How I Split AI Work Between Opus, Sol, And Jev

🔍 Read the full analysis: How I Split AI Work Between Opus, Sol, And Jev on ThorstenMeyerAI.com

Prime Big Deal Days · Oct 6–7Offer from Amazon

Get hardware and tech essentials delivered free — and shop member deals

• Fast, free delivery on millions of items
• Access to Prime Big Deal Days deals on October 6–7
• Prime Video, Amazon Music and more included

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

In a report dated Sept. 29, 2026, Thorsten Meyer said he uses Claude Opus 5.5 for software building, GPT-6.1 Sol for detail work and review, and Jev for high-volume yes-or-no and routing decisions. The split reflects a cost-and-capability calculation based on Artificial Analysis benchmark data; it matters to teams deciding whether to use one model for every task or route work among several.

Meyer identifies Opus 5.5 at high effort as his main model for features, APIs, multi-file work and refactors. He gives it an Intelligence Index score of 54 and an estimated cost of $1.82 per task. For architecture, migrations and trust-boundary work, he uses xhigh: a score of 56 at $3.46 per task. At the max setting, Opus scores 58 and costs $5.98 per task.

He assigns GPT-6.1 Sol to close examination of files and code changes, and to review work produced by Opus. The report lists Sol at $0.32 per task on high and $0.39 on xhigh, with index scores of 50 and 51. Meyer says that price makes it practical to run a review on each meaningful change. He describes Jev as unable to write prose, using it instead for high-volume binary decisions and routing; the supplied material gives no benchmark score or per-task price for Jev.

The report lists six models and settings in its comparison: Opus 5.5, Sonnet 5.5, Fable 5.1, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. It says their top-setting scores span 37 to 58, while listed costs range from $0.07 to $7.63 per task. Meyer recommends Sonnet at high effort for scoped work such as documents and slides, and Luna for classification, extraction and routing. Astra and Fable are alternatives when task-specific tests favor them.

Why Meyer Adds a Review Model

The workflow treats model choice as a task-level cost decision. Meyer argues that paying for the highest available setting across all work can raise costs sharply for small score gains. In his figures, Opus rises from $1.82 per task at high effort to $5.98 at max, a 73% increase from xhigh to max for two additional index points. He also reports Sonnet 5.5 at max costing $7.60 per task for a score of 56, compared with Opus at 58 for $5.98.

A separate model reviewing generated work may catch issues the authoring model misses, and Meyer says the lower listed cost of Sol makes repeated reviews affordable in his workflow. That is his operational judgment, not evidence that a second model will reliably find defects. He cautions that review quality depends on the requirements and evidence provided: two models can share the same mistaken assumptions, and passing tests alone do not settle whether a change should ship.

The cost figures also put human review time in the calculation. Meyer says even a small saving on model charges can be erased by an extra minute of human review, but labels that example illustrative rather than measured. The practical outcome will depend on task complexity, correction rates and how much oversight a team needs.

The Benchmarks Behind the Split

Meyer says the scores come from the Artificial Analysis Intelligence Index v4.3.x, except where otherwise noted. He describes it as a map of general capability rather than a verdict on a reader’s workload, and advises shadow-testing before switching systems. The figures in the report are presented as estimates for defined benchmark tasks, not as a guarantee of production cost or output quality for a particular team.

The report’s effort-level table shows that settings change both scores and costs. For Opus 5.5, the listed score moves from 42 at low effort to 58 at max, while estimated task cost rises from $0.55 to $5.98. For Sonnet 5.5, the score moves from 36 to 56 and cost from $0.41 to $7.60. Meyer favors medium for everyday documents and high for Sonnet’s scoped work; he says max is rarely justified in his use.

For GPT-6.1 Sol, the report says Artificial Analysis listed medium, high and xhigh settings when it was published. Medium scored 48 at $0.21 per task, high scored 50 at $0.32 and xhigh scored 51 at $0.39. Meyer compares Sol xhigh with Astra and Fable, whose listed scores are one or two points higher, while their stated task costs are substantially greater. He notes that a one-point gap may fall within measurement noise.

Limits of the Cost Comparison

The supplied report does not include the benchmark methodology, uncertainty ranges or enough detail to independently reproduce its per-task cost estimates. It does not define Jev’s provider, price, evaluation results or routing accuracy. Its figures should therefore be read as the report’s account of one benchmark and workflow, not as independently verified results for every use case.

Some measurements may also change as model settings and benchmark listings are updated. Meyer says the index had not published low or max results for GPT-6.1 Sol at the time of writing. He reports first-token delays of 57 seconds at high and 69 seconds at xhigh, which may limit interactive use, but the material does not describe the full testing conditions or how those delays vary across workloads.

The report does not provide a controlled comparison of the complete workflow against a single-model setup. It remains unclear how often Sol catches errors, how many Opus outputs need rework, or whether the savings in model charges outweigh added review and routing overhead for other teams.

Test the Workflow on Real Tasks

Meyer’s stated next step for readers is to shadow-test models on their own work before changing defaults. A team could compare outputs against its existing process, track error rates and review time, and record total cost per completed task rather than relying on index scores alone. Those results would help determine whether Opus, Sol, Jev or the listed alternatives fit particular tasks.

The report does not announce a formal follow-up test or a date for updated results. Further detail on Jev’s performance, the benchmark’s assumptions and the treatment of human review would help readers assess the full workflow. Until those details are available, the model assignments remain Meyer’s reported practice and a proposal to evaluate against local requirements.

Key Questions

What jobs does Meyer assign to Opus 5.5?

He uses Opus 5.5 at high effort for features, APIs, multi-file changes and refactors. He reserves xhigh for harder work such as architecture and migrations.

Why does he use GPT-6.1 Sol for review?

The report lists Sol at $0.32 per task on high and $0.39 on xhigh. Meyer says that cost makes a separate review pass practical in his workflow; the report does not show that Sol will catch every error.

What does Jev do in the workflow?

Meyer describes Jev as a decision model that cannot write a sentence. He uses it for high-volume yes-or-no decisions and routing, but the supplied report does not provide its benchmark score or price.

Do the benchmark scores show which model is best for every team?

No. Meyer says the Artificial Analysis Intelligence Index measures general capability and is not a verdict on a particular workload. He advises testing models against a team’s own tasks before making a switch.

Source: ThorstenMeyerAI.com