Subscribe

SignalIndustrialsData story

Exowatt says its fine-tuned market model beat Claude Opus 4.7 at $0.60 per 1,000 answers vs $29.26

On Exowatt's own test of 2,000 energy-market cases, a fine-tuned Qwen3-14B scored 0.975 macro-F1 to 0.625 for Claude Opus 4.7.

Tom Bedwell, founder of 50 Foot Media. Tom Bedwell 3 min read

Key figures

  • $0.60 vs $29.26 Cost per 1,000 answers, energy-market model vs Claude Opus 4.7, about 49x lower. List prices times measured tokens. (AWS Machine Learning Blog, 27 Oct 2026)
  • 0.975 vs 0.625 macro-F1 Same comparison, 2,000 held-out test cases. Claude Sonnet 4.6 scored 0.538. (AWS Machine Learning Blog, 27 Oct 2026)

An AWS blog post, co-written by AWS staff and two Exowatt engineers. Exowatt's own test sets and list-price cost estimates, not independently verified.

Exowatt's solar and heat-storage fleet is growing toward thousands of field units. In an AWS post published October 27, 2026, co-written with Exowatt's Tim Child and Felipe Arce, the team says sending every routine decision to a frontier model "across thousands of sites is not economical."

Top-tier rates make little sense for one narrow job that repeats continuously. A trained specialist can cost less and do it better. Exowatt fine-tuned Qwen3-14B on five years of public ERCOT data to flag one of five market events from six hours of live data. On 2,000 held-out cases it scored 0.975 macro-F1, against 0.625 for Claude Opus 4.7 and 0.538 for Claude Sonnet 4.6, at $0.60 per 1,000 answers to $29.26. The post calls that "about 49x lower cost per 1,000 answers," and its headline claim of up to 49x rests on this comparison alone.

Bar chart of cost per 1,000 answers: $0.60 for Exowatt's fine-tuned energy-market model against $29.26 for Claude Opus 4.7, on Exowatt's own 2,000-case test set at list prices. Source: AWS Machine Learning Blog, 27 Oct 2026.
AIVI chart of figures from the AWS Machine Learning Blog, 27 Oct 2026. Exowatt's own held-out test set, with costs at Amazon Bedrock list prices; not independently verified.

A second model handles alarm triage. Before training, the base Qwen3-1.7B "couldn't produce a valid answer." After about 32 minutes of training, at roughly $20 to $25, it scored 0.911 macro-F1 on 320 held-out cases, ahead of all three Claude models tested. It is built to run offline on the site's existing CPU server.

For a small team, speed matters too. The market model went from initial plan to champion in six days, across 23 training jobs. Separately, the post says Exowatt moved its customized model "from experimentation to production in only nine days."

Exowatt plans to reuse the approach to build "smaller, task-specific models across the fleet."

Get new AI Value Index records in your inbox.

Subscribe

You might also like...