DVx Signal/Builder/Issue 007

GPT-6 Sol, Luna, Opus 5.5 and Jev: what to test

New models offer lower API costs and typed decisions. Compare completed tasks before switching.

DVx Research2 min read

Sol and Luna give builders cheaper GPT-6 options

September 22 launch · checked September 26Official source ↗
API modelInput / 1MOutput / 1M
GPT-6 Sol$2$10
GPT-6 Luna$0.10$0.50
Standard API rates for requests at or below 272K input tokens. Cached input is cheaper; tools, retries and taxes are extra. View official source ↗
What changed

OpenAI launched both models September 22. Sol targets complex coding and agent work; Luna targets focused, high-volume tasks. Both accept text and images and return text.

Why it matters

For 1 million uncached input and 200,000 output tokens spread across short requests, Sol costs $4 and Luna $0.20 at standard rates. That is a price comparison, not a quality ranking.

What to do

Replay a few accepted tasks on both model IDs. Compare answer quality and total usage before routing volume to Luna.

Opus 5.5 costs less per token than Opus 5

September 22 launch · checked September 26Official source ↗
Per 1M tokensOpus 5.5Opus 5
Input$4$5
Output$20$25
Cache read$0.20$0.50
Anthropic’s published API rates. Fast mode costs $8 input and $40 output per million. View official source ↗
What changed

Claude Opus 5.5 is the first 5.5 model. Its input and output token rates are 20% below Opus 5; cache reads are 60% lower.

Why it matters

Anthropic reports roughly 40% lower cost on its typical tasks because the model also uses fewer tokens. Treat that as a vendor result; your cost per completed task may differ.

What to do

Run the same coding or research task on Opus 5 and 5.5. Record accepted result, tokens, cache use and retries.

Jev returns decisions your code can use directly

September 15 launch · checked September 26Official source ↗
QuestionReturnsUseful for
ChoiceOption + probabilitiesRouting
ScoreNumber + confidencePrioritizing
Noul0–1 valueYes/no signals
TypeSafe’s documented output types. A valid type does not guarantee a correct judgment. View official source ↗
What changed

Jev takes state and typed questions, then returns structured values rather than prose. It launched in early access. TypeSafe lists $0.042 per million input tokens; output is unmetered.

Why it matters

This is a narrower tool for classification, scoring and routing. TypeSafe’s speed and accuracy comparisons are its own tests, not independent proof.

What to do

Try it on labeled support tickets. Measure wrong routes, low-confidence cases and total cost before allowing automatic action.

Also since the last issue

Useful updates

Copilot local sandboxing

GitHub’s September 23 public preview limits local agents’ file, network and credential access by project. It is off by default and applies to new sessions.

More models in Copilot

GitHub’s September 25 roundup lists Sol, Luna and Opus 5.5. Availability varies by Copilot plan; check the picker and usage terms.

Gemini speech API

Google’s September 23 release adds Flash and Flash-Lite TTS. Current API rates are $0.50 text input and $9 or $6 audio output per million tokens through December; both double January 1. The AI at Work edition has a sample.