Vector Lab
VECTOR LAB

EST. 2025

WEEKLY UPDATE2026
BY ANDREW MEAD

GPT 5.6

Is GPT 5.6 better than Claude Fable? Cursor helps Grok make a comeback, and Meta takes another swing

Read in

Releases

GPT 5.6

GPT 5.6 has been cleared by the US government and is now available to everyone. The model comes in 3 sizes, named Luna, Terra, and Sol (in order from smallest to largest).

ArtificialAnlys benchmarks

Benchmarks from Artificial Analysis, since OpenAI did not give a good benchmark image to use.

Based on my usage this week and also the public sentiment I would put GPT 5.6 Sol in the same tier as Claude Fable 5. It is extremely capable on agentic tasks, coding and otherwise. It also has the added benefit of being one third the cost, due to its lower cost and lower token us versus Fable.

Fable does still seem to have the best intent inference and judgement (big model smell) but for most tasks this is not necessary. If you really need the extra juice from Fable’s planning, I would use Fable as the planner and have it use GPT 5.6 as the implementer. GPT 5.6 also doesn’t seem to have as many or as oppressive safeguards compared to Fable.

Historically the GPT models have been poor at frontend design as evidenced by Design Arena and Frontend Code Arena. GPT 5.6 changes this, as it is now tied for first with Fable in the Frontend Code Arena. Previously I had only used Claude for frontend tasks, as GPT 5.x was better at software engineering in my experience, and now I no longer need to go to Claude for anything now for day to day coding.

AA coding scores per tier

From rasbt on Twitter

We need to talk about reasoning effort levels and the smaller Terra and Luna models.

There are 6 reasoning levels to choose from: light, medium, high, extra high, max, and ultra. The first 5 use increasing amounts of reasoning effort to get better results, and the ultra tier uses a large number of subagents to work in parallel.

Normally for the GPT series I would just say to use the best model for all tasks and not worry about the smaller models. From benchmarks, and some initial community testing, this doesn’t seem to be the case anymore.

For instance, looking at the AA Coding Agent index shown above, GPT 5.6 Terra Ultra should be cheaper and better than using Sol with high reasoning. This of course is on benchmarks, which may not be indicative of real world performance. Members of the community are still testing to see what combination of model and reasoning level is the best for day to day use. We will be covering any advances or learnings about this at AI Tools Club, our weekly community meeting about how and what AI tools we are using (livestreamed on Tuesdays at 5PM on the Vector Lab YouTube channel).

For now the makers of Codex recommend using Sol with medium reasoning, so I would use that as the default for now. Also note that the Codex rate limits get burned faster with 5.6 Sol than with 5.5, so be aware of that when using it.

Model$ per million (input)$ per million (output)Tokens per second
GPT 5.6 Sol$5$3045
GPT 5.6 Terra$2.50$1563
GPT 5.6 Luna$1$684
GPT 5.5$5$3051
Claude Fable 5$10$5040
Claude Opus 4.8$5$2560
GLM 5.2$1.40$4.4047
Grok 4.5$2$688
Muse Spark 1.1$1.25$4.25no data
Data from OpenRouter, except for Muse Spark, which comes from AA

Grok 4.5 Fast

xAI has released their first model as a part of SpaceX, and it is a real surprise. The previous Grok models were all rather lackluster and saw little usage outside of @Grok on Twitter.

Their new Grok 4.5 model seems actually competitive with the Chinese frontier (GLM 5.2) and the American mid tier (Sonnet 5). This is probably due to the new collaboration that they have with Cursor team, since Cursor was bought by SpaceX shortly after they IPO’d.

Grok 4.5 coding benchmarks

For real world use it seems very usable, similar to GLM 5.2. The most notable thing about it is its speed, which coupled with its low amount of reasoning that it uses allows it to feel much faster than many of the frontier models.

This low token usage also makes it so that it’s around the same cost as GLM 5.2 with max reasoning, even though its nominal price is ~1.5x more.

Overall, an unexpectedly solid release from xAI, we will see if they will be able to continue this momentum into the future and become a contender as one of the top US labs.

Quick Hits

Meta Muse Spark 1.1

Meta also released a model this week, but unfortunately for them it was unable to outshine GPT 5.6 or Grok 4.5.

Muse Spark 1.1 seems to sit around the same level of usability as Grok 4.5 and GLM 5.2, but is the worst of the 3, being about 5-10% worse on most benchmarks. The one thing it has going for it is that it does seem to be 25-50% cheaper than the other two.

It does seem to be good in some random fields like legal and medical administrative work, but my guess this is overfitting to the benchmarks.

Finish

I hope you enjoyed the news this week. If you want to get the news every week, be sure to join our mailing list below.

Vibes of the day
Early computational art from Joseph Alessio on Twitter

Stay Updated

Subscribe to get the latest AI news in your inbox every week!

← BACK TO NEWS