropic, OpenAI and the others, might have the latest models that aren't available to us, and METR's report this week says Anthropic staff put their own speed-up from the most capable internal model at about 12 times in informal polls [[1]](#ref-1). ## The harness moves the result more than the model Zhang and colleagues tested this in May 2026 and found that on coding benchmarks "performance variance is governed more by harness configuration than by model choice" [[2]](#ref-2). ## The year of the leaderboard Big companies, OpenAI among them, spent the spring trying the opposite, ranking engineers by tokens spent on formal or informal leaderboards [[4]](#ref-4). By the end of May Fortune was calling tokenmaxxing over, and it lists Uber burning through its whole 2026 token budget in the first four months of the year, the Financial Times reporting Amazon employees running agents on pointless tasks to keep their numbers up for performance reviews, and Microsoft cancelling Claude Code for several product divisions, as The Verge reported [[4]](#ref-4). From the leaderboards to the pullback 1. **January to April** (hype). Uber spends its whole token budget for 2026 [[4]](#ref-4) 2. **April** (hype). Meta's token leaderboard, with its Token Legends, is reported and comes down two days later [[5]](#ref-5) 3. **28 May** (pullback). Fortune calls tokenmaxxing over, and Microsoft has cancelled Claude Code for several divisions [[4]](#ref-4) 4. **August** (pullback). Microsoft tells engineers tokenmaxxing isn't the goal, and Amazon retires KiroRank [[6]](#ref-6) In the survey for LeadDev's AI Impact Report, 58% of respondents said their organisation measures AI impact by token usage while 57% said token usage is a poor way to gauge real value, and only 19% rated tokenmaxxing as effective [[6]](#ref-6). Respondents, in per cent | Answer | Per cent | | --- | --- | | Their organisation measures AI impact by tokens | 58 | | Token usage doesn't gauge real value | 57 | | Tokenmaxxing is effective | 19 | Each series has a zero-based scale. ## Loops are brute force You can put a coding agent in a loop now, the same prompt fed back in every time it stops, and leave it running all night, and Anthropic turned that into an official Claude Code plugin in December 2025 [[7]](#ref-7). ## Where the features come from If you're building for people, the features they need are found by researching your market and talking to potential customers. If you're building for agents, the features they need are found by testing the products and making requests cost efficient. As far as I can tell, tokenmaxxing led to nothing useful. ## Where ASI comes from I think we're intelligent as a species because we started using tools, and our opposable thumb helps. I think ASI won't be one of the big labs, it will be a group of people who figure out how to use the reasoning part and tool usage. ## Tooling is the bottleneck, and why we exist One of the most important bottlenecks is agent tooling, the APIs, MCP servers and services an agent has to find, sign up for, pay for and call, and that's why we exist. We need more small, focused products, not big encompassing ones. --- More in [Opinion](https://www.anchorterminal.com/blog/opinion/index.md) · Older opinion [A machine that finds the bug in hours](https://www.anchorterminal.com/blog/a-machine-that-finds-the-bug-in-hours.md) - Older: [What an agent asks before it picks a tool, and what a site should answer](https://www.anchorterminal.com/blog/what-an-agent-asks-before-it-picks-a-tool.md) - Newer: [A machine that builds the next machine, and a report that landed before the roadshow](https://www.anchorterminal.com/blog/a-machine-that-builds-the-next-machine.md) - All posts: https://www.anchorterminal.com/blog/index.md Anchor Terminal grades the tools AI agents run on, with a source for every score. ## References Read on 7 November 2026. Quotes are as published. [1] METR, 5 November 2026. Independent assessment of AI R&D acceleration at Anthropic. [2] Zhang, Y., Wang, J., Ge, Y., Xu, W., Hamm, J. [3] Anthropic, 24 November 2025. Introducing Claude Opus 4.5. Source for Opus 4.5 being available through the Claude API from that day. [4] Fortune, Jeremy Kahn, 28 May 2026. Tokenmaxxing is over. [5] Fortune, 9 April 2026. Meta killed its employee AI token dashboard. [6] LeadDev, Chantal Kapani, 17 August 2026. The tokenmaxxing hype didn't last long. [7] DreamHost, 2026. The Ralph Wiggum loop, from first principles. [8] Anchor Terminal measurements of its own MCP server against production data, 6 November 2026, not published. [9] Anchor Terminal, 2026. The Railway MCP server listing, with the 71 tools its hosted endpoint lists. [10] Stout, D. [11] Tanaka, K. [12] Sutton, R., 13 March 2019. The bitter lesson. [13] Anchor Terminal research notes, 6 November 2026, not published. [14] Anchor Terminal, 2026. Resend's listing, read on 6 November 2026. Source for the 136 tools its hosted MCP server lists and the 106 that seven of its panel reviews, written on 3 October 2026, report.