How I Assign AI Tasks Across Opus, Sol, And Jev

🔍 Read the full analysis: How I Assign AI Tasks Across Opus, Sol, And Jev on ThorstenMeyerAI.com

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

Get smart everyday buys 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 says he assigns software building to Claude Opus 5.5, detailed investigation and review to GPT-6.1 Sol, and high-volume yes-or-no decisions to Jev, a decision model that cannot write sentences. In his Sept. 29 account, Meyer argues that per-task costs vary widely among models with relatively close benchmark scores, making the choice of model and effort setting a practical cost decision for teams using AI.

In the comparison Meyer published Sept. 29, he cites the Artificial Analysis Intelligence Index v4.3.x for general capability scores and task-cost estimates. The index’s listed top settings put Opus 5.5 at 58 points and an estimated $5.98 per task, and GPT-6.1 Sol at xhigh at 51 points and $0.39. These are index results, not measurements of Meyer’s production workload. Meyer cautions that the index should not be treated as a verdict on any particular use case.

Meyer describes a division of work based on those trade-offs. He says he uses Opus 5.5 at high for features, APIs, multi-file work and refactors, and xhigh for harder work such as architecture, migrations and trust boundaries. He assigns Sol at high or xhigh to focused file or code-diff investigations and independent review. He says Astra or Fable provide second opinions when Sol and Opus disagree; Sonnet handles scoped subtasks, while Luna is used for classification, extraction and routing.

The Artificial Analysis estimates change with the effort setting. Its index lists Opus at 54 points and $1.82 per task at high, compared with 58 points and $5.98 at max. Meyer says he rarely uses max. He identifies Sonnet’s high setting, at 47 points and $1.08, as its best value among the listed results; its max setting costs $7.60 for a score of 56. These are index estimates, not a guarantee of equivalent quality or savings in other work.

Why Task Costs Shape Model Choice

Meyer’s approach shifts model selection from a single ranking to a cost-per-task decision. If a cheaper model meets a team’s quality bar for routine classification or review, using the most expensive available model for every step could add cost without a matching benefit. Conversely, a higher-scoring model may be justified for complex work where errors or rework carry greater consequences.

Meyer also argues that an inexpensive review pass can make a different model family practical as a second set of eyes. He says another model may catch issues the builder misses. That is his rationale for review, not proof that a model’s review is independent or sufficient: the reviewer can share assumptions embedded in the same brief, and Meyer says human judgment remains part of his process.

Meyer cautions that token prices alone do not measure the full cost of work. He says human review time can outweigh a model-price saving, though his example is explicitly illustrative rather than measured. The figures therefore frame a purchasing and workflow question; they do not show realized savings for a particular company.

The Scores Behind Meyer’s Stack

According to the figures Meyer attributes to the Artificial Analysis Intelligence Index, six models sit within about 20 index points of one another while their estimated task costs differ by roughly 100 times. The table he cites lists Opus 5.5, released Sept. 22, with a score of 58 and cost of $5.98 per task; Sonnet 5.5, released Sept. 28, at 56 and $7.60; and Fable 5.1, released Sept. 1, at 53 and $7.63. It lists GPT-6 Astra at 53 and $3.26, GPT-6.1 Sol at 51 and $0.39, and GPT-6 Luna at 37 and $0.07.

Meyer says GPT-6.1 Sol launched Sept. 29 at the same listed token prices as its predecessor: $2 per million input tokens and $10 per million output tokens. The Artificial Analysis index lists medium, high and xhigh settings for the model. Meyer reports their respective scores as 48, 50 and 51, and estimated task costs as $0.21, $0.32 and $0.39. He says the index has not yet published Sol’s low or max settings.

Meyer reports that, in the index results, moving Opus from xhigh to max adds two points while raising estimated task cost by 73%. From medium to max, he reports a 4.46-fold cost increase for seven points. Those comparisons apply to the specific index results and settings he cites; benchmark points alone do not establish how much better a model will perform on a given assignment.

What the Benchmark Cannot Settle

The comparison Meyer published does not show how the models perform on his own tasks or another organization’s workload. He recommends shadow-testing before switching, but provides no results from such a test. He also says a one-point score difference falls within the noise, so small gaps should not be read as decisive evidence of a capability advantage.

Several factors remain unresolved in Meyer’s account: the index’s estimates may not match an organization’s actual prompts, output lengths or pricing; he does not detail how its task-cost figures were calculated; and Sol’s low and max effort results are not yet listed. His statements about which models suit particular roles describe his current practice and judgment, rather than controlled comparisons across users.

Meyer says Sol’s high and xhigh settings take 57 to 69 seconds to produce a first token in the index. That is a reported benchmark measurement, not a promise of response time in every product or setup. His account does not say whether that delay is acceptable for all the review tasks he assigns to Sol.

Test the Split on Real Work

Meyer recommends that teams considering a similar arrangement shadow-test candidate models against their existing workflow before making a switch. That would let teams compare quality, task costs and review time on representative assignments. He does not announce a formal trial, release schedule or further benchmark update.

For his own process, Meyer says a failed review should go back to Opus with the failing case and supporting evidence, rather than a vague instruction to try harder. He also says passing tests should not be treated as approval to ship by themselves. His account does not specify what review criteria he applies beyond that guidance.

Key Questions

Which model does Meyer use to build software?

He says Opus 5.5 at high is his main setting for features, APIs, multi-file work and refactors. He uses xhigh for harder assignments such as architecture, migrations and trust boundaries.

What does Meyer use GPT-6.1 Sol for?

He assigns Sol at high or xhigh to focused investigations and independent review of files or code diffs. He says its estimated index cost is $0.32 to $0.39 per task at those settings.

Does the benchmark show which model is best for every team?

No. The Artificial Analysis index measures general capability, and the source says teams should shadow-test models on their own work before switching. Its task costs and scores may not predict results for a particular workload.

Why does Meyer rarely use Opus at max effort?

The index lists Opus at $5.98 per task and 58 points at max, compared with $3.46 and 56 points at xhigh. Meyer says the additional cost is rarely worthwhile for his assignments.

Source: ThorstenMeyerAI.com