A Wednesday with a new benchmark, a public contract with a discount and several texts about who answers for what agents write.

Every PM who has ever defended a model choice in a meeting knows the question: which one is the best? Yesterday a benchmark came out that made me think the question is badly framed.

Artificial Analysis launched version 5 of its Intelligence Index. About 60% of the tests are now private, to make it harder for models to be tuned for the leaderboard. Claude Opus 5.5 leads with 58 points, and GPT-6 Astra is right behind with 57. No model wins all nine evaluations.

The math per task

What caught my attention most was the cost. On average, each task in the index costs US$ 10.04 with Opus 5.5 at maximum effort. GPT-6.1 Sol sits 4 points lower, at US$ 1.27. That is roughly one eighth of the price.

And then there is reasoning effort. The same Opus 5.5 scores 35 points at low effort, at US$ 0.98, and 58 at maximum. In other words, the model is one decision, but its configuration is another, and it changes the math.

What changes for the product

On the product side, this changes the question. Instead of looking for the best model, I want to know which is the cheapest one that delivers the quality that feature requires. Some tasks call for the top, and plenty do not. Measuring that with our own cases, and not just with the leaderboard, becomes part of the job for whoever looks after the product. The guide on AI agent cost management and the one on how to evaluate AI agents help structure that measurement, within AI product management.

I think it is great that the market is showing cost next to quality. The more transparent the benchmark, the easier it is to choose well and the more room is left to build.

For those who got curious and want the full breakdown, here is the reading.

The rest of the radar

  • Claude in drug discovery — at Enveda, the review of 1,061 pages dropped 89% (from more than 25 h to under 3 h) and a regulatory document was done in 4 hours instead of about 3 months; still early work and mostly preclinical, with data from Anthropic itself. Anthropic.
  • Anthropic and GSA — 40% off Claude Enterprise, Claude for Government and the API for the US government, with no minimum spend and usage-only billing; a good example of public-sector pricing. Anthropic.
  • Genie 4 — Google DeepMind’s world model generates interactive environments with sound and dialogue; it is still a preview for testers, with no pricing, and coherence drops in long sessions. DeepMind.
  • LFM3 — six models of about 2B parameters with open weights; the text one fits in about 1 GB at 4 bits, but the speed numbers come from the vendor and the license is not on the page. Liquid AI.
  • Here be monsters — when agents write the code, whoever answers for it in an incident needs to be defined and to have the map of the system. Blog.
  • The Mythical Man-Month in the age of AI — agents do not eliminate the cost of coordination: with vague requirements, you reach the wrong answer faster. j11g.
  • AI content cleanup on Wikipedia — about 20% of new articles were flagged in late 2025 and 15–16% in early 2026, with the caveat that the drop may be a detector limit. IEEE Spectrum.

Tomorrow there is more, and the idea stays the same: separate what changes the product decision from what only makes noise.