Building With Opus, Researching With Sol, Deciding With Jev

🔍 Read the full analysis: Building With Opus, Researching With Sol, Deciding With Jev on ThorstenMeyerAI.com

Thorsten Meyer published a Sept. 29 comparison recommending Claude Opus 5.5 for software development and GPT-6.1 Sol for research and review, arguing that large differences in estimated cost per task now matter alongside model scores. His figures cite the Artificial Analysis Intelligence Index v4.3.x and describe a personal workflow, not a controlled comparison across every use case.

In Meyer’s table, Opus 5.5 scores 58 index points at its top setting and costs an estimated $5.98 per task. GPT-6.1 Sol at xhigh scores 51 and is listed at $0.39 per task. Meyer assigns Opus to features, APIs, refactors and other building work, while using Sol to examine specific files or diffs and review changes. He says Sol launched on Sept. 29 at the same token prices as the prior GPT-6 Sol: $2 per million input tokens and $10 per million output tokens.

The comparison includes four other models: Claude Sonnet 5.5, Claude Fable 5.1, GPT-6 Astra and GPT-6 Luna. Meyer places Sonnet at high effort for scoped subtasks and documents, Astra or Fable as occasional second opinions, and Luna in classification, extraction and routing. He reports Luna at $0.07 per task and 37 index points, while Astra scores 53 at its top setting and costs $3.26 per task. These are figures as presented in the source, not independently verified here.

Effort settings change the estimated price as well as the score. Meyer lists Opus at high as scoring 54 for $1.82 per task, and xhigh as scoring 56 for $3.46. At max, it reaches 58 for $5.98. For Sonnet 5.5, he says max costs $7.60 per task for a score of 56, while high costs $1.08 for a score of 47. The source says Sol at high and xhigh takes 57 to 69 seconds to first token, a potential drawback for interactive use.

Cost Shapes Meyer’s Model Choices

Meyer’s recommendation reflects a shift in how teams might compare general-purpose models: once several score within a narrow band on a benchmark, the estimated cost per task can affect how often a model is used and which work it handles. In his figures, Sol’s xhigh estimate is roughly one-eighth of Astra’s and one-twentieth of Fable’s, while scoring one to two points below those models. A low-cost review pass could make it practical to check more changes with a second model, though the source does not provide independent evidence that this workflow improves outcomes.

The approach also separates model selection from effort selection. Meyer’s table shows higher Opus settings costing more for incremental score gains. That makes the task’s quality requirements and latency tolerance relevant to the choice: a slower, more expensive setting may be reserved for difficult work, while routine tasks may use a lower setting or a less costly model. These conclusions depend on the tasks and measurement method behind the cited index.

How the Comparison Scores Models

The figures are attributed to the Artificial Analysis Intelligence Index v4.3.x, which Meyer describes as a measure of general capability rather than a verdict on a team’s specific workload. His cost-per-task figures are presented alongside token prices and index scores, but the supplied source does not lay out the full task mix, calculation method or uncertainty range for every estimate. A benchmark ranking therefore cannot establish which model will perform best on a particular codebase or document workflow.

The comparison covers releases from Sept. 1 to Sept. 29, 2026: Fable 5.1, Astra, Opus 5.5, Luna, Sonnet 5.5 and GPT-6.1 Sol. Meyer says six models fall within about 20 index points while task costs differ substantially. He advises teams to shadow-test models against their existing process before switching. The article also cautions that effort cannot compensate for missing requirements, and that a different model reviewing work is not fully independent if both models receive the same flawed specification.

„“The question from ‘which model is smartest?’ to ‘which model clears my quality bar at the lowest cost per task?’”“

— Thorsten Meyer

Workload Results Still Need Testing

The published figures do not establish how the models compare on each reader’s software, prompts or review criteria. The source does not specify the full methodology behind all per-task costs, so those estimates should be treated as reported benchmark figures rather than expected bills for every workflow. It also says one index point is within the noise and that Artificial Analysis had not published low or max effort results for GPT-6.1 Sol at the time of writing.

It remains unclear whether Sol’s lower estimated cost would translate into better total workflow economics once waiting time, retries, human review and errors are counted. Meyer notes that a minute of additional human review can erase modest savings in model price, but labels his example illustrative rather than measured. The source gives no broader deployment data to confirm the recommended stack’s performance across teams.

Shadow Tests Before Switching

Meyer’s immediate recommendation is to shadow-test candidate models on a team’s own work before changing defaults. That means comparing outputs against current workflows and checking quality, cost and latency against the task’s requirements. For his own setup, he says Opus handles building, Sol supplies research and review, and Astra or Fable may be used when the two disagree.

Further index results could add effort settings that were not available for Sol in the published comparison, while practical adoption will depend on the team’s own measurements. The next decision point is whether Sol’s reported price and review role hold up on real changes without adding enough delay or human checking to outweigh the savings.

Key Questions

What model does Meyer recommend for building?

Meyer names Claude Opus 5.5 at high or xhigh effort as his main model for features, APIs, refactors and other development work.

What role does GPT-6.1 Sol have in his workflow?

He uses GPT-6.1 Sol at high or xhigh to investigate details in files or diffs and to review work produced by Opus.

How much cheaper is Sol in the cited comparison?

The source lists Sol xhigh at $0.39 per task, compared with $3.26 for Astra and $7.63 for Fable. These are estimates tied to the cited index and may not match a reader’s workload.

Does the index identify the best model for every team?

No. Meyer describes the Artificial Analysis Intelligence Index as a map of general capability and recommends testing models on the team’s own work before switching.

Source: ThorstenMeyerAI.com