The Three-Part AI Workflow I’m Using This September

🔍 Read the full analysis: The Three-Part AI Workflow I’m Using This September 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 current AI workflow uses Claude Opus 5.5 for building and newly released GPT-6.1 Sol for detailed investigation and review, with other models assigned narrower tasks. The account matters to developers weighing model costs, but its benchmark comparisons do not establish which model is best for any particular project.

Meyer’s workflow has three main parts: Opus 5.5 as the primary builder, GPT-6.1 Sol as a second model for investigating specific files or diffs and reviewing work, and a set of alternatives for defined jobs. He uses Opus at high effort for features, APIs, multi-file work and refactors, and at xhigh for harder work such as architecture, migrations and trust boundaries. He assigns documents and scoped subtasks to Sonnet 5.5, while Luna handles classification, extraction and routing.

The figures Meyer cites come from Artificial Analysis Intelligence Index v4.3.x, which he describes as a general-capability index rather than a measure of performance on a reader’s own workload. In that index, Opus 5.5 scores 54 at high effort and costs $1.82 per task; at xhigh, it scores 56 and costs $3.46. GPT-6.1 Sol scores 50 at high effort for $0.32 per task and 51 at xhigh for $0.39. These are the source’s reported task costs; results and costs may differ on other work.

Meyer says Sol’s advantage for review is its lower reported task cost, which makes it practical for him to use routinely as another model examines Opus’s output. He says he turns to Astra or Fable for a second opinion if Sol and Opus disagree. His guidance is to shadow-test models before switching and to treat passing tests as evidence rather than automatic approval to ship.

A Lower-Cost Seat for Code Review

Meyer’s example shows how developers can divide AI work by task and cost rather than choosing one model for every stage. A lower-cost model may be easier to run routinely as a reviewer, while a higher-cost model remains assigned to work where its results justify the expense. That approach could affect a team’s model budget, but the cited benchmark alone cannot show whether the review catches defects in a real codebase.

He also draws attention to effort settings. On the index figures he cites, raising Opus from medium to max increases its per-task cost from $1.34 to $5.98 while its score rises from 51 to 58. Meyer therefore uses high or xhigh for development and says max is rarely worthwhile for his needs. The trade-off is specific to the reported benchmark and his workflow; teams would need to compare quality, latency and cost on their own tasks.

Benchmarks Behind the Model Split

Meyer’s report frames model selection as a cost-per-task decision. Its table lists six models with scores within about 21 index points, while their reported costs range from $0.07 to $7.63 per task. The comparison includes Opus 5.5, Sonnet 5.5, Fable 5.1, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. Meyer says Opus 5.5 has the highest score in that group and Sol and Luna offer lower-cost options for work they can handle.

The report says GPT-6.1 Sol launched Sept. 29 at the same listed token prices as GPT-6 Sol: $2 per million input tokens and $10 per million output tokens. Artificial Analysis had listed medium, high and xhigh settings in the figures Meyer cites. Sol’s high and xhigh results came with reported times to first token of 57 and 69 seconds, respectively, a limitation for interactive use. The comparison is a dated snapshot; the source notes that index results and model settings can change.

““The practical reading: Sol is not the model I ask to build. It is the model I can afford to run on everything.””

— Thorsten Meyer

How the Scores Translate to Projects

The source reports benchmark scores and task costs, but does not provide enough information here to establish how representative those figures are for different software teams, project types or production usage. It does not report results from an independent evaluation of Meyer’s workflow, such as how often Sol catches issues or how the model split affects overall development time.

Meyer also notes that one index point may fall within measurement noise, that low and max settings for GPT-6.1 Sol were not yet listed, and that its higher-effort settings have long times to first token. It remains unclear how the models compare on a reader’s own code, how their prices or benchmark entries may change, and whether a separate model review catches errors when both models share the same flawed requirements.

Test the Split on Real Work

Meyer advises readers to shadow-test before switching models. A team following that approach could compare outputs on representative tasks and record quality, latency, model cost and the human time required to review results. Those checks would show whether the proposed split fits its work better than a single-model setup.

As GPT-6.1 Sol’s index entries and model offerings develop, the reported snapshot may change. Meyer’s account does not give a date for a follow-up or a scheduled benchmark update, so the next meaningful evidence will depend on further published evaluations or results from teams testing the models on their own workloads.

Key Questions

What is the three-part workflow?

Opus 5.5 builds, GPT-6.1 Sol investigates details and reviews work, and alternatives such as Sonnet 5.5 or Luna handle narrower tasks.

What benchmark supports Meyer’s comparisons?

He cites the Artificial Analysis Intelligence Index v4.3.x and says it measures general capability, not performance on a reader’s specific workload.

Why does Meyer use GPT-6.1 Sol for review?

In his cited figures, Sol costs $0.32 to $0.39 per task at high or xhigh effort. Meyer says that price makes routine review practical for him; it does not prove the model will catch issues in every codebase.

What limitation does the report give for Sol?

Meyer reports that Sol at high and xhigh effort took 57 to 69 seconds to produce a first token on the index, which may make those settings less suitable for interactive work.

Source: ThorstenMeyerAI.com