What Opus Builds, What Sol Digs Into, And How Jev Decides

🔍 Read the full analysis: What Opus Builds, What Sol Digs Into, And How Jev Decides on ThorstenMeyerAI.com

Age 18–24?Offer from Amazon

Prime made for students and young adults

• Fast, free delivery for dorm and study essentials
• Prime Video and Amazon Music included
• Member-only deals

Try Prime for Young Adults →Free trial for eligible 18–24 year olds

As an affiliate, we earn on qualifying purchases.Hide

Thorsten Meyer says he now uses Claude Opus 5.5 to build, GPT-6.1 Sol to inspect details and review work, and Jev for high-volume yes-or-no decisions, in a model stack described Sept. 29. His account argues that differences in task cost may matter more to choosing a model than small gaps in benchmark scores, while the index he cites measures general capability rather than performance on a reader’s specific work.

Meyer bases most of his comparison on the Artificial Analysis Intelligence Index v4.3.x. In the figures he reports, Claude Opus 5.5 scores 58 at its top setting and costs $5.98 per task. GPT-6.1 Sol scores 51 at xhigh and costs $0.39 per task. The table also lists GPT-6 Astra at 53 and $3.26 per task, Claude Fable 5.1 at 53 and $7.63, Claude Sonnet 5.5 at 56 and $7.60, and GPT-6 Luna at 37 and $0.07. These are the source’s reported index and task-cost figures, not independently verified results.

Meyer says he uses Opus 5.5 at high for general development, including features, APIs, multi-file work and refactors. He reserves xhigh for harder tasks such as architecture, migrations and trust boundaries. At those settings, the reported index scores are 54 and 56, with task costs of $1.82 and $3.46. He assigns Sol at high or xhigh to focused investigation and review, citing reported costs of $0.32 and $0.39 per task.

The article also compares effort settings. Meyer reports that Opus at max costs $5.98 per task for a score of 58, while xhigh costs $3.46 for 56. For Sonnet 5.5, he reports $7.60 at max for a score of 56, compared with $2.74 at xhigh for 52. He says he uses Sonnet at high for scoped subtasks, documents and slides, and Luna for routine checks and bulk classification. He names Astra and Fable as alternatives for cases where his own tests show an advantage.

Why Task Costs Shape His Stack

The account makes cost per task, rather than token price alone, the central operational comparison. In Meyer’s figures, Sol at xhigh is much less expensive per task than Astra or Fable, while its index score is one or two points lower. If those figures apply to a particular workflow, running a second model to inspect important changes could be affordable enough to do routinely.

Meyer’s case for a separate review model is also about using a different model family as a second check. He says Sol reviews Opus’s output because a model checking its own work may miss the same problems. That review does not settle whether a change is safe or correct: he cautions that passing tests alone is not approval to ship and says the reviewer needs the requirements and evidence for a failure.

The comparison has limits for teams weighing a switch. Index scores summarize performance across the index’s tasks, and the article itself says readers should shadow-test models on their own work. Meyer also says one index point is within the noise. Costs and scores in the article can inform a trial, but they do not establish the best model for a given team, task or quality threshold.

The Model Releases Behind the Comparison

The source describes a cluster of six models whose top-setting index scores fall within about 21 points, from Luna at 37 to Opus 5.5 at 58. Meyer says their task costs differ by roughly 100 times. That range frames his question as which model meets a task’s quality requirements at an acceptable cost, rather than which model leads a single ranking.

The release dates in Meyer’s table run from Fable 5.1 on Sept. 1 to Opus 5.5 on Sept. 22, Luna on Sept. 22, Sonnet 5.5 on Sept. 28 and GPT-6.1 Sol on Sept. 29. Astra is listed as released Sept. 3. The timing matters because the comparison was written as Sol launched, and the available index results for it were still limited.

Meyer reports token prices separately from cost per task: Opus at $4 per million input tokens and $20 per million output tokens, Sol at $2 and $10, and Luna at $0.10 and $0.50. He lists Fable and Astra at $10 and $50. Those rates do not directly describe the total cost of completing a task, which depends on usage; the article’s per-task figures are the comparison he uses for choosing roles.

“Opus 5.5 at high or xhigh is my main model for building.”

— Thorsten Meyer, in the Sept. 29 article

What the Index Cannot Establish

The reported scores do not show how the models perform on every real-world workload. Meyer calls the index a map of general capability and recommends shadow-testing before switching. The source does not provide details of his own testing methods, sample sizes or uncertainty ranges for the task-cost measurements.

Some Sol settings were not yet listed in the index when Meyer wrote: he says low and max results had not been published. He also cautions that a one-point score difference is within the noise. High and xhigh Sol reportedly took 57 and 69 seconds to produce a first token, which may affect interactive use; the article does not say how those timings vary across conditions.

The source introduces Jev as a decision model that cannot write sentences and assigns it high-volume yes-or-no and routing judgments. It does not give Jev’s scores, cost per decision, evaluation method or error rate. The supplied text also begins a cost example about model pricing and human review but ends mid-sentence, so its figures and conclusion cannot be established from the material provided.

Testing the Roles on Real Work

Meyer recommends shadow-testing candidate models against existing work before making a switch. For a team applying his framework, the next step would be to compare outputs against its own quality requirements and measure the full cost of completing tasks, including review time. The article does not give a schedule for further index updates or say when more Sol effort settings will be published.

His stated process sends a failed review back to Opus with the failing case and supporting evidence. That makes the quality of the task specification and review evidence part of the next decision, alongside model choice and effort setting. Meyer says the final decision to ship should not rest on passing tests alone.

Key Questions

What work does Meyer assign to Opus 5.5?

He uses Opus 5.5 at high for general development such as features, APIs and refactors, and xhigh for more demanding work such as architecture and migrations.

Why does he use GPT-6.1 Sol for review?

Meyer says Sol provides a second check from a different model family at a reported $0.32 to $0.39 per task at high or xhigh. Those figures come from his source article and are not a guarantee of cost on other workloads.

Does the index prove which model is best?

No. Meyer says the Artificial Analysis Intelligence Index measures general capability, not performance on every user’s workload, and recommends shadow-testing before switching.

What is Jev used for?

The source describes Jev as a decision model for high-volume yes-or-no and routing judgments that cannot write sentences. It provides no benchmark scores, error rates or cost figures for Jev.

Source: ThorstenMeyerAI.com