How Opus, Sol, And Jev Collaborate In My AI Stack

🔍 Read the full analysis: How Opus, Sol, And Jev Collaborate In My AI Stack on ThorstenMeyerAI.com

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

Get the latest gadgets 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

Thorsten Meyer said Sept. 29 that his AI stack uses Opus 5.5 as its main model for building, GPT-6.1 Sol for detailed review and Jev for high-volume yes-or-no and routing decisions. The arrangement reflects a shift in his model selection criteria from benchmark rank alone toward cost per task, though the published figures come from a general capability index and do not establish how the models perform on every workload.

Meyer’s account draws on the Artificial Analysis Intelligence Index v4.3.x for model scores and task-cost estimates. In that index, Opus 5.5 scored 58 at its top setting and was estimated at $5.98 per task. GPT-6.1 Sol scored 51 at xhigh and $0.39 per task. The source lists Luna at $0.07 per task and 37 points, while GPT-6 Astra scored 53 at its top setting for $3.26 per task. These are index figures, not results from Meyer’s own controlled comparison across all jobs.

The described division of work assigns Opus 5.5 at high effort to features, APIs, multi-file changes and refactoring. Meyer reports an index score of 54 and task cost of $1.82 at that setting. He uses xhigh, scored at 56 and $3.46 per task, for architecture, migrations and trust boundaries. GPT-6.1 Sol at high or xhigh handles detailed inspection of files and code changes and provides a separate review pass.

Meyer also lists Jev as a decision model that cannot write a sentence, assigning it high-volume binary decisions and routing judgments. The source gives no Jev benchmark score, price, release date or measurement method. Sonnet 5.5, Luna, Astra and Fable are described as alternatives for narrower tasks, with Astra or Fable reserved for cases where Sol and Opus disagree.

Cost Shapes Each Model’s Role

The account illustrates how reported differences in cost per task can affect where models are used. Meyer says Sol’s xhigh index estimate is $0.39 per task, compared with $3.26 for Astra and $7.63 for Fable 5.1. If those estimates translate to his work, running a second model as a routine review step becomes more affordable. The source does not report an audited total for the stack or evidence that this process catches more defects.

The division also assigns separate models to production and review. Meyer argues that another model family can provide a more useful second perspective on Opus’s work than asking Opus to assess its own output. That is a description of his practice, not proof of independent verification: a reviewer can still share assumptions or miss problems, especially if both models receive the same incomplete requirements.

For readers choosing models, the account makes a practical point: benchmark position alone may not determine fit. Latency, effort setting and human review time also affect whether a cheaper task estimate leads to a lower overall cost. Meyer cautions that one extra minute of human review can erase a saving from cheaper model tokens; he labels the example illustrative rather than measured.

Benchmark Scores Meet Task Costs

Meyer frames his Sept. 29 report around what he calls a changing AI model market: several models are close on a general capability index, while their estimated task costs differ substantially. He says six models sit within about 20 index points of one another, but that comparison depends on the models and settings included in his table. The index is a broad capability measure, and Meyer advises shadow-testing models on a target workload before switching.

GPT-6.1 Sol was released on Sept. 29, according to the source, at the same listed token prices as its predecessor: $2 per million input tokens and $10 per million output tokens. The index lists medium, high and xhigh settings for Sol, with estimated costs of $0.21, $0.32 and $0.39 per task. At high and xhigh, the source reports first-token times of 57 and 69 seconds, respectively, which may make those settings unsuitable for interactive use.

Effort settings also change the reported scores and costs. For Opus 5.5, Meyer lists $1.82 per task at high and $5.98 at max, alongside scores of 54 and 58. For Sonnet 5.5, he lists $2.74 at xhigh and $7.60 at max, with scores of 52 and 56. The source says index measurements show Sonnet producing about 193,000 output tokens per task at max. These estimates describe the index’s evaluation setup; actual usage and costs can vary by task.

““which model clears my quality bar at the lowest cost per task?””

— Thorsten Meyer

Jev’s Performance Is Not Reported

The source does not provide Jev’s cost, accuracy, benchmark results or evaluation method, so its role in the stack cannot be compared quantitatively with Opus 5.5 or GPT-6.1 Sol. It also does not say which Jev version is in use, how the yes-or-no judgments are checked, or how routing errors are handled.

The reported index scores and per-task estimates are not evidence of performance on Meyer’s own projects. The source says Sol’s high and xhigh settings take 57 to 69 seconds to produce a first token, but does not give response times across typical real-world jobs. Meyer also notes that the index has not published Sol’s low or max settings and that a one-point score difference falls within noise.

It remains unclear whether routine Sol review reduces defects, how often reviewers disagree, and what the total cost is after human checking. Meyer says models should be shadow-tested before a switch, but the article does not publish results from such tests or define a quality threshold for each type of work.

Workload Tests Will Guide Changes

Meyer’s stated next step is to compare candidate models against real tasks through shadow testing before changing defaults. That would allow teams to check output quality, latency and cost against their own requirements rather than relying only on a general index. The source does not give a timetable or announce a formal test program.

For the described stack, the next decision points are whether Sol’s review catches useful issues at its reported cost and whether Jev’s routing judgments meet the required accuracy at volume. The account does not say when new Jev data, further Sol index settings or results from workload tests will be available.

Key Questions

What roles does Meyer assign to Opus 5.5, Sol and Jev?

He uses Opus 5.5 for building, GPT-6.1 Sol for detailed review and Jev for high-volume yes-or-no and routing judgments.

Why does Meyer use Sol for review?

Meyer says Sol’s reported cost per task makes a second-model review affordable to run routinely. The source does not provide measured evidence that the review catches more defects.

How much does Jev cost or score?

The source provides no cost or benchmark score for Jev. It describes Jev as a decision model that cannot write sentences but gives no evaluation details.

Are the reported costs guaranteed for other users?

No. They are estimates from the Artificial Analysis Intelligence Index v4.3.x evaluation. Actual costs and results can depend on task, settings and usage.

What should teams do before switching models?

Meyer recommends shadow-testing candidate models on their own workloads. This can show whether quality, latency and cost meet the team’s requirements.

Source: ThorstenMeyerAI.com